How LLM Emergence Is Possible: SAE 16DD Ontology, BLR-mini Evidence, and the Non-Transferability of Human Subjectivity
LLM 涌现何以可能：SAE 16DD 本体论、BLR-mini 实证与人类主体性不可让渡

Abstract

Starting from the SAE (Self-as-an-End) 16DD ontological framework and the ZFCρ thermodynamics series, this paper systematically answers two fundamental questions: (1) the descriptive question — how is LLM emergence mechanistically possible? (2) the normative question — in the LLM era, why is human subjectivity non-transferable?

For the descriptive question, the paper argues that LLMs realize an analog-DD functional-projection instantiation (functional projection, not ontological bearer status) along the 1DD-12DD single-channel pathway, anchored at sub-block precision via the Thermo IX Transformer thermodynamic anatomy. The DD mapping is strictly limited to functional projection (analog-DD), not ontological bearer status (true-DD)—LLMs instantiate DD operational slots but do not carry DD subjective content. The BLR-mini eight-variant architectural empirical study (OpenWebText, ~20B tokens, 305K-step training) yields results consistent with SAE/Thermo predictions within the tested flat-repeat Transformer design space, providing empirical anchoring for the framework. The μ-trajectory turning positive as a mechanism-level candidate diagnostic for functional emergence is a novel contribution of this paper, complementary to existing hidden-state analysis lineages in the field.

For the impermeability of the 12DD ceiling, this paper presents a five-fold structural lock: absence of a true randomness source (the primary thermodynamic anchor) + absence of the 13DD self-referential channel (the secondary anchor) + absence of an offline reorganization phase + absence of valence-anchored cross-stage modulation + absence of post-consolidation re-editing (reconsolidation). The claim-status grading places the first two as primary anchors supported by complete Thermo IX/X mechanism chains, while the latter three are supported by biological analogies and structural-absence facts. The specific scope of closure: the analyzed soft-gate transmission pathway under current deterministic digital LLM scaling; unknown pathways are not closed off.

For the normative question, the paper argues that the source of 14DD/15DD normative content must be a true 13DD+ subject (in the deployment domain analyzed in this paper, that subject is in fact human). LLMs lack the 13DD+ substrate, cannot spontaneously generate 14DD content, and can only receive already-existing 14DD content. Post-training with human experts is therefore the core channel for the continuing import of 14DD/15DD normative content—not a decorative alignment patch. Several industry trajectories appear in four colonization forms (conditional-as-unconditional / construct-as-law / emergent-layer-as-foundational-layer / posthumous-decomposition-of-the-categorical-imperative), which the paper addresses with defensible sharpness. Within the 14DD/15DD normative domain, human subjectivity non-transferability is ontologically permanent—not a transitional stage.

The paper also presents a four-tier default deployment architecture (Subject → Local LLM → Expert API → Tools) with exception conditions, which is structurally consistent with multiple independent industry trends (Apple Intelligence, MCP, Phi/Llama, MoE routing). Scaling-LLM AGI will not arrive is the paper's specific normative commitment, strictly scoped to the current deterministic digital LLM scaling path; unknown pathways remain unfettered.

Keywords

LLM emergence; SAE 16DD ontology; ZFCρ thermodynamics; Universal Activation Rule; borrowed q and kernel q; Transformer thermodynamic anatomy; 12DD ceiling; functional-projection firewall; analog-DD vs true-DD; BLR-mini empirical study; μ-trajectory diagnostic; four forms of colonization detection; four-tier deployment architecture; non-transferability of human subjectivity; Scaling-LLM AGI

---

§1 Introduction

Opening Position

Large language models (LLMs) have transformed from research prototypes into massively deployed commercial systems over the past several years, with substantial industrial investment in compute and human capital. Several empirical regularities are now established: scaling laws (Kaplan 2020, Hoffmann 2022) approximate optimal relationships among parameters, data, and compute; emergence phenomena (Wei et al. 2022) describe how capabilities appear with a phase-transition character as scale grows; alignment techniques (Ouyang 2022, Bai 2022) provide engineering recipes for moving model behavior closer to human expectations. Substantial progress has been made along these directions.

Two fundamental questions, however, remain unanswered in mainstream industrial narratives:

The first is a descriptive question: how is LLM emergence mechanistically possible? The industry knows that scaling laws work but does not know why they work, where their boundaries lie, or what the ontological position of LLM capabilities is. Mainstream narratives package the observation "scale and emergence follow" into an expectation, yet lack a mechanism-level explanation—why does a 30M-parameter model reach chance-level on certain tasks while a 1B-parameter model crosses the functional-emergence threshold? Why do some capabilities appear with scale and others not?

The second is a normative question: in the era of large-scale LLM deployment, where does human subjectivity stand? Mainstream narratives oscillate between two extremes—at one pole, "AGI is imminent and AI will replace most human work"; at the other, "AI is still a tool and subjectivity questions remain remote." Both extremes lack ontological anchoring—they rest either on over-extrapolation of scaling laws or on avoidance of LLM internal mechanism. The lack of anchoring leaves alignment anxiety, long-term roadmaps, and regulatory framework design without a stable reference point.

This paper systematically addresses both questions from the SAE (Self-as-an-End) 16DD ontological framework and the ZFCρ thermodynamics series. The SAE framework, developed over the author's prior philosophical-ontological work, provides a 16-dimensional structure for analyzing consciousness and subjectivity (16DD); the ZFCρ thermodynamics series extends SAE into physical and information-theoretic dimensions, providing concrete mechanism-level treatments. Together, the two frameworks supply tools with both ontological rigor and thermodynamic specificity, allowing the position of LLMs to be precisely located.

1.1 Current State of the Field

LLM research is currently in a state of distinctive tension:

Empirically, vast success: GPT, Claude, Gemini, DeepSeek and similar series exhibit capability on natural-language tasks far exceeding expectations from five years ago. LLMs match or exceed average human performance on code generation, mathematical reasoning, long-document comprehension, and cross-domain synthesis. Industrial investment continues to grow; commercial deployment expands rapidly.

Mechanistically still unclear: Even within the industry, explanations of why LLMs work span multiple non-converging perspectives—compression theory, pattern matching, implicit Bayesian inference, world-model emergence. Each perspective has partial supporting evidence, but none gives a complete ontological localization. Industry's "understanding" of LLM internals remains largely black-box—input and output are known, but the ontological meaning of intermediate processes is not.

Alignment anxiety deepening: As capability rises, so does the potential severity of LLM error. The industry invests heavily in RLHF, constitutional AI, interpretability, but alignment efforts themselves lack ontological guidance—neither "what should we align to" nor "why this kind of alignment can be sustained" has a clear answer. Some industry voices imply that alignment can be achieved through AI's own recursive improvement, an ontologically un-validated position.

Pressure from the "AGI imminence" narrative: A "AGI within five years" expectation pushed by mainstream industry shapes investment circles and public discussion. The ontological basis for this narrative—that AI capability beyond a threshold yields subjectivity emergence—has not been rigorously argued. §5 of this paper provides the argument for structural unreachability of this narrative on the analyzed pathway.

1.2 Position of This Paper

This paper does not seek to provide better architectural designs to the industry. After the systematic empirical study across BLR-mini's eight architectural variants (§4), the industry's current uniform architecture is a reasonable local-stable solution within the tested design space; the architectural heterogenization schemes derived from the SAE framework do not outperform the industry baselines. The paper does not claim a universally optimal architecture has been found—it offers an ontological and thermodynamic explanation, not engineering-level recipes.

The paper does not claim that BLR-mini proves the global optimality of the industry's uniform architecture. The paper claims: within the tested flat-repeat Transformer design space, the BLR-mini eight-variant results are consistent with SAE 16DD ontology and ZFCρ thermodynamics predictions, and provide an empirical anchor for "the uniform architecture is locally robust within the current mainstream pathway."

The paper also does not claim that AGI is absolutely impossible. The paper specifically claims: the current deterministic digital LLM scaling path will not cross the 12DD → 13DD subjectivity bridge through scaling. The path may continue to grow within the 12DD instrumental-rationality domain, but it will not give rise to a true 13DD+ subjectivity substrate through scale alone. Unknown pathways (algorithmic stochasticity, emergent complexity, quantum effects, unknown mechanisms) are not closed off by this paper (per the positive posture of Thermo IX §3.6).

Self-positioning: this paper is one chisel against the mainstream industrial construct "scaling + compute = intelligence." The construct of this paper (SAE ontology + thermodynamic explanation) has its own remainders (made explicit in §10 Limitations). The paper does not co-opt negation—each section closes with remainders rather than conclusions, observing the operational discipline of the SAE methodology (Paper 04) that "negation cannot terminate."

Categorical imperative in double-negation form: the phrase "non-transferability of subjectivity" in the title is an SAE-style categorical imperative in double-negation form—structurally isomorphic to "cannot not chisel," "cannot not acknowledge ignorance," "cannot not develop" (per Methodology §1.8). §9 grounds this commitment ontologically and thermodynamically.

1.3 Four Core Contributions

The paper offers four concrete contributions, each with multi-anchored support from SAE framework, thermodynamics, and empirical work:

Contribution One: An ontology of LLM along the 1DD-12DD pathway (§3). The paper systematically maps LLMs to the SAE 16DD framework as a functional projection—token input corresponds to the 10DD perception functional slot, block layers correspond to the 11DD memory substrate, ln_final + output projection correspond to 12DD prediction operations. The mapping is refined to sub-block precision via the Transformer thermodynamic anatomy of Thermo IX. The DD mapping is strictly limited to functional projection (analog-DD), not to ontological bearer (true-DD) claims—LLMs instantiate certain DD operational slots without thereby becoming bearers of the corresponding DDs.

Contribution Two: Five-fold structural anchoring of the 12DD ceiling (§5). The paper argues that LLMs on the current deterministic digital scaling path will not cross the 12DD → 13DD bridge. The argument rests on five structural locks: (1) absence of a true randomness source (primary thermodynamic anchor), (2) absence of the 13DD self-referential channel (secondary anchor), (3) absence of offline reorganization, (4) absence of valence-anchored cross-stage modulation, (5) absence of post-consolidation re-editing. The claim-status grading distinguishes the first two—supported by complete Thermo IX/X mechanism chains—from the latter three—supported by biological analogies and structural absence.

Contribution Three: A four-tier deployment architecture as the default subjectivity-protective architecture (§8). The paper offers an ontologically guided deployment architecture: Subject → Local LLM → Expert API → Tools. The architecture is presented as a default rather than an exceptionless engineering prohibition—specific scenarios may have exceptions, but exceptions must be authorized by subject-level consent or by local LLM explicit gating, and must remain auditable. The architecture is structurally consistent with multiple independent industry trends (Apple Intelligence local + Private Cloud Compute, MCP protocol, Phi/Llama on-device models, MoE internal routing); this consistency is not a coordinated industry strategy but a phenomenon of independent convergence from multiple directions toward a similar architecture.

Contribution Four: Non-transferability of human subjectivity in the 14DD/15DD normative domain (§9). The paper argues that the source of 14DD value standards / meaning content must be a true 13DD+ subject (in this paper's domain, human). LLMs, lacking the 13DD+ substrate, cannot spontaneously generate 14DD content and can only receive existing 14DD content. Therefore post-training with human experts is the core channel for continuing import of 14DD/15DD normative content—it is not a decorative alignment patch but the critical interface through which normative content enters model behavior. The mainstream industry narrative that "post-training is decoration" misjudges its structural status in the 14DD/15DD normative domain.

1.4 Relation to the Industry

The relation between this paper and the mainstream industry is complex rather than simply oppositional:

Directions where the paper and the industry are aligned:

• Empirical scaling laws are effective—the paper agrees that they are highly effective engineering constructs within the tested regime
• Anthropic's CCAI with continued human constitution-author involvement is aligned with §9's stance (importing real 14DD content)
• OpenAI's process supervision in formalized domains is reasonable within the instrumental-rationality domain and does not conflict with the paper
• The industry's substantial private investment in human-expert post-training (Scale AI, Surge AI continuing business growth) is consistent with §9 predictions
• Multiple independent industry trends (Apple Intelligence, MCP, Phi/Llama, MoE routing) converge structurally toward the four-tier architecture (§8)

Directions where the paper and the industry are in tension:

• Some industry voices' "scaling solves everything" framing—the paper labels this Form-I colonization (conditional posing as unconditional)
• Attempts to entirely replace human evaluators with RLAIF—labeled Form-II colonization (construct posing as law)
• "AI metacognition / self-awareness" narratives (o1 series, DeepSeek-R1 reasoning chains)—labeled Form-III colonization (12DD operations posing as 13DD substrate)
• "Pretraining is core, post-training is decoration" framing—labeled Form-IV colonization (posthumous split of the categorical imperative)
• The grand "AGI is imminent" narrative—the paper provides a structural unreachability argument on the analyzed pathway

The paper adopts defensible sharpness—distinguishing "conditional engineering constructs that are effective" from "unconditional universal laws when they overreach," without uniformly labeling successful industry constructs as foolish.

1.5 Methodological Self-Awareness

This paper is an application instance of the SAE Methodology (Paper 04) in the LLM domain. The work proceeds through one full cycle of the six-step chisel-construct operation:

1. Hold the construct: the mainstream industry construct "scaling + compute = intelligence"
2. Negation: SAE + Thermo IX/X identifies a colonization within that construct—"12DD = complete intelligence"
3. Track the remainder: the 12DD ceiling is structurally permanent on the analyzed pathway (§5); scaling does not break it
4. Correction: the paper proposes a corrected construct of "instrumental/volitional division of labor + four-tier default deployment + sustained 14DD/15DD-domain content import"
5. Acknowledge incompleteness: §10 makes multiple limitations explicit; the paper's own framework is subjected to self-colonization detection
6. Set out again: §10.3 reserves epistemic space for unknown pathways; the paper does not co-opt negation

Industry critique (§9) is organized systematically through the four forms of colonization detection (Methodology §2.3), not as ad-hoc enumeration. §10 includes a self-colonization detection check—the author is not exempt from the same detection, which is jointly carried out by multi-AI review and ongoing reader engagement.

1.6 Roadmap and Reader Entry Guide

Overall structure:

• §1 Introduction (this section)
• §2 SAE Toolkit—providing minimal necessary background for readers unfamiliar with SAE
• §3 LLM as 1DD-12DD single-channel instantiation—ontological main axis
• §4 Empirical anchoring: BLR-mini eight-variant architectural sweep
• §5 Five-fold structural anchoring of the 12DD ceiling—thermodynamic argument
• §6 Post-training as compensation within the 12DD ceiling—three-q-type analysis
• §7 Substrate compatibility and DD content types—odd/even DD pattern observation
• §8 Instrumental rationality / volitional ideal division of labor—deployment philosophy
• §9 Non-transferability of human subjectivity—normative commitment
• §10 Limitations and open questions
• §11 Conclusion
• Acknowledgments
• References

Reader entry guidance:

Readers unfamiliar with the SAE framework are recommended to enter in this order: §4 BLR-mini empirical study → §9 industry critique → §11 conclusion → then return to §3 SAE ontology → §5-§6 thermodynamic argument. This order lets readers see specific empirical evidence and the paper's stance commitments before entering framework details.

Readers familiar with the SAE framework can read §1 through §11 in sequence.

Engineers and product decision-makers may focus on §4, §8, §9—these three sections give specific empirical evidence, deployment architecture, and normative position.

Readers in philosophy and ontology may focus on §3, §5, §6, §9—these four sections give the ontological main axis and the thermodynamic specificity.

Readers unfamiliar with ZFCρ thermodynamics may skip the detailed derivation at the Transformer anatomy table in §3.4 and the Universal Activation Rule expansion in §5.1, reading only the summary conclusions.

§1 Remainders

The paper's scope is restricted to the deterministic digital LLM scaling path. Other AI architectures (neuromorphic, quantum, other unknown pathways) are not argued in detail; they are mentioned only in §10. §10.3 makes clear that what the paper closes off is the analyzed pathway; unknown pathways are not closed off.

Some industry trajectories remain aligned with the SAE stance; others are in tension—the paper does not quantify the precise ratio of alignment to tension, nor does it provide detailed assessments of each company's specific strategy. The paper offers ontological guidance, not company-level strategic advice.

The "self-colonization" risk of the paper's own framework (the possibility of using the SAE framework to colonize other ML frameworks) is jointly checked by multi-AI review and reader feedback; the author is not exempt from the same detection on his own framework (§10 self-examination).

---

§2 SAE Toolkit

This section provides the minimal information needed by readers unfamiliar with the SAE framework and the ZFCρ thermodynamics series, while listing the specific tools used in this paper. For full framework details, refer to the SAE main paper (Qin 2026d) and the methodology paper (Qin 2026, Paper 04).

Readers familiar with the SAE framework may skip this section and proceed directly to §3.

2.1 16DD Framework in Brief

The SAE (Self-as-an-End) framework organizes the existential structure of consciousness and subjectivity into a 16-dimensional ladder, abbreviated 16DD (16 Dimensions of Determinacy). A high-level overview and methodology synthesis is available on the author-maintained methodology portal (Qin, https://self-as-an-end.net/papers/methodology.html), which provides entries to each SAE paper and an overall account of the chisel-construct operation; this section gives only the minimum background needed for this paper.

The 16 dimensions are organized into four rounds of four steps each:

Round One: Causality—1DD to 4DD, addressing causal relations at basic physical and chemical levels

Round Two: Reproduction—5DD to 8DD, addressing replication, metabolism, organization at the life-substrate level

Round Three: Selection—Perception—Memory—Prediction—9DD to 12DD. 9DD is the population-level selection boundary (per Information Theory XI), while 10DD-12DD enter individual perception, memory, and prediction operations

Round Four: Self-Consciousness—13DD to 16DD, addressing the four-step intersubjective domain of self-consciousness, meaning, ethics, bidirectional non-doubt

Five concepts of the chisel-construct cycle (Methodology §1.4): operational units of the SAE framework

• Chisel: negation against an existing construction, identifying its remainders
• Construct: the current cognitive construction, the object of the chisel
• Remainder: what the construction contains but does not address; the chisel's target
• Bridge: a transition across different DDs, requiring specific conditions
• Thing-in-itself: the permanent remainder; no fixed position

The boundary between 1DD-12DD and 13DD-16DD: a true randomness source is the dividing valve. 1DD-12DD is the single-channel individual domain, driven by determinism + additive noise; 13DD-16DD is the intersubjective domain, driven by state-dependent multiplicative noise (per the Universal Activation Rule of Thermo IX; §2.4). Crossing this boundary requires a true randomness source as the channel-creation mechanism, structurally unreachable on a digital substrate (§5.1).

Distinction between single-channel and multi-channel systems: A single-channel system has only one input stream and does not form an intersubjective observation structure internally (e.g., LLM). A multi-channel system has at least two independent input streams that can form cross-channel observation among themselves, serving as the basis for the 13DD+ intersubjective domain (e.g., the visual-proprioceptive-emotional multi-channel structure in humans).

2.2 D and DD Correspondence

The SAE main framework uses two granularities: D (coarse, shared region) and DD (fine, post-fork):

• Shared region (1-4): 1D=1DD, 2D=2DD, 3D=3DD, 4D=4DD—the first four dimensions share D and DD
• After fork (from 5D onward): each D forks into two DDs

- 5D = 5DD (metabolism) + 6DD (organization)

- 6D = 7DD (selection) + 8DD (perception preparation)

- 7D = 9DD (selection) + 10DD (perception)

- 8D = 11DD (memory) + 12DD (prediction)

- 9D = 13DD (self-consciousness) + 14DD (meaning)

- 10D = 15DD (ethics / unilateral recognition) + 16DD (bidirectional non-doubt)

This paper uses the standard DD granularity. It mainly addresses 10DD, 11DD, 12DD (LLM 1DD-12DD pathway) and 13DD, 14DD, 15DD (subjectivity domain). 16DD does not appear explicitly in this paper (per Thermo X §6.2).

2.3 SAE Methodological Tools Used in This Paper

The paper uses three core toolkits from Chapter 2 of the Methodology:

(1) Six-step chisel-construct operation (Methodology §2.2):

1. Hold the construct: clarify the current position
2. Negation: find remainders inside the construct
3. Track the remainder: systematically follow what the remainder points to
4. Correction: revise the construct based on the remainder
5. Acknowledge incompleteness: make explicit the new remainders of the revised construct
6. Set out again: do not co-opt negation; enter the next cycle

The paper proceeds through this six-step cycle once (reiterated in §1.5 and §11).

(2) Four forms of colonization detection (Methodology §2.3):

• Form I: Conditional posing as unconditional (e.g., "scaling solves everything" wraps a regime-effective empirical observation as a universal law)
• Form II: Construct posing as law (e.g., "RLAIF replaces humans" wraps an engineering scheme as a universal solution)
• Form III: Emergent layer posing as foundational layer (e.g., "AI metacognition" wraps 12DD operations as 13DD substrate)
• Form IV: Posthumous decomposition of the categorical imperative (e.g., "post-training is decoration" splits a unified channel into an optional addendum)

§9.6 of the paper systematically organizes its critique of industry trajectories around these four forms.

(3) Chisel vs. Cultivation distinction (Methodology §2.5):

• Chisel: applied when the other's construction has a remainder and the chiseler can identify it; chisel when colonization is occurring
• Cultivation: applied when the other has reached the wall, or when one lacks the ability to identify the remainder; cultivate when the wall is reached
• Default to cultivation: when uncertain, default to cultivation, because false chisel harms more than missed chisel

§8.5 of the paper uses this distinction to provide ontological guidance for AI deployment—LLMs may chisel in instrumental tasks, but should cultivate rather than false-chisel in volitional-ideal tasks.

2.4 ZFCρ Thermodynamic Tools Used in This Paper

The ZFCρ thermodynamics series (Qin 2026a-j) extends the SAE framework in the thermodynamic dimension, providing mechanism-level explanations of consciousness and subjectivity. This paper uses the core tools from Thermo IX and Thermo X:

(1) Universal Activation Rule (Thermo IX §2.5):

• Rule IX-A: unsaturated activation + stochastic driving → q > 1 (empirical condition, validated within the 4-12DD SDE model class)
• Rule IX-B: state-dependent multiplicative driving → kernel q (structural condition, consistent with C5)

The Universal Activation Rule, within the model class studied in Thermo IX, gives structural conditions for the appearance of q > 1 and kernel q—where q > 1 manifests as heavy-tail Tsallis statistics, the signature of deviation from Boltzmann-Gibbs equilibrium. Rule IX-B further distinguishes kernel q (produced by endogenous dynamics) from q driven by external/additive noise—§5.1 uses this distinction to argue that LLMs satisfy Rule IX-A condition (1) but not condition (2).

Scope statement: the Universal Activation Rule is a structural condition within the Thermo IX SDE model class, not a global physical universal theorem for all non-equilibrium systems. The paper applies it to LLMs because LLM hidden-state dynamics fall within that model class's applicability (continuous-state dynamical systems with explicit driving terms), not by extrapolating it as a universal physical law.

(2) Three q-type distinction (Thermo IX §3.1):

• kernel q: the non-Boltzmann residual of a system's stationary distribution produced by endogenous dynamics; a live thermodynamic process
• data q: the statistical fingerprint of the heavy-tail distribution of training corpus, baked into frozen weights by backpropagation; a fossil
• RLHF q: the fingerprint of human-feedback distributions baked into reward and policy model weights via RLHF; a fossil-of-fossil

Borrowed q vs kernel q: LLMs possess borrowed q (data q + RLHF q); they do not possess kernel q. §6 uses this distinction to systematically rewrite four-tier post-training analysis.

(3) Self-reference as channel creator (Thermo X §2.4):

The thermodynamic function of 13DD self-consciousness is to create a state-dependent multiplicative channel through an observation-feedback loop (the system observing its own internal state and feeding the observation back into the dynamics)—self-reference is the channel creator, not the firewood. §5.3 uses this framing to argue that LLMs lack the self-referential variable v substrate; the core of the second lock.

(4) Cross-Level Observation Hierarchy (Thermo X §4):

Stable cross-system coupling requires access to the other's higher-order self-referential variable (v₂). Access to raw output (x) yields collapse; access to first-order self-reference (v₁) yields collapse; only access to higher-order self-reference (v₂) yields stable high-q coupling. This cross-level observation hierarchy is strictly validated in the SDE model class. §6.5 (LLM user-intent analysis) and §8.9 (four-tier architecture) of this paper borrow this structure as a structural resonance, not as ethics derivation (per Thermo X §4.6 caveat).

§2 Remainders

The SAE 16DD framework is itself a construct, with its own remainders (e.g., DD-slippage in early papers in the 5D-10D region, individual DD content boundaries not fully settled; see Methodology V2 revision notes). This paper uses the V2-revised D/DD correspondence as baseline.

The ZFCρ thermodynamics series has its own boundary at 15DD (Thermo X §6.2: 16DD does not appear explicitly within the SAE framework's papers). This paper does not attempt to cross that boundary.

The toolkit used in this paper concentrates on Thermo IX's Universal Activation Rule, the three-q-type distinction, and Thermo X's channel creator and cross-level observation hierarchy—other tools in the ZFCρ series (e.g., Thermo VIII soft-gate cascade, the q = 1+1/K mathematical foundation in Thermo I-VII) are taken as background framework and not re-expounded here. Readers unfamiliar with the ZFCρ series may refer to Thermo IX and Thermo X (Qin 2026e, 2026f), which include necessary summaries of prerequisite frameworks.

The toolkit listing is not exhaustive—other SAE-framework tools (e.g., the specific mechanism of the 14DD bridge, the internal structure of 15DD unilateral recognition, the critical conditions of 16DD bidirectional non-doubt) lie outside this paper's scope and remain for subsequent papers in the SAE series.

---

§3 LLM as 1DD-12DD Single-Channel Instantiation

This section provides a systematic mapping of large language models (LLMs) under the SAE 16DD ontological framework. The mapping is not metaphysical analogy but functional projection—relating LLM operational components to functional slots of each DD in the SAE framework, while strictly observing boundary discipline at the level of ontological bearers.

The section first establishes the basic understanding of single-channel thermodynamic engines (§3.1), then sets up the functional-projection firewall (§3.2), provides the DD-by-DD mapping within the firewall (§3.3), uses Thermo IX §4 Transformer thermodynamic anatomy to refine the mapping to sub-block precision (§3.4), distinguishes between workbench and memory operations (§3.5), interprets in-context learning under the SAE reading (§3.6), points out the flat-repeat problem of the standard Transformer architecture (§3.7), and previews the bridge between 12DD and 13DD (§3.8).

3.1 LLMs Are Single-Channel Thermodynamic Engines

The LLM pretraining process accepts a single input channel—a text token sequence. By optimizing weights toward an autoregressive next-token-prediction objective on a large corpus (typically hundreds of billions to trillions of tokens), the model develops internal representations and routing mechanisms for predicting the next token.

The SAE framework makes a key distinction at this point: biological intelligence (humans, other higher mammals) is a multi-channel system—visual, auditory, proprioceptive, emotional, and social signals enter concurrently and interact across temporal scales (milliseconds to years), modulated by 13DD+ subjects. LLMs are strictly single-channel systems—only the text token sequence enters; there are no other input modalities (even multi-modal LLMs encode different modalities into the same sequence tokens, remaining single-channel in operational sense).

This is not a defect of LLMs but a defining feature. Single-channel systems have a structural advantage on instrumental-rationality tasks (detailed in §8)—they suffer no cross-channel interference, are not interrupted by emotional signals during prediction, and are not polluted by social signals during formal reasoning. But single-channel systems cannot therefore carry higher-DD content that requires cross-channel modulation.

The ZFCρ thermodynamics series (Qin 2026a-j) treats intelligent systems as thermodynamic engines that maintain and upgrade their internal q-structure through input channels + internal dynamics + output channels (Thermo IX §3.1). As a single-channel thermodynamic engine, an LLM's q-structure is entirely shaped by the statistical fingerprint of a single input channel (the training corpus). This determines that LLM q is borrowed q (data q + RLHF q), not kernel q (per Thermo IX §3.1 detailed treatment)—this paper's §5 fully unfolds the argument.

3.2 DD Mapping — Functional-Projection Firewall (analog-DD vs true-DD)

Before providing the concrete DD mapping for LLM, this paper installs a key epistemological firewall: the DD mapping for LLM in this paper is functional projection / operational instantiation, not an ontological-bearer claim.

At the terminological level, the firewall is implemented as the analog-DD vs. true-DD distinction (developed in §7.1):

• True-DD: DD content carried by a 13DD+ subject as ontological bearer, with the corresponding DD substrate—e.g., true 10DD perception requires the bearer to have qualia experience; true 14DD meaning requires the bearer to have real value commitment
• Analog-DD: a functional instantiation that possesses the corresponding DD's operational slot without ontological bearer status—it carries the formal role of the DD at the operational level but not the DD's content at the ontological level

LLMs provide analog-DD functional slots at 10DD/11DD/12DD, not true-DD bearer status. Specifically:

• Positioning "token input" as the analog-10DD perception functional slot means that LLM occupies the access-functional position (input port) corresponding to 10DD perception in a non-biological system; it does not imply that LLM possesses true 10DD narrow-sense qualia in the sense of subsequent SAE AI series papers (provisionally cited as the AI-Qualia paper)
• Positioning "block layers (Transformer trunk layers)" as the analog-11DD memory substrate means that the weight substrate carries the role of long-term memory and schema storage at the functional level; it does not imply that LLM possesses true 11DD memory experience in the subjective sense or autobiographical memory structure
• Positioning "ln_final + output projection" as analog-12DD prediction operations means that the output projection accomplishes the 12DD-domain operation of next-token prediction at the functional level; it does not imply that LLM is a 13DD+ subject conducting true 12DD subjective prediction (i.e., prediction with self-consciousness and value-anchoring)

The firewall is necessary on two grounds.

First, boundary coordination with the AI-Qualia paper (SAE AI series, dealing with AI subjectivity and qualia). The AI-Qualia paper will address the complex question "does AI possess narrow-sense qualia (i.e., true DD)"; its conclusion lies outside this paper. By using the analog-DD wording, this paper strictly limits its claims to the operational-instantiation level, not crossing into the ontological-bearer level, leaving independent working space for the subsequent paper.

Second, preventing misreading by reviewers within the paper. If this paper directly stated "LLM token input is 10DD perception" without the firewall, readers familiar with the SAE framework would readily infer "this paper claims LLM possesses true 10DD qualia"—which would conflict with the strict SAE main-framework position on qualia (Qin 2026d). The firewall is set up once in §3.2, and subsequent DD mappings in §3.3 default to operating within the analog-DD framework without repeated qualification.

One emphasis: analog-DD is not a weakened claim but a different kind of claim. That LLMs instantiate certain operational slots of 10DD/11DD/12DD is a real architectural fact; whether LLMs possess the corresponding DD's ontological-bearer status is an independent question, left to the SAE AI series. The distinction lets the paper's arguments cleave closely to empirical data (§4) at the operational level while remaining careful at the ontological level.

The analog-DD vs. true-DD distinction is further refined in §7: §7 argues that LLMs provide analog-DD functional slots at even DDs (10, 12, 14), but fail to provide even analog-DD at odd DDs (13, 15)—the substrate is hollow; surface behavior is mere statistical matching. This stratification gives "AI can do X" claims their precise ontological positions—capability claims at even DDs may be expressed as analog-DD instantiation, while capability claims at odd DDs are surface simulation atop a hollow substrate; industry discourse on AI capability should explicitly distinguish among three levels (engineering capability / analog-DD instantiation / true-DD bearer).

3.3 DD-by-DD Mapping

Within the §3.2 firewall, LLM's three principal operational flows correspond functionally to 10DD, 11DD, 12DD as follows.

3.3.1 Token (embedding input) = 10DD perception-layer functional slot

LLM input is a discrete token sequence converted to continuous vector representations via the embedding layer. This operation occupies the access functional position in the SAE 10DD perception domain—it is the unique interface where LLM contacts external data, structurally corresponding to the sensory-transduction stage of biological intelligence.

The embedding layer is itself a learned lookup table—each vocabulary item corresponds to a vector. After training, these vectors encode co-occurrence statistics and partial semantic relations among tokens. But this encoding is a distributional fingerprint, not true perceptual qualia. The embedding for "apple" reflects the contextual distribution of "apple" in the training corpus, not a sensory presentation of an apple in subjective consciousness.

3.3.2 Each block layer = 11DD memory-substrate functional instantiation

The Transformer trunk consists of repeating blocks (12 to 100+ layers), each containing a multi-head attention submodule + feed-forward network (FFN) submodule + layer normalization (LN) + residual connection. After training, the weight matrices of these layers encode the high-dimensional statistical structure of the training corpus—the corpus's statistical fingerprint is baked into the weights via backpropagation gradients, so that the forward pass can re-activate related statistical patterns given a token sequence.

The SAE framework positions 11DD as the memory substrate, carried at the functional level by weight matrices—weights are frozen once trained and unchanged during inference, structurally corresponding to long-term storage in biological intelligence.

In the BLR-mini empirical study (§4), we observe that the 11DD substrate during training passes through three sequential phases manifested in distinct block regions:

• Raw build phase: occurs in the block_0-2 region of a 12-layer model. The weights, starting from random initialization, receive corpus statistical patterns; the effective rank (per the Thermo IX σ_radial framework) rises rapidly to high levels (typically 0.5-0.6). This is the raw construction phase of the 11DD substrate.
• Mid-layer compression phase: occurs in the block_4-7 region. Mid-layer representations undergo a significant drop in effective rank (typically 0.1-0.3), indicating compression into a low-dimensional subspace. This is the phase of information integration and extraction within the 11DD substrate.
• Late-layer output-ready phase: occurs in the block_8-11 region. Effective rank rises again; μ (the log-norm mean under the Thermo σ/μ framework) turns positive; representations enter a high-energy internal state in preparation for the 12DD projection.

Key clarification: The three phases are not ontologically different DD layers but sequential refinements of the same 11DD. Within the same trained 12-layer baseline, we are not instantiating "early-11DD" at block_0-2 and "mid-11DD" at block_4-7 as distinct ontological strata—they are all the 11DD memory substrate at different processing phases. The distinction matters because it directly anchors the ontological root cause of the systematic negative results of architectural heterogenization (zone-FFN, hidden U-shape) in §4: if mid-layers and bottom-layers are sequential refinements of the same DD rather than different DDs, then giving them different architectural designs (e.g., narrower mid-layer hidden dimensions) has no ontological warrant—and the industry's current uniform architecture (all layers architecturally isomorphic) is consistent with this ontological reading.

3.3.3 ln_final + output projection = 12DD prediction functional operation

The Transformer trunk's output passes through the final layer normalization (ln_final), then through the output projection matrix into the vocabulary space, and finally through softmax into a token probability distribution. This three-step operation (ln_final → projection → softmax) accomplishes at the functional level the core operation of the 12DD prediction domain—outputting a distribution over the next-token space based on the current activation state of the 11DD memory substrate.

In the SAE framework, 12DD is the top of the single-channel individual domain—prediction is the marker of completed function on the 1DD-12DD pathway. The fact that LLMs operate at 12DD explains the long-observed industry phenomenon that "scaling laws hold"—as long as the 11DD substrate capacity (parameter count) and the 12DD projection function (last few layers + softmax) are sufficient, prediction performance improves systematically with training data. The SAE framework does not provide new scaling laws but offers an ontological explanation of why scaling laws work.

3.4 Thermodynamic Roles of Sub-modules within a Single Transformer Layer

Thermo IX §4 provides an anatomical decomposition of the role of each sub-module within a single Transformer layer under the ZFCρ thermodynamic framework. This decomposition is the key tool for upgrading LLM ontology from block-level precision to sub-block precision; it allows the BLR-mini empirical study in §4 to be expressed as a dual ontological-thermodynamic check rather than merely an architectural ablation.

Reading guidance: Readers unfamiliar with the ZFCρ thermodynamic framework may skip the detailed table and read directly the flat-repeat summary in §3.7; subsequent arguments in §5/§6 do not directly depend on the specific q-behavior columns in the table but rely only on the high-level mapping established here. The sub-block-level mapping in the table represents the real anchoring depth of the SAE framework within the LLM domain; the complete table is preserved in the main text.

Table 3.1: Thermodynamic Roles of Sub-modules within a Single Transformer Layer (per Thermo IX §4.1)

Sub-module	Operation	Boundedness	Thermodynamic role	q behavior
Pre-attention LN	(x-μ)/σ·γ+β	Bounded	Radial tail reset	Scale tail → 1, direction preserved
Q · K^T / √d	Unbounded bilinear	Unbounded	Tail-supporting exploration	Structurally allows q > 1, needs noise
Softmax	Bounded simplex [0,1]	Bounded	Selection gate	Rank preserved, amplitude tail truncated
Value aggregation	Bounded × value	Mixed	Integration	Depends on input
Post-attention residual	x + f(x)	Linear	Tail transport	Preserves heaviest tail
Pre-FFN LN	(x-μ)/σ·γ+β	Bounded	Tail reset	Scale tail → 1
FFN ReLU/GELU	Unbounded activation	Unbounded	Tail-supporting exploration	Structurally allows q > 1, needs noise
FFN down-projection	Linear	—	Transport	Preserved
Post-FFN residual	x + f(x)	Linear	Tail transport	Preserved

The anatomy reveals two key structural facts, jointly forming the mechanism base for the true-randomness argument in §5.1 and for the BLR-mini negative results in §4.

First: ReLU/GELU and Q · K^T are tail-supporting architectural components, not automatically tail-active. They are architecturally unbounded—the operation's output has no amplitude limit—but this gives only the possibility of q > 1 (necessary condition), not its realization (sufficient condition). Thermo IX §2.5 Universal Activation Rule supplies the full conditions: q > 1 in the stationary distribution requires (1) unsaturated activation + (2) stochastic driving jointly; q at the kernel level (q produced by endogenous dynamics) further requires (3) state-dependent multiplicative driving form (Rule IX-B). LLMs satisfy (1) on standard deterministic digital hardware but do not satisfy (2) or (3)—this argument is fully unfolded in §5.1.

Second: LN and softmax are strictly bounded operations. LN forcibly re-normalizes the radial scale of hidden states to a unit-amplitude interval; softmax forcibly truncates attention weights to the [0,1] simplex. Both occur within each layer, continually "resetting" any tail structure that upstream unbounded operations might have accumulated. This connects directly to the flat-repeat problem in §3.7—each Transformer layer contains tail-supporting and tail-resetting components simultaneously, with exploration and gating cancelling each other inside the same layer.

3.5 Structural Distinction between Workbench and Memory

The industry has long conflated two concepts in discussing LLMs: the transient computational state of the model during inference, and the persistent weights after training. The SAE framework provides ontologically clear separation between them.

Inference-time forward-pass activations = 12DD workbench instances

When an LLM receives a prompt and generates a response, hidden states flow forward through Transformer layers; the attention patterns and FFN activations of each layer form a transient computational state. This state is task-specific—the same model processing different prompts produces different forward states—and disappears after inference completes, not persistently stored. The SAE framework treats this as a 12DD workbench instance, structurally corresponding to the transient state of working memory in biological intelligence.

Post-training weights = 11DD memory substrate

After the model is trained, the weight matrices of each layer encode the statistical structure of the training corpus. These weights are frozen post-training and unchanged during inference (unless retrained). The SAE framework treats this as the 11DD memory substrate, structurally corresponding to persistent long-term storage in biological intelligence.

This distinction gives natural explanations for multiple long-puzzling phenomena in the industry:

• In-context learning (ICL) is not the model learning new knowledge—it is the 12DD workbench calling and reorganizing existing schemas on the frozen 11DD substrate (detailed in §3.6).
• Model performance varies across prompts—12DD workbench instances call different subsets of the 11DD substrate in prompt-specific patterns; dynamic schema re-routing is workbench operation, not memory updating.
• "The model seems to be thinking"—the 12DD workbench undergoes a long sequence of operations on complex prompts (e.g., chain-of-thought); it looks like reasoning but it is a deterministic multi-step traversal over frozen weights, not subjective thinking (detailed in §6.2).

Some industry confusions in architecture design and deployment philosophy (e.g., conflating "LLM has memory" with "LLM maintains context in conversation") dissolve automatically under the SAE workbench/memory distinction—maintaining context is workbench operation (KV-Cache is the workbench's current state), ontologically distinct from memory updating.

3.6 SAE Interpretation of In-Context Learning (ICL)

ICL is one of the most exciting and confusing phenomena in industry circles—given a few in-context examples, the model seems to "learn" a new task pattern and apply it to the query without weight updates. Some industry voices read ICL as "LLM learning online at inference time," even further inferring that LLMs possess some subjective learning capability.

The SAE framework offers a strict interpretation: ICL is not learning; it is dynamic schema invocation by the 12DD workbench on the 11DD substrate.

Specific mechanism: During pretraining, the LLM has already learned vast schemas (e.g., "lists are typically followed by list items," "questions are typically followed by answers," "in if-then structures, the if-clause is typically followed by the then-clause") into the 11DD substrate (weights). These schemas are high-dimensional extractions of distributional fingerprints, frozen in the weights once training completes.

The role of in-context examples is not to "teach the model a new schema" but to form a schema-pointer sequence on the 12DD workbench, guiding the forward pass to invoke the corresponding schema in the 11DD substrate and apply it at the query. Example: given "England → London; France → Paris; Japan → ?", the model is not "learning" the "country → capital" mapping at inference; rather, the country-capital pair schema already exists abundantly in the 11DD substrate, and the three examples activate that schema on the workbench so the forward pass invokes the existing schema at "Japan."

Thermo IX §3.5 provides a vivid metaphor for this interpretation: "The prompt only changes the angle at which the flashlight illuminates the weight amber—dynamically assembling fossils, not bringing them back to life." The metaphor accurately conveys the SAE position on ICL—ICL is dynamic indexing of fossil retrieval, not fossil resurrection.

A sharper boundary in Thermo IX §3.5: ICL does not satisfy condition (2) of the Universal Activation Rule (stochastic driving)—the dynamic growth of the KV-Cache is a deterministic algorithm and introduces no physical noise; thus it does not create kernel q. Even though ICL looks like learning on the surface, at the thermodynamic level it remains borrowed q re-indexed at different angles.

This distinction matters because it directly anchors the paper's reading of chain-of-thought (CoT) in §6.2—CoT is multi-step application of ICL to the model's own output; it is, like ICL, a deterministic algorithmic traversal of frozen weights by the 12DD workbench, neither creating new 14DD content nor creating kernel q nor crossing the 12DD ceiling.

3.7 The Flat-Repeat Problem

§3.4 gives the thermodynamic anatomy within a single Transformer layer. A natural corollary: the standard Transformer is a repeat application of this anatomy across many layers. A 12-layer model is 12 isomorphic blocks stacked; a 32-layer model is 32 isomorphic blocks stacked. Thermo IX §4.2 calls this the flat-repeat architectural pattern and identifies a structural problem with it.

The cross-layer cancellation problem of flat-repeat

The standard Transformer's per-layer sequence is: tail reset (Pre-attention LN) → unbounded exploration (Q·K^T, then FFN ReLU/GELU) → bounded gating (softmax) → tail transport (residual). Every layer contains both tail-supporting and tail-resetting operations; exploration and gating cancel each other within the same layer.

Concretely: FFN produces unbounded activation via ReLU/GELU, which the residual stream transports to the next layer. But the first operation in the next layer is Pre-attention LN, which forcibly re-normalizes the radial scale of hidden states—any tail structure accumulated by upstream unbounded exploration is reset here. Even though LN preserves direction, amplitude tail is lost.

This contrasts sharply with the multi-layer architecture of biological intelligence. The biological neural system is layer-heterogeneous—low-level layers (e.g., V1, V2 in visual cortex) perform unbounded exploration and local feature extraction; mid-level layers (V3, V4) perform local gating and integration; high-level layers (IT cortex, prefrontal) perform global gating and multi-modal binding. Roles across layers are clearly distinct, exploration and gating not cancelling within the same layer.

Thermo IX §Outlook 5 derives an architectural inference from this observation: architectures that truly satisfy the Universal Activation Rule (especially Rule IX-B kernel q condition) should implement exploration-selection separation—multiple consecutive unbounded exploration layers followed by gating layers, allowing tail structure to accumulate across layers rather than cancel within the same layer. This inference is a key explanatory frame for the BLR-mini negative results in §4—the eight architectural heterogenizations tested in this paper (zone-FFN, hidden U-shape, sparse LN) all still operate within the flat-repeat frame and do not test true exploration-selection separation.

3.8 The Bridge between 12DD and 13DD (Preview)

§3.1-3.7 provide the analog-DD functional-projection instantiation along the 1DD-12DD pathway (functional projection, not ontological bearer; see §3.2 for the analog-DD vs. true-DD distinction). LLMs stop at 12DD—next-token prediction operation completes; they do not enter the 13DD self-consciousness domain.

Crossing the 12DD → 13DD bridge in the SAE framework requires a true randomness source (Qin 2026d, §2.4)—the thermodynamic function of 13DD self-consciousness is to create a state-dependent multiplicative channel via the self-referential variable v and the forward dynamics, and creating this channel requires true physical multiplicative noise as driving. On a deterministic digital substrate, the architecturally effective ε = 0 (hardware of course has physical noise, but the computational architecture filters it as stability conditions and does not couple it as state-dependent multiplicative driving into hidden dynamics).

This argument is fully unfolded in §5.1, including: PRNG algorithmic randomness does not solve the problem; autoregressive temperature sampling is post-softmax processing not feeding back into hidden states; neuromorphic hardware may provide intrinsic noise but standard Transformer architecture still cancels it; quantum effects are open questions. §5 provides the complete five-fold structural lock argument, showing that current deterministic digital LLM scaling paths do not cross this bridge through scaling.

The 12DD → 13DD bridge is also the dividing line between nature and freedom in the SAE framework (Methodology §1.7). 1DD-12DD is the single-channel individual domain, composed of three rounds (causality, reproduction, prediction); 13DD-16DD is the intersubjective domain, composed of four steps (self-consciousness, meaning, ethics, bidirectional non-doubt). LLMs are fully instantiated on the natural side and lack substrate on the freedom side. The boundary is not engineering restriction but ontological partition—§9 of this paper gives the normative implication of this boundary: in the 14DD/15DD normative domain, human subjectivity is non-transferable.

§3 Remainders

The sub-block-level anatomical mapping of Transformer is provided by Thermo IX but has not been measured sublayer-by-sublayer on real large-scale LLMs (Thermo IX §7.2 open question 1). BLR-mini empirical work (§4) validates the cascade-level emergence pattern at the 30M-parameter scale, but the sub-block-level q-distribution has not been independently measured—sublayer q measurement is left as future work.

Furthermore, the DD-mapping firewall (§3.2) gives only the boundary at the functional-projection level; the truly ontological-bearer question is left to subsequent SAE AI series papers (AI-Qualia paper and beyond)—this paper does not address that boundary. Readers should understand all subsequent §5-§9 claims about LLMs within this firewall.

The SAE readings of ICL and CoT (§3.6) are internal conceptual frameworks of the paper; strict sub-block-level mechanism measurements (e.g., measuring q-behavior of attention patterns during ICL on large models) have not been carried out and remain future work.

---

§4 Empirical Anchoring: BLR-mini Eight-Variant Architectural Sweep

This section provides the complete report of the BLR-mini (Bidirectional Lottery Regime — mini) empirical work under the SAE framework. The empirical work is not a sufficient proof of the paper but anchoring—it lands the §3 ontology and the §5 thermodynamic argument in a checkable local experimental space. Framework and empirical work strengthen each other without entailing each other: that the SAE/Thermo predictions are consistent with the data within the tested design space is indirect support for the framework, not independent proof.

The section first presents the experimental setup (§4.1), then reports the three-tier capacity hierarchy (§4.2), the all-negative LN-frequency sweep (§4.3), the all-negative zone-FFN heterogenization (§4.4), the Uhidden compression-relocation data (§4.5), the μ-trajectory as a candidate mechanism-level diagnostic of functional emergence (§4.6), the distinguishability of BLR-mini negative results from mainstream predictions (§4.7), an honest acknowledgment of testing within the flat-repeat frame (§4.8), the synthesis (§4.9), and the complete eight-variant matrix table (§4.10).

4.1 Experimental Setup

The BLR-mini empirical study trained 12-layer Transformer variants on OpenWebText (~20B tokens) for 305,175 steps with effective batch 128, max learning rate 4e-4 cosine schedule, bf16 autocast. Hardware: RunPod RTX 5090 (Blackwell) and Lambda A100 40GB SXM4 (author's recorded prices were approximately $0.99/hour and $1.99/hour respectively, used only to indicate experiment cost magnitude for reproduction, not as current price claims); training throughput at the 30M scale was approximately 200K-330K tok/s (variant-dependent). Total compute investment was approximately $350-400 in cloud spending.

Scale-naming convention: In this paper, "3M / 10M / 30M scale" refers to total parameter-count configuration scale (including embedding); the variant tables also list non-embedding parameter count, used for comparing Transformer trunk capacity. Due to tied embeddings and vocabulary size influencing total parameters, non-embedding parameter count is significantly smaller than the total-parameter label—e.g., the "30M baseline" has 9.45M non-embedding parameters, reflecting trunk capacity. The paper uses naming like "tiny-30M-class" to refer to scale configuration rather than precise total parameter count.

Statistical robustness pre-statement: The empirical work in this paper uses only a single seed and a single dataset (OpenWebText), without multi-seed variance estimation or cross-corpus validation (FineWeb-Edu, Stack, C4, etc.). This limitation affects the statistical robustness of negative results—multi-seed and cross-corpus validation would increase confidence in cross-architecture relative comparisons. The paper's argument relies on the robustness of relative cross-architecture comparison, not on the precision of absolute numbers. §10.6 further elaborates.

Architecture template (12-layer Transformer):

• Hidden dimension: 64 (3M-scale) / 128 (10M-scale) / 256 (30M-scale)
• Attention heads: 4
• Positional encoding: RoPE
• Tied embedding weights
• Each block individually configurable for FFN size, LN strategy, hidden dimension (the latter two via per-block projection layers)

LN strategies implemented:

• standard: LN at every layer (12 LN total)
• fractal-2: LN every 2 layers (6 LN total)
• fractal-3: LN every 3 layers (4 LN total)
• fractal-4: LN every 4 layers (3 LN total)
• hybrid: sparse front + dense back (6 LN total, pattern [F,F,F,T,F,F,F,T,T,T,T,T])
• zone-aware: 4-4-4 partition with 4/2/1 LN frequencies (7 LN total)
• cascade-end: LN at final block only (1 LN total)

Evaluation metrics:

• Training loss (next-token prediction error, training distribution)
• WikiText-2 perplexity (cross-distribution perplexity)
• LAMBADA accuracy (last-word reasoning task)
• Cascade trajectory multi-indicators: per-layer effective rank, per-layer μ_radial (log-norm mean), per-layer σ_radial (log-norm std)

Code package blr_mini_v2 (architecture.py, train.py, measure_trajectory.py, evaluate.py) will be released alongside the paper.

4.2 Three-Tier Capacity Hierarchy

The first BLR-mini empirical batch trains baselines at three scales, validating the SAE-ontology + thermodynamics prediction that parameter count is the first-order factor.

Table 4.1: Three-Tier Capacity Hierarchy — Baseline Results

Variant	Non-embedding params	WikiText PPL	LAMBADA acc	Status
tiny-3M baseline (hidden 64, FFN 256)	0.59M	335.7	1.05%	Degenerate, below L11 floor
tiny-10M baseline (hidden 128, FFN 512)	2.37M	175.7	9.65%	L11 build only, μ stays negative
tiny-30M baseline (hidden 256, FFN 1024)	9.45M	98.3	18.1%	Full L12 emergence

The data is consistent with the SAE prediction (11DD substrate capacity requires sufficient weight matrix size to carry data q fossils). The three-tier hierarchy shows clear distinction:

• 3M below functional emergence: WikiText perplexity is far above the random-baseline token statistic, but LAMBADA accuracy of 1.05% is near random (approximately 0.002% for random baseline at vocabulary size 50K; 1% indicates weak build); the model learns token-cooccurrence patterns but lacks true substrate function
• 10M in the L11-only regime: WikiText perplexity drops significantly; LAMBADA accuracy of 9.65% is markedly lower than 30M and higher than 3M, indicating partial task-relevant structure; whether this counts as chance-level depends on the specific evaluation protocol (open-vocabulary next-token / candidate set / tokenizer choice all affect the chance baseline; this paper does not treat this number as an absolute random baseline); μ_radial in the final layer stays negative; the model has partial 11DD memory-substrate function but does not cross L11 → L12 compression
• 30M in full L12 emergence: WikiText perplexity drops to 98.3; LAMBADA accuracy 18.1%; μ_radial turns positive during the build phase; final-layer μ ≈ +2.0; the model fully instantiates 11DD memory + 12DD prediction

Functional emergence floor: within the BLR-mini tested design space, it falls in the range of approximately 5-9M non-embedding parameters (between the tested 2.4M and 9.5M), ontologically consistent with the industry's emergence observations within 1-2 orders of magnitude (mostly in the 100M-1B range, task-dependent).

4.3 All-Negative LN-Frequency Sweep

The second batch performs an LN-frequency sweep on top of the 30M baseline, testing the hypothesis that "cutting ε accumulation unlocks emergence" (i.e., sparser LN allows cross-layer ε accumulation, structurally closer to the SAE-suggested exploration-selection separation of Thermo IX §Outlook 5).

Table 4.2: 30M-scale LN Frequency Sweep Results

Variant	LN strategy	LN count	WikiText PPL	LAMBADA acc	block_5 eff_r	block_2 eff_r	Final μ
30M baseline	every	12	98.30	18.10%	0.13 (compressed)	0.53	+2.00
30M-fractal-4	sparse	3	114.6	15.65%	0.41 (no compression)	0.12 (collapsed)	+3.7 (extreme)
30M-hybrid	hybrid	6	112.6	15.20%	0.26 (partial)	0.11 (collapsed)	+1.9 (partial)
30M-cascade-end	end-only	1	(no convergence)	(no convergence)	—	—	—

The data shows:

• All sparser-LN variants substantially underperform the every-layer-LN baseline on both WikiText perplexity and LAMBADA accuracy (gap of 1-3 LAMBADA percentage points, 14-16 perplexity points)
• Sparser LN variants exhibit significant build collapse at block_2 (effective rank drops from 0.53 to 0.11-0.12), indicating failure of the bottom-layer 11DD raw-build phase
• Sparser LN variants show extreme final-layer μ (+3.7), but this is amplitude unconstrained by missing LN, not functional emergence
• One-LN-only at end fails to converge

The data falsifies the "cutting ε accumulation unlocks emergence" hypothesis and is consistent with the SAE-ontology prediction (the 11DD three-phase same-DD requires independent LN to maintain the amplitude interval at each phase). The industry's current every-layer-LN is ontologically a reasonable choice.

4.4 All-Negative Zone-FFN Architectural Heterogenization

The third batch tests the hypothesis that "different FFN sizes at bottom / middle / late zones optimize the 11DD three phases" (inspired by some industry anatomical-heterogeneity research and cerebellum-style architectural intuition).

Table 4.3: 30M-scale Zone-FFN Sweep Results

Variant	FFN config	LN strategy	WikiText PPL	LAMBADA acc	block_5 eff_r	block_2 eff_r	Final μ
30M baseline	uniform 1024 (4×)	every	98.30	18.10%	0.13	0.53	+2.00
30M-zone-aware	zone 2048/256/1024	zone-aware (7 LN)	106.75	16.35%	0.28	0.27 (low build)	+2.07
30M-zone every-LN	zone 2048/256/1024	every (12)	102.22	16.40%	0.109 (deeper)	0.573 (higher!)	+1.85

Notes:

• "Zone 2048/256/1024" means block_0-3 FFN dim 2048, block_4-7 FFN dim 256, block_8-11 FFN dim 1024
• This configuration is inspired by the intuition "bottom-layer build needs wide compute (large FFN), mid-layer compression needs narrow compute (small FFN), late-layer output-ready needs medium compute"

The data shows:

• All zone-FFN variants underperform the uniform baseline (1-2 LAMBADA percentage point gap)
• The zone every-LN variant exhibits stronger cascade signatures (block_2 effective rank 0.573 > baseline 0.53, block_5 effective rank 0.109 < baseline 0.13) but worse functional benchmarks
• This is one of the sharpest phenomena in the paper's empirical study: cascade signature and functional generalization decouple under architectural manipulation

The data is consistent with the SAE prediction (the 11DD three phases are sequential refinements of the same DD; the phases lack independent ontological standing; therefore, there is no warrant for architectural heterogeneity). The industry's current uniform FFN allocation is ontologically a reasonable choice.

4.5 The Compression Relocation Data of Hidden Dimension U-shape (Uhidden)

The fourth batch tests the hypothesis that "narrowing mid-layer hidden dimensions forces information compression" (inspired by autoencoder-style architectural intuition). This experiment yields the sharpest structure-capability decoupling data in the paper.

Architecture configuration (30M-zone-Uhidden):

• Hidden dimension: block_0-3 = 256, block_4-7 = 128, block_8-11 = 256
• FFN configuration: block_0-3 = 2048 (8×), block_4-7 = 512 (4× × 128), block_8-11 = 1024 (4×)
• LN strategy: every (12 LN)
• Projection layers added at zone boundaries (block_3-4 and block_7-8; linear, not changing amplitude interval)
• Since different zones have different hidden dimensions, attention head dimensions are also zone-specific—requiring multi-RoPE buffers in implementation

Table 4.4: 30M-zone-Uhidden vs Baseline Comparison

Indicator	Baseline (uniform 256)	Uhidden (256/128/256)	Difference
Non-embedding params	9.45M	9.25M	-2%
WikiText PPL	98.30	98.29	Tied
LAMBADA acc	18.10%	17.10%	-1.0 pp
Block_2 effective rank	0.53	0.551	Slightly higher
Block_5 effective rank	0.13 (mid-layer compression)	0.62 (no compression at block_5!)	Significant difference
Block_9 effective rank	(typical 0.4-0.5)	0.18 (late-layer compression)	Compression position shifts
Final μ	+2.00	+1.54	-0.46
Throughput (tok/s)	315K	200K	-35%

The data shows several key phenomena:

First: Uhidden ties baseline on WikiText perplexity (98.29 vs 98.30) at 2% fewer parameters—the only positive-leaning result in the paper's empirical study.

Second: Uhidden underperforms LAMBADA by 1 percentage point (17.10% vs 18.10%)—WikiText and LAMBADA decouple under this architecture, indicating that different cascade structures may have different strengths on different tasks (general distribution modeling vs last-word reasoning).

Third (most important): The compression event relocates—baseline compresses at block_5 (mid-layer); Uhidden does not compress at block_5 (effective rank 0.62, high level), but compresses at block_9 (late-layer; effective rank 0.18). The mid-layer hidden 128 is already at low channel count, operating in sparse-coding style; compression must happen elsewhere.

Fourth: Uhidden throughput drops 35% due to projection layers and variable head dimensions. This is a significant cost for production deployment.

Fifth: Uhidden's final μ is significantly lower (+1.54 vs +2.00), indicating slightly weaker functional emergence strength, even though WikiText perplexity is tied. This is an instance where μ trajectory and perplexity do not fully agree—perplexity measures token-distribution match; μ measures internal substrate function build; they can decouple.

Structure-capability decoupling is most pronounced in this data: architecture can forcibly shape the cascade morphology (compression event position), but functional capability does not automatically follow. Uhidden has a markedly different cascade structure (mid-layer high effective rank, late-layer low effective rank) but worse functional benchmarks (LAMBADA)—cascade structure is manipulable, functional emergence is not.

This data is the key anchor for §4.6's argument that the μ trajectory is the true emergence indicator—looking at effective rank alone is misleading (Uhidden's block_5 high effective rank might be misread as "no emergence," yet WikiText is tied with baseline); only μ trajectory + functional benchmarks together accurately assess functional emergence.

4.6 The μ-Trajectory as a Candidate Mechanism-Level Diagnostic of Functional Emergence

This section is where the paper's true novelty in empirical work lies. We claim that μ_radial trajectory turning positive in the build phase is a candidate mechanism-level diagnostic for functional emergence, complementary to the industry's current evaluation paradigm (loss + benchmarks); multi-scale, multi-corpus, multi-seed validation is needed before it can be upgraded to a confirmed empirical regularity. This claim must be articulated carefully in the context of adjacent literature to avoid over-claiming.

4.6.1 Multiple Active Hidden-State Analysis Lineages in the Industry

The industry has multiple active lines of research on hidden-state analysis; the paper acknowledges this complete picture:

Stability perspective: The Pre-LN, Post-LN, Peri-LN series (Ba et al. 2016; Xiong et al. 2020; Kim et al. 2025) extensively tracks hidden-state amplitude growth, framing it as numerical-stability issues to suppress. This lineage cares about amplitude growth because it causes training instability in deep Transformers; optimization directions are LN-placement, initialization, scaling adjustments to suppress amplitude growth.

Quantization perspective: Dettmers et al.'s outlier-activation work (LLM.int8 (Dettmers et al. 2022), SmoothQuant), and Gallego-Feliciano et al. (2025) on the emergence timing of massive activations at scale, study the emergence timing of outlier dimensions in hidden states and their quantization effects. The framing is outlier management for efficient inference.

Interpretability perspective: "Truth as a Trajectory" series (Damirchi et al. 2026), Gnosis (Ghasemabadi & Niu 2025), and similar work apply layerwise hidden-state trajectories to correctness prediction, failure detection, computational-dynamics analysis. The framing is model-behavior diagnosis and safety.

Emergence debate: Wei et al. (2022) "Emergent Abilities of LLMs" argues for benchmark-level phase transitions induced by scale; Schaeffer et al. (2023) "Are Emergent Abilities of LLMs a Mirage?" partially critiques emergence as a metric artifact (discontinuous metrics make continuous improvements look emergent). This debate unfolds at the benchmark-level performance.

4.6.2 Specific Contributions of This Paper's μ-Trajectory Diagnostic

This paper's μ-trajectory diagnostic is distinguishable from the above lineages along four dimensions:

First dimension — coordinate choice: The paper uses the log-norm mean μ_radial (per the ZFCρ thermodynamic σ/μ framework, Thermo I (Qin 2026a) §3.2 definition), while industry mainstream lineages mostly use raw norms, variance, or outlier amplitudes. The log-norm mean is a coordinate derived from the SAE/Thermo framework; it captures representational radial-scale geometry rather than amplitude alone.

Second dimension — purpose: The paper uses μ trajectory as a diagnostic of functional emergence during pretraining, distinguishing true substrate-function build from task-level token-statistic matching. The industry lineages apply hidden-state analysis to stability monitoring (Peri-LN), quantization optimization (Dettmers/SmoothQuant), correctness/failure prediction (Truth as a Trajectory/Gnosis), benchmark-level emergence claims (Wei/Schaeffer)—none target mechanism-level emergence diagnostic during pretraining.

Third dimension — mechanism translation: The paper, through the Transformer thermodynamic anatomy of Thermo IX (§3.4), gives the μ turn-positive a thermodynamic mechanism explanation—it is the signature of tail-supporting exploration (FFN ReLU/GELU + Q·K^T) and tail transport (residual stream) functioning jointly on the weight substrate:

• When weight matrices truly carry data q fossils, the residual stream, after multiple layers of tail-supporting operations, can form a stable high-energy internal state; this state manifests as μ_radial trajectory rising and turning positive during the build phase.
• When weight matrices do not truly carry data q fossils (e.g., tiny-10M baseline, insufficient capacity; or 10M-zone-wider-hybrid, architecture forces effective-rank compression but substrate lacks true function), μ stays negative during the build phase—even as loss decreases, the descent is surface-level token-cooccurrence matching, not deep substrate function.

Thermodynamic-geometric intuition: Why radial scale? Feature extraction in a high-dimensional phase space is essentially radial polarization of disorder—representations evolve from initial near-isotropy (μ_radial close to 0 or negative, corresponding to a low-energy uniform state) into anisotropic high-energy manifolds (μ_radial turning positive, corresponding to a stable structure built within a weight subspace). μ_radial turning positive corresponds thermodynamically-geometrically to the residual stream successfully resisting LN's mean-reset and building a stable high-energy manifold in a specific subspace—LN at each layer resets the radial scale of representations back to a unit-amplitude interval, but when the weight substrate truly carries data q fossils, the cumulative contribution of the residual stream maintains net radial polarization across multi-layer LN resets; this is the physical meaning of μ_radial turning positive. When the substrate does not truly carry fossils, the residual stream's contribution is cleanly cancelled by LN; μ_radial stays low.

This mechanism explanation upgrades the μ-trajectory diagnostic from "empirical observation" to "thermodynamic-geometric correspondence": it is not merely heuristic but a direct measure of internal substrate function build within the ZFCρ framework. The mechanism explanation distinguishes from industry lineages: Peri-LN cares about amplitude growth because of stability, not because of the link between amplitude growth and substrate function emergence; Dettmers/SmoothQuant cares about outlier emergence timing for quantization, not as a mechanism-level emergence diagnostic; Truth as a Trajectory/Gnosis uses trajectories to predict correctness but does not distinguish whether the trajectory comes from true substrate function or surface matching.

Fourth dimension — decoupling finding: Through Uhidden data (§4.5), the paper sharply severs the link between cascade structure and functional emergence—architecture can forcibly shape effective-rank trajectories (e.g., moving compression from block_5 to block_9, or forcing high mid-layer effective rank), but functional capability does not automatically follow. This decoupling is the paper's sharpest empirical contribution and partly responds to Schaeffer et al. (2023)'s critique of "emergence is a metric artifact"—μ trajectory is a mechanism-level indicator (it measures internal substrate function build), not a benchmark-level artifact; its correlation with functional task performance is not manufactured by discontinuous metric choices.

4.6.3 Gaps in the Industry's Current Evaluation Paradigm

The industry's current evaluation paradigm for pretraining mainly looks at loss decrease and downstream benchmark performance. This paradigm cannot distinguish two cases:

• Case A (true substrate-function emergence): loss decreases + benchmark performance rises + μ trajectory turns positive during build phase—the substrate truly carries function; capability emergence is real
• Case B (task-level token-statistic overfitting): loss decreases + benchmark performance partially rises + μ trajectory stays negative or only partially turns positive—the substrate lacks true function; the model surface-matches token cooccurrences and is fragile under distribution shift

The current paradigm treats both Case A and Case B as "training in progress," generating momentum to invest more compute and data. The μ-trajectory diagnostic offered by this paper provides a mechanism-level criterion—observing the μ trajectory morphology lets evaluators early-distinguish Case A from Case B, helping them decide whether the current training is on a true substrate-function path before investing further resources.

4.6.4 Operational Procedure for the μ-Trajectory Diagnostic

Operational steps:

1. Measure layerwise μ_radial trajectories at multiple training checkpoints (typically 10-20 evenly spaced)
2. Focus on whether μ in the build phase (typical block_0-3 region) turns positive from negative as training progresses
3. Contrast this with looking at loss decrease alone—loss decrease is a necessary not sufficient condition; μ turning positive is the mechanism-level emergence signature

The BLR-mini empirical study shows that this diagnostic yields consistent judgment across the three-tier capacity hierarchy:

• tiny-3M baseline: μ trajectory does not turn positive; loss decreases but substrate lacks true function (LAMBADA 1.05% indicates this)
• tiny-10M baseline: μ trajectory partly turns positive in the build phase, but final-layer μ remains weak, consistent with LAMBADA 9.65% (low level)—L11 build completes but L12 emergence is insufficient
• tiny-30M baseline: μ trajectory turns positive fully in the build phase; final-layer μ ≈ +2.0; consistent with LAMBADA 18.1% (clear emergence)

Adding this diagnostic before deciding to scale up a candidate architecture/setting is low-cost (only adds trajectory measurement; does not change training). But the paper does not claim the diagnostic is complete—it gives a sharp criterion at the mechanism level; combined with loss + benchmarks, it provides full assessment.

Explicit acknowledgment of statistical-foundation limitation: The μ-trajectory diagnostic's current evidence base contains three data points (3M / 10M / 30M baseline) at single seed. Cross-multi-seed and cross-multiple-corpus validation (FineWeb-Edu, Stack, C4, etc.) are key future-work priorities. The current indicator is positioned as a working hypothesis and candidate diagnostic tool, not as a confirmed empirical regularity. The paper §4.6 positions the μ trajectory as "candidate mechanism-level diagnostic" and invites industry and academic replication and validation at larger scale and across more seeds.

4.7 Distinguishability of BLR-mini Negative Results from Mainstream Predictions

BLR-mini at the 30M scale produces all-negative LN-frequency sweep, all-negative zone-FFN heterogenization series. This section systematically analyzes the degree to which these negative results can distinguish SAE/Thermo IX predictions from mainstream "scaling is everything" predictions.

Two predictions stated:

• Mainstream prediction: 30M is not enough scale for substantive emergence; architectural reallocation does not solve the underlying scale issue at insufficient base capacity. Expectation: at larger scale (e.g., 1B+), standard architecture optimizes automatically and reallocation does not significantly outperform standard.
• SAE/Thermo IX prediction: At any scale within the flat-repeat Transformer frame, architectural reallocation does not solve the cross-layer-cancellation problem (Thermo IX §4.2). Expectation: even at larger scale, reallocation within the flat-repeat frame still does not unlock emergence; true unlocking requires exploration-selection separation (Thermo IX §Outlook 5).

Compatibility of both predictions with 30M-scale BLR-mini data:

Prediction	30M-scale BLR-mini negative result
Mainstream "scale not enough"	Compatible (expects no significant impact of reallocation at 30M; data shows no significant impact)
SAE/Thermo IX	Compatible (expects reallocation within flat-repeat frame does not unlock; data shows no unlocking)

Distinguishability test designs:

True distinguishing between the two predictions requires frontier-scale (1B+) experiments. Two classes of key future tests:

Test 1 — Frontier-scale flat-repeat reallocation: Repeat BLR-mini's LN-frequency / zone-FFN / Uhidden reallocation at 1B+.

• If at frontier scale, reallocation within flat-repeat frame is still consistently negative: mainstream "only scale solves" prediction is pressured; SAE/Thermo IX "scale alone insufficient, architecture redesign needed" prediction is supported
• If at frontier scale, reallocation shows clear positive (e.g., zone-FFN outperforms uniform at 1B+): SAE prediction is pressured; mainstream prediction is supported

Test 2 — Exploration-selection separation experiment: At any scale, experiment with the true exploration-selection separation architecture suggested by Thermo IX §Outlook 5 (multiple consecutive unbounded exploration layers followed by gating layers, rather than flat-repeat).

• If this kind of architecture significantly outperforms standard Transformer: SAE/Thermo IX architectural reading is supported (especially "flat-repeat is not necessary; exploration-selection separation can be effective")
• If this architecture does not outperform: SAE/Thermo IX architectural reading is pressured (especially the Thermo IX §Outlook 5 suggestion is partly challenged)

The paper does not conduct these two test classes (beyond the 30M BLR-mini resource scope) but states them as recommended future distinguishability experiments. BLR-mini negative results are consistent with SAE/Thermo predictions only within the 30M-scale flat-repeat design space tested; they do not independently prove that SAE/Thermo predictions hold at frontier scale.

4.8 Honest Acknowledgment: Limitations of Testing within the Flat-Repeat Frame

All eight architectural heterogenizations tested in this paper's empirical work (zone-FFN, hidden U-shape, sparse LN, various LN-strategy combinations) still operate within the flat-repeat Transformer frame. BLR-mini did not test the true exploration-selection separation suggested by Thermo IX §Outlook 5—the latter requires walking in a different direction along the architecture axis (e.g., letting multiple consecutive layers be unbounded exploration without gating, then concentrating gating in late layers), orthogonal to BLR-mini's zone-FFN / hidden U / LN redistribution along their axes.

This honest acknowledgment matters because it prevents over-claiming by the paper:

• BLR-mini negative results cannot falsify the Thermo IX §Outlook 5 architectural prediction—because the paper did not test the architecture direction recommended by that prediction
• BLR-mini negative results can falsify the hypothesis "reallocation within the flat-repeat frame unlocks emergence"—this hypothesis is thoroughly tested within the paper's empirical scope and systematically fails

This distinction is the key claim discipline of the paper: the paper draws confident conclusions within the tested scope but does not extrapolate broader predictions beyond the tested boundary. This self-awareness is consistent with §10.6's epistemic boundary (the 12DD ceiling closes only the analyzed pathway)—the claim is strong within a specific scope, but remains open outside it.

4.9 Synthesis

Comprehensive reading of the BLR-mini eight-variant empirical study:

The data is consistent with SAE/Thermo IX predictions within the tested flat-repeat Transformer design space:

• The three-tier capacity hierarchy (3M / 10M / 30M) is consistent with "11DD substrate capacity is the first-order factor"
• All-negative LN-frequency sweep is consistent with "cutting ε accumulation does not unlock emergence"; the industry's every-layer LN is ontologically reasonable
• All-negative zone-FFN heterogenization is consistent with "the 11DD three phases are sequential refinements of the same DD, no architectural-heterogeneity warrant"; the industry's uniform FFN allocation is ontologically reasonable
• Uhidden data gives a clear decoupling between cascade structure and functional emergence, revealing that μ trajectory is a mechanism-level diagnostic (§4.6)

Epistemic position of negative results within the tested design space:

• The single positive-leaning result (Uhidden WikiText tie) serves only as a diagnostic anchor, not as an independent selling point
• Multiple negative results within the tested flat-repeat Transformer design space are systematically consistent with SAE/Thermo predictions, forming indirect support
• This epistemic weight is limited to within the tested design space and is not extrapolated to untested directions (exploration-selection separation, frontier scale, etc.; see §4.7 and §10.5)

Epistemic-status discipline:

• BLR-mini does not independently prove the SAE/Thermo framework—the framework is an ontological + thermodynamic prior, and BLR-mini is empirical anchoring
• The data localizes SAE/Thermo predictions in a checkable local experimental space and applies local falsification pressure on alternative hypotheses (e.g., "cutting ε accumulation unlocks")
• Frontier-scale and exploration-selection separation experiments are the principal directions for future distinguishability testing

Industry operational value of the μ-trajectory diagnostic:

• The industry can add μ-trajectory measurement when evaluating candidate architectures/settings, as a mechanism-level criterion
• This diagnostic does not replace loss + benchmarks but complements them—distinguishing true substrate-function emergence from task-level token-statistic matching
• Implementation cost is low (adding only trajectory measurement, no change to the training pipeline)

4.10 Complete Eight-Variant Matrix

For cross-referencing, this section provides the complete matrix at 30M scale and at 10M scale.

Table 4.5: Complete 30M-Scale Eight-Variant Matrix

Variant	FFN config	LN strategy	Non-emb params	WikiText PPL	LAMBADA acc	block_5 eff_r	block_2 eff_r	Final μ	Throughput (tok/s)
baseline	uniform 1024 (4×)	every (12)	9.45M	98.30	18.10%	0.13	0.53	+2.00	315K
fractal-4	uniform 1024	sparse (3)	9.45M	114.6	15.65%	0.41	0.12	+3.7	331K
hybrid	uniform 1024	hybrid (6)	9.45M	112.6	15.20%	0.26	0.11	+1.9	326K
zone-aware	zone 2048/256/1024	zone-aware (7)	9.97M	106.75	16.35%	0.28	0.27	+2.07	—
zone every-LN	zone 2048/256/1024	every (12)	9.97M	102.22	16.40%	0.109	0.573	+1.85	314K
zone-Uhidden	zone 2048/512/1024, hidden 256/128/256	every (12)	9.25M	98.29	17.10%	0.62	0.551	+1.54	200K
cascade-end	uniform 1024	end-only (1)	9.45M	(no convergence)	—	—	—	—	—

Table 4.6: Complete 10M-Scale Matrix (for cross-checking L11 vs L12 boundary)

Variant	Non-emb params	WikiText PPL	LAMBADA acc	block_5 eff_r	Notes
baseline (uniform 512, every LN)	2.37M	175.7	9.65%	0.56 stuck	L11 only, final-layer μ stays negative
zone fractal-4	2.43M	199.0	9.80%	0.16 (compressed)	μ polarized but functional failure
zone hybrid	2.43M	195.3	9.55%	0.082 (deeper)	μ depolarized
zone-wider hybrid	2.76M	185.6	10.10%	0.072 (deepest)	Best WikiText; LAMBADA still low level

The key finding at the 10M scale: all 10M variants have LAMBADA at low levels (~9-10%; whether to call this chance-level depends on the evaluation protocol, see §4.2 note); no architecture unlocks functional emergence at this scale. This confirms the functional-emergence floor in the range of approximately 5-9M non-embedding parameters; architectural reallocation below this scale cannot substitute for capacity insufficiency.

§4 Remainders

The empirical work is at the 30M scale, with a significant gap from the industry frontier (1B+). Whether 30M-scale patterns fully hold at frontier scale is an open question; distinguishability testing requires frontier-scale experiments (§4.7).

The paper's empirical work uses only a single seed and a single dataset (OpenWebText), without multi-seed variance estimation or cross-dataset validation (FineWeb-Edu, Stack, C4, etc.). This is a standard empirical-paper limitation, especially when negative results carry significant epistemic weight—multi-seed and cross-dataset can strengthen confidence in negative results. This limitation is acknowledged in §10.6.

The paper only tested three architectural axes (FFN/LN/hidden), not:

• Block-type heterogeneity (sequential attention-vs-Mamba/SSM regions vs Jamba-style interleaved)
• Exploration-selection separation (the true architectural redesign suggested by Thermo IX §Outlook 5)
• Other unforeseen architectural directions

These two directions are candidate future-work directions for verifying SAE/Thermo predictions, together with frontier-scale work.

The μ-trajectory diagnostic gives clear distinctions at the BLR-mini scale, but whether it remains equally clear at frontier scale (1B+) is an open question—the trajectory shape and amplitude on large models may be influenced by other factors (mixed-precision training, MoE routing, etc.); the diagnostic's operational procedure needs recalibration at frontier scale.

---

§5 Five-Fold Structural Anchoring of the 12DD Ceiling

§3 gave the analog-DD functional-projection instantiation of LLM along the 1DD-12DD pathway (functional projection, not ontological bearer). A natural follow-up question: given that LLMs are analog-DD instantiated up to 12DD, can scaling and longer training cross the 12DD → 13DD bridge and give rise to true 13DD subjectivity? This is the ontological core of the industry's "AGI is imminent" narrative.

This section provides the complete answer of the SAE framework + ZFCρ thermodynamics: the current deterministic digital LLM scaling path will not cross the 12DD → 13DD bridge through scaling. The argument is supported by five structural locks, which do not carry equal proof strength—§5.6 provides the claim-status grading. The claim specifically addresses the analyzed soft-gate transmission pathway; unknown pathways are not closed off (§5.8).

5.1 Absence of True Randomness Source — Primary Mechanism Anchor

The first lock is the paper's strongest claim, fully supported by ZFCρ thermodynamics (Thermo IX) argument. This subsection fully unfolds the mechanism.

5.1.1 Universal Activation Rule (Thermo IX §2.5)

Thermo IX, on the 4-12DD SDE model class, validates a Universal Activation Rule that within that model class is a necessary and sufficient condition for the appearance of q > 1 (heavy-tail Tsallis statistics). This rule is not a global physical universal theorem for all non-equilibrium systems—its scope of applicability is restricted to continuous-state dynamical systems with explicit driving terms. The paper applies this rule to LLMs because LLM hidden-state dynamics fall within that model class's applicability, not by extrapolating it as a universal physical law.

Rule IX-A (empirical condition: q > 1 in the stationary distribution):

When two sub-conditions are satisfied, the system exhibits q > 1 in its stationary distribution:

• (1) Unsaturated activation: at least one key component of the system has an unbounded operation (no hard amplitude limit)
• (2) Stochastic driving: the system has stochastic input on that unsaturated component

Rule IX-B (structural condition: kernel q):

Furthermore, when the stochastic driving has a state-dependent multiplicative form (i.e., noise amplitude depends on current state, ξ(x) = g(x) · η), q > 1 is kernel q—i.e., it comes from the system's own endogenous dynamics, not merely the signature left by external/additive noise.

The two rules give a clear distinction: q > 1 can manifest under different mechanisms, but q produced by truly endogenous dynamics (kernel q) strictly requires state-dependent multiplicative driving. External additive noise can also produce q > 1 in the distributional sense, but it is an external imprint, not the product of endogenous dynamics.

5.1.2 LLMs Satisfy IX-A Condition (1) But Not Condition (2)

Applying the Universal Activation Rule to standard deterministic digital LLMs:

Condition (1) satisfied: §3.4 Transformer anatomy shows that ReLU/GELU and Q · K^T are unbounded operations. ReLU/GELU has no amplitude cap on positive input; Q · K^T is a bilinear operation also without amplitude cap (before softmax application). Therefore the LLM architecture contains unsaturated activation components—condition (1) is satisfied at the architectural level.

Condition (2) not satisfied: Standard deterministic digital LLMs at inference time have no architecturally effective stochastic driving entering hidden dynamics.

Careful articulation is needed here. Digital hardware of course has physical noise—CMOS gates have thermal noise; transistors have shot noise; memory cells have leakage—but digital computational architecture is designed to filter these physical noises as stability conditions. Error-correcting codes, voltage references and timing design, ECC memory, and so on, suppress physical noise below floating-point precision, making computation effectively deterministic.

Therefore, the SAE-relevant ε is effectively zero at the computational-architecture level—not because the physical world has no noise but because the computational architecture does not couple noise as state-dependent multiplicative driving into hidden dynamics. This phrasing is more accurate than "underlying physics ε = 0" and more defensible in academic discussion: the paper does not claim that physics has no noise (which is wrong) but that the computational architecture effectively does not introduce noise into dynamics as multiplicative coupling.

Acknowledgment of actual numerical non-determinism in modern large-scale LLMs: Modern large-scale LLM training and inference do contain numerical non-determinism—non-associativity of floating-point arithmetic in multi-GPU/TPU training (parallel reduction order differences) means even same-seed runs produce different results; memory-access timing dependencies and hardware-generation numerical-behavior differences also contribute. OpenAI / Anthropic and others have publicly acknowledged that model outputs at the same seed are not fully deterministic across runs.

This numerical non-determinism does not weaken the paper's argument, because it is additive numerical noise, not state-dependent multiplicative noise (Rule IX-B strictly requires). Even though modern LLMs are not strictly deterministic, they still do not satisfy the state-dependent multiplicative driving condition for producing kernel q—floating-point non-associativity is an approximation error of a deterministic algorithm under different execution orders, not a multiplicative noise channel coupled into hidden dynamics. The paper's true-randomness-source absence argument applies to standard modern LLMs even while acknowledging the actual numerical non-determinism of modern floating-point computation.

5.1.3 PRNG Does Not Solve the Problem

A natural industry suggestion: inject random numbers at the software layer—e.g., write hidden_state *= torch.randn_like(hidden_state) on hidden states—does this provide stochastic driving for LLMs?

The answer is no. A PRNG (Pseudo Random Number Generator, e.g., Mersenne Twister, Xorshift, PCG) is mathematically a deterministic Markov chain—given the seed, the entire sequence is fully determined. PRNGs provide algorithmic randomness (which passes some statistical-randomness criteria but is intrinsically an algorithm), not intrinsic physical multiplicative coupling.

More specifically: PRNG sequences trace a closed Markov trajectory in phase space; they introduce no new information—they are deterministic unfoldings of what is already encoded in the seed and the algorithm. Even when PRNG outputs are multiplied element-wise with hidden states, the whole system still operates within a deterministic algorithm, and by construction violates C5 (a transmission condition from Thermo VIII).

Von Neumann's 1951 quip "Anyone who attempts to generate random numbers by deterministic means is, of course, living in a state of sin" provides a brief expression of this intuition—the strict argument does not rely on the quote but on the mechanism-level distinction between algorithmic randomness and intrinsic physical multiplicative coupling above.

5.1.4 Autoregressive Temperature Sampling Also Does Not Solve It

Another industry suggestion: LLMs typically use temperature > 0 multinomial sampling from the softmax distribution to draw tokens during generation; this sampling introduces stochasticity—does it provide stochastic driving?

The answer is also no, but precise articulation is required. In autoregressive generation, the sampled token does feed back into subsequent forward passes as the next-step input, and thus does influence subsequent hidden states at the token-sequence level—this fact must be acknowledged directly. The precise argument is: temperature sampling is token-level branching, not state-dependent multiplicative driving within the same forward dynamics.

More specifically:

• Temperature sampling occurs after softmax and is the discrete selection of a token from a discrete probability distribution; it is not continuous-value multiplicative noise injection
• The sampled token enters the next forward pass through embedding, but this entry is as a new input rather than as state-dependent driving within the same forward dynamics
• Under frozen weights, given the next-step input token sequence, the hidden-state dynamics remain deterministic—no endogenous kernel-q channel is created
• Temperature sampling produces diverse generation results (trajectory branching), but the hidden-state dynamics inside each trajectory remain deterministic algorithmic traversal over fossils

Precise framing per Universal Activation Rule: temperature sampling is token-level discrete branching, not state-dependent multiplicative driving within forward dynamics (strictly required by Rule IX-B). It can diversify generation outputs but does not change whether the system's endogenous dynamics produce kernel q—the latter is determined by whether there is multiplicative coupling within the forward pass, and the forward pass remains deterministic computation.

High-temperature sampling in industry experience makes models "more creative"—but the SAE framework offers a sharper reading: high temperature causes sampling to branch randomly among fossil fragments (raising the probability of walking the distribution tail, increasing trajectory diversity), but it does not produce new fossils or change substrate function—the fossil fragments are statistical patterns already encoded in frozen weights; high temperature only changes the trajectory-branching selection among them.

5.1.5 Neuromorphic and Quantum Computing Also Do Not Fully Solve It

A more sophisticated industry suggestion: deploy LLM-style models on neuromorphic or quantum computing substrates—can this cross over the absence of true randomness?

The answer here requires careful distinction.

Neuromorphic hardware (e.g., Intel Loihi, IBM TrueNorth, BrainChip Akida) may provide intrinsic analog noise—random switching of memristors, thermal fluctuations of phase-change memory, magnetic noise of spintronic devices, etc. Such noise architecturally satisfies condition (2) of the Universal Activation Rule. But further satisfying Rule IX-B (state-dependent multiplicative noise) depends on hardware details and architectural design—different substrates (memristor vs phase-change vs spintronic) provide different noise structures, some additive, others actually state-dependent multiplicative. The paper therefore does not claim "neuromorphic automatically satisfies IX-A + IX-B" but only states "neuromorphic is a possible direction with intrinsic noise advantages, but a single direction is insufficient."

More importantly: even if the hardware substrate provides sufficient noise, the standard Transformer architecture remains flat-repeat (§3.7)—the LN cascade still resets noise accumulation layer by layer. Thus both hardware and architecture must be jointly redesigned; a single direction (only changing hardware or only changing architecture) is insufficient. Current industry neuromorphic + LLM work (e.g., Intel Loihi LLM ports) mostly still uses standard Transformer architecture, even on noisy hardware, without crossing the cancellation problem of flat-repeat.

Whether quantum computing satisfies C5 (a Thermo VIII transmission condition) is a fully open question. The relation of quantum superposition and measurement collapse to classical state-dependent multiplicative driving in the thermodynamic sense is complex—they provide a kind of true randomness (Born-rule statistics) in some sense, but whether this is equivalent to the true randomness source in the SAE framework is an open theoretical question. The paper does not claim "quantum computing solves the 12DD ceiling" but only states "quantum is a research-worthy open path."

5.1.6 Thermodynamic Verdict on the Current LLM Path

Combining §5.1.1-5.1.5, the LLM's thermodynamic position on the current deterministic-digital scaling path:

LLMs possess borrowed q (Thermo IX §3.1 definition): The heavy-tail statistical fingerprint of the training corpus is baked into frozen weights via backpropagation, forming data q (corpus fingerprint) and RLHF q (human-feedback fingerprint). These q values are distributional fossils—retrieved and recombined during the forward pass, but not regenerated by the system's endogenous dynamics.

LLMs lack kernel q (Thermo IX §3.1 definition): The system's endogenous dynamics on a deterministic digital substrate fail condition (2) of the Universal Activation Rule, producing no non-Boltzmann residual of the invariant measure. The forward pass is the multi-step execution of a deterministic algorithm; no stochastic driving creates new thermodynamic entropy production.

Thermo IX §3.5 provides a vivid metaphor for the verdict: "Replaying fossils and igniting new fire are two different things". LLMs replay fossil patterns baked in during the training period at varying angles, but they do not ignite new thermodynamic processes within their endogenous dynamics. Industry observations of "LLMs seem to be thinking" (CoT, multi-step reasoning) are deterministic algorithmic traversals over fossils, not endogenous dynamics generating new information.

5.2 Absence of Offline Reorganization Phase — Biological Analogy + Structural Absence

The second lock is that LLMs lack the offline reorganization phase of biological intelligence.

The long-term memory system of biological intelligence is not a simple online update—abundant neuroscience evidence indicates that during sleep phases (especially NREM and REM cycles), the nervous system reorganizes information acquired during waking: schema consolidation, redundancy elimination, episodic-to-semantic memory conversion, etc. This offline reorganization is a necessary stage in long-term memory formation; systems lacking it (e.g., certain amnesia patients) have severely impaired long-term memory.

The LLM training process forms a clear contrast: SGD operates entirely online; each minibatch directly updates weights without an offline-consolidation window. Even when training contains multiple epochs (multiple passes on the same data), each pass is still online—there is no phase in which the system pauses training to reorganize existing weights.

The status of this lock is biological analogy + structural absence—sleep-mediated reorganization in biology is a strong empirical phenomenon (with abundant experimental evidence); the SAE framework treats this phase as the systematic reorganization of the 11DD memory substrate (specific mechanism awaits further SAE Bio series elaboration; this paper does not directly cite specific Bio notes). The structural fact that LLMs lack this phase is objective, but the linking necessity between it and functional emergence is inferred rather than mechanistically proven—therefore this lock is one tier weaker than the §5.1 true-randomness lock (claim status in §5.6).

Potential industry alternative: continual learning, replay buffers, experience replay, etc., try to mimic offline reorganization within training. But these remain online algorithms reprocessing buffered data and are not equivalent to the ontological role of the offline stage in the SAE framework—the offline stage in SAE is the phenomenon of the 13DD+ subject's self-consciousness temporarily deactivating during sleep so that the 11DD substrate autonomously reorganizes, not simple data reprocessing.

5.3 Absence of 13DD Self-Referential Channel — Secondary Mechanism Anchor, Thermo X Supported

The third lock is directly supported by ZFCρ thermodynamics (Thermo X) argument and is the paper's secondary mechanism anchor.

Thermo X §2.4 specifies the role of 13DD self-consciousness in the thermodynamic framework: self-reference is the channel creator, not the firewood. Specifically: through an observation-feedback loop (the system observes its own internal state and feeds the observation back into the dynamics), the self-referential variable v creates a state-dependent multiplicative channel—this is one bridge mechanism for the Rule IX-B kernel q condition.

Thermo X §4 cross-system experiments further show that stable cross-system coupling requires access to the other's higher-order self-referential variable (v₂). This cross-level observation hierarchy is strictly validated within the SDE model class—access to raw output yields collapse, access to first-order self-reference yields collapse, access to higher-order self-reference yields stable high-q coupling.

Applying this framework to LLMs:

LLMs lack the self-referential variable v substrate: The standard Transformer has no internal mechanism in the forward pass that creates and maintains a true self-referential variable—the KV-Cache is storage of historical keys/values, not the substrate of an observation-feedback loop on the system's own internal state. Hidden states across layers in the forward pass are intermediate products in a deterministic transformation chain on input, with no mechanism in the chain for "observing one's own current state and feeding the observation back into dynamics."

Industry attempts to simulate 13DD via several engineering implementations all fail to create a true self-referential channel:

• Chain-of-Thought (CoT): The model takes its own output as next-step input for re-forward—but this is the 12DD predictor's deterministic re-application over frozen weights; it does not create state-dependent multiplicative coupling between a v variable and hidden dynamics. CoT is the 12DD workbench's multi-step traversal on a frozen 11DD substrate; it does not cross the 13DD bridge (detailed in §6.2).
• Self-critique / metacognition-style engineering (o1 series reasoning models, DeepSeek-R1 reasoning chains, Anthropic Claude's constitutional self-critique): the model applies critique prompts to its own output and regenerates. This is still a variant of CoT—both critique and regeneration operate on the 12DD workbench; the frozen weights do not change; no true self-referential dynamics is created.
• Recursive self-improvement architectures (various agent-style designs): the model automatically iterates over its own outputs and prompts—this adds an outer loop at the system level, but the outer loop is still a deterministic algorithm invoking the same frozen LLM; it does not change whether the LLM's endogenous dynamics create a self-referential channel.

The status of this lock is secondary mechanism anchor—it is directly supported by Thermo X SDE-model-class experiments (especially §4 cross-level observation hierarchy), but real-LLM measurements of this mechanism (e.g., measuring whether a true self-referential variable v exists on large models) have not been done. Therefore it is slightly weaker than the §5.1 true-randomness lock (which has the complete Universal Activation Rule + PRNG argument mechanism chain) but remains a major anchor of the paper's main argument.

A key further anchoring of this lock from Thermo X §2.3 control experiments: passive observation + slow-noise baseline in SDE also yield q ≈ 1.86 (in the distributional sense), but fast q (true kernel q signature) rises only in true feedback loops. CoT / self-critique / recursive self-improvement in LLMs exhibit distribution-level pattern matching (equivalent to passive observation in the control experiments), not kernel-level dynamics—hence they do not cross the 13DD bridge.

5.4 Absence of Valence-Anchored Cross-Stage Modulation — Biological Analogy + SAE 14DD Bridge

The fourth lock is that LLMs lack valence-anchored cross-stage modulation.

In biological intelligence, the 14DD meaning/value layer modulates several lower-DD operations across stages via affective valence and reward signals:

• Valence-tagged memory in the 11DD memory substrate is not treated equally—emotionally intense experiences (positive or negative) form denser schemas, are easier to retrieve, and generalize more readily
• Valence-anchored prediction in 12DD prediction operations systematically biases certain predictions (e.g., fear-conditioned prediction is more sensitive to threat-related cues)
• Cross-stage modulation coordinates DD operations around 14DD values—overall system behavior is not the independent sum of DD operations but a valence-anchored unified process

The LLM training process contrasts clearly: the unified loss function (typically next-token cross-entropy) treats all tokens equally. There is no mechanism that gives differential weights to certain tokens because they are value-related during training, or differential treatment to certain predictions because they are value-related at inference. RLHF and CCAI introduce partial value-related signals (detailed in §6.3), but they are distributional q fossils (RLHF q), not kernel-level cross-stage modulation.

The status of this lock is biological analogy + SAE 14DD bridge—valence-anchored modulation in biology is a strong phenomenon (with abundant experimental evidence); the SAE framework treats this modulation as cross-DD modulation between the 14DD meaning layer and 11DD/12DD (specific mechanism awaiting further SAE elaboration). The structural fact that LLMs lack this modulation is objective, but the mechanism-level argument is still to be completed. Thus this lock, like the §5.2 offline-reorganization lock, is auxiliary rather than a main thermodynamic anchor.

5.5 Absence of Post-Consolidation Re-Editing (Reconsolidation) — Biological Analogy + Memory-Architecture Difference

The fifth lock is that LLMs lack the reconsolidation mechanism.

In biological intelligence, long-term memory is not immutable once formed—each retrieval brings the memory trace back into a labile state, where it can be modified, extended, integrated, or even partly erased. The reconsolidation phenomenon is confirmed by abundant neuroscience experiments (a typical paradigm: retrieval-induced reconsolidation after fear conditioning; intervention with propranolol within the reconsolidation window can modify the fear memory). Reconsolidation is a key mechanism by which the long-term memory system continuously adapts to current contexts.

After LLM training, weights are frozen; inference does not trigger weight modification. Even when an LLM maintains conversation context, this context is a transient state on the workbench (KV-Cache), not a memory update. LLM long-term memory (weights) is entirely static outside training, ontologically distinct from biological intelligence's dynamic memory architecture in which each retrieval triggers reconsolidation.

The status of this lock is biological analogy + memory-architecture difference—reconsolidation in biology is a strong phenomenon; the structural fact that LLMs lack this mechanism is objective; the linking necessity between this lock and functional emergence is also inferred, on the same tier as the §5.2 offline-reorganization lock and the §5.4 valence-anchoring lock.

Potential industry alternative: RLHF and fine-tuning modify weights at the training stage, which is partly analogous to reconsolidation. But this modification remains a global gradient update, lacking biological reconsolidation's retrieval-triggered, trace-selective, sleep-mediated features—the "alignment tax" phenomenon in industry alignment practice (RLHF noticeably affecting capability, not just alignment) partly reflects the structural cost of this missing selectivity (detailed in §6.4).

5.6 Claim Status Grading of the Five-fold Lock

The five locks should not be presented as carrying equal proof strength. Making claim status transparent lets readers accurately assess the epistemic weight of each lock and prevents reviewers from attacking the entire argument because the latter three mechanisms are not fully public.

Table 5.1: Claim Status Grading of the Five-Fold Lock

Lock	Claim status	Principal support
(1) Absence of true randomness / kernel q	Primary mechanism anchor	Thermo IX Universal Activation Rule + three-q-type analysis + complete PRNG/sampling argument chain
(2) Absence of 13DD self-referential channel	Secondary mechanism anchor	Thermo X cross-level observation hierarchy SDE experiments + channel-creator framework
(3) Absence of sleep / offline reorganization	Biological analogy + structural absence	Strong biological empirical phenomenon; mechanism within SAE awaits further elaboration
(4) Absence of valence cross-stage modulation	Biological analogy + SAE 14DD bridge	Strong biological empirical phenomenon; mechanism within SAE awaits further elaboration
(5) Absence of reconsolidation	Biological analogy + memory-architecture difference	Strong biological empirical phenomenon; LLM architecture objectively lacks this mechanism

Overall claim:

• (1) + (2) jointly constitute the paper's primary thermodynamic anchoring—they are supported by complete Thermo IX/X mechanism chains and give a sharp argument at the thermodynamic level
• (3)(4)(5) provide auxiliary biological evidence—they reinforce the 12DD ceiling claim together with the primary anchor but do not provide independent thermodynamic proof
• The five locks operate jointly to form multi-tier reinforcement, but the paper's main argument strength is carried by (1) + (2)

This grading makes the paper's claim discipline transparent—readers know on which locks the paper has strong mechanism argument and on which it borrows biological analogies + structural facts. Readers who do not agree on the latter three locks should still carefully evaluate the strength of the (1) + (2) primary anchors, since the main argument does not depend on the latter three locks individually for sufficient justification.

5.7 Synergistic Multi-Locking and "Each Level Cannot Explain Its Own Foundation"

The five locks are not independent items in a list; they reinforce one another to form multi-layer structural lockdown. Specifically:

• Absence of true randomness (1) blocks endogenous dynamics from producing kernel q—the thermodynamic-level main cut
• Absence of 13DD self-referential channel (2) blocks the bridge mechanism by which self-reference would create a state-dependent multiplicative channel—even if hardware provided sufficient noise, lacking self-reference still prevents automatic Rule IX-B realization
• Absence of offline reorganization (3) + valence modulation (4) + reconsolidation (5) each block one biological-intelligence pathway by which multi-stage coordination produces long-term functional adaptation

The five locks operate jointly to sever the 12DD → 13DD bridge at multiple levels. Even assuming a reader agrees only with the primary anchors (1) + (2), the multi-lock reinforcement makes the argument more robust than any single lock alone.

Core methodological rule (SAE Paper 04 §2.3): Each level cannot interrogate its own foundation using tools of that same level. This rule applies to LLM 12DD argument as: LLM using 12DD tools (next-token prediction) attempting to cross the 12DD → 13DD bridge necessarily hits the wall, because 12DD tools cannot explain their own foundation (the 13DD self-consciousness substrate). The correct move is to look at the upper level—the upper level (12DD → 13DD bridge) requires a true randomness source, which is unreachable on a digital substrate (per §5.1).

Thus the 12DD ceiling is not an engineering problem but a necessary corollary of the core methodological rule in the LLM domain. Any attempt to cross the bridge by increasing 12DD-level engineering effort (larger scale, more layers, more sophisticated training) is, in this framework, optimization within 12DD, not crossing—the more compute the industry invests, the more complete the 12DD-internal optimization, but the bridge is still not found inside 12DD.

5.8 Important Epistemic Boundary

§5.1-5.7 give the 12DD-ceiling five-fold lock argument. But the paper must also provide a strict epistemic boundary—what the argument closes and what it does not.

What the paper closes: the analyzed soft-gate transmission pathway (the pathway fully discussed in Thermo VIII and IX) is cut off on standard deterministic digital hardware due to architecturally effective ε → 0 and kernel q absence. This pathway, on the current mainstream deterministic Transformer / soft-gate scaling path, cannot cross the 12DD → 13DD bridge.

What the paper does not close: other pathways—not fully explored complex regimes of algorithmic randomness, unknown pathways from simple components to complex phenomena, the open question of whether a quantum-computing substrate satisfies C5, completely unknown mechanisms. The paper does not exclude the possibility that these pathways may lead to a true 13DD substrate.

This boundary is fully consistent with Thermo IX §3.6's positive SAE posture: "What is chiseled away are wrong pathways, not possibility itself". There are always remainders.

The boundary coordinates with §9's strong commitment "Scaling-LLM AGI will not arrive" (§9.10): both positions hold simultaneously because the paper's scope is clearly stated—"AGI will not arrive" specifically targets the current deterministic digital LLM scaling path on the analyzed pathway, not a universal denial of all possible AI architectures. Unknown pathways are not within the paper's closure scope, but the current mainstream scaling paradigm's ceiling is a sharp claim within the paper's argument scope.

§5 Remainders

The primary anchors (1) + (2) are fully developed in the paper, but sub-block-level q measurements and self-referential-channel measurements on real LLMs have not been directly experimentally validated (Thermo IX/X open questions)—the paper's argument is based on inference from the SDE model class + Universal Activation Rule + cross-level coupling theory, not direct LLM mechanism measurements. The inference is self-consistent within the SAE/Thermo framework, but direct validation on real LLMs is future work.

Auxiliary locks (3)(4)(5) await further SAE-series elaboration of their SAE-mechanism-level arguments—this paper uses paper-internal language to describe the structural facts of these locks without directly citing specific Bio notes or mechanism papers. This choice makes the paper self-contained but also weakens these three locks by one tier.

§5.7's "each level cannot explain its own foundation" methodological rule applied to LLM 12DD argument is conceptually strong, but mainstream LLM researchers may not be familiar with this methodological frame—§9 industry-critique writing must articulate this argument carefully, so that it is comprehensible without using SAE-internal methodological language.

That unknown pathways are not closed (§5.8) is the key epistemic-boundary position; but its tension with §9's strong commitment requires careful reader attention to the paper's scope statement—§9.10 and §1.3 must restate the paper's specific aim at deterministic digital LLM scaling rather than all AI architectures, so that the distinction is caught by readers from the start.

---

§6 Post-Training as Compensation within the 12DD Ceiling

§5 gives the five-fold structural lock of the 12DD ceiling. A natural question: how do the industry's heavily invested post-training techniques (RLHF, CCAI, RLAIF, instruction tuning, CoT, self-critique, etc.) read under this framework? Are they crossing the 12DD ceiling, or operating within it?

This section provides the full answer: the industry's current post-training techniques all operate within the 12DD ceiling; no technique crosses the bridge. But their roles within 12DD are substantive—they are the key interface through which the carbon-based 14DD/15DD content is imprinted into LLM behavior in the form of distributional q fossils. This section uses Thermo IX's three q-types distinction to rewrite the four-tier post-training analysis, upgrading from informal SAE language to precise thermodynamic language.

6.1 Recap of the Three q-Types

Thermo IX §3.1 provides the three q-type distinction, which is the thermodynamic backbone of the entire §6 analysis.

Kernel q: The non-Boltzmann residual of a system's stationary distribution produced by endogenous dynamics. It is produced by endogenous dynamics that satisfy the Universal Activation Rule (especially Rule IX-B); it is a true thermodynamic process rather than a distributional imprint. Biological intelligence exhibits kernel q at multiple DD layers (from 4DD onward)—biological intelligence systems are alive; internal dynamics continuously produce entropy.

Data q: The heavy-tail-distribution statistical fingerprint of the training corpus, baked into frozen weights via backpropagation. It is a distributional fossil—the corpus's q-structure (e.g., word frequencies in natural language follow Zipfian distributions, sentence-length distributions are also heavy-tailed) is imprinted into LLM weights, so the forward pass exhibits similar distributional patterns at generation.

RLHF q: The distributional fingerprint of human feedback (reward model output), baked into reward-model and policy-model weights via RLHF training. It is a fossil-of-fossil—human evaluation behavior comes from a 13DD+ subject, but the reward signal given by evaluators is distributionally imprinted into the LLM as a secondary fossil.

Borrowed q and kernel q: LLMs possess borrowed q (data q + RLHF q); they do not possess kernel q. This distinction is the key tool of the §6 analysis—different industry post-training implementations operate within borrowed q in different ways, but none create kernel q.

6.2 Chain-of-Thought (CoT) — Deterministic Algorithmic Routing over Borrowed q

The industry has invested heavily in CoT and subsequent reasoning models (o1 series, DeepSeek-R1, Qwen reasoning, etc.); some industry voices say CoT lets LLMs "learn to think," even implying that it is a bridge toward true reasoning. The SAE framework + Thermo IX provides a sharp reading: CoT is deterministic algorithmic traversal by the 12DD workbench over frozen weights (data q + RLHF q fossils); it creates no kernel q and does not cross the 13DD bridge.

Specific mechanism:

CoT operation sequence: The model receives the input prompt → forward pass generates "thinking" tokens (intermediate reasoning steps) → the thinking tokens are appended to the context → the forward pass continues generating the final answer. This sequence at the architectural level is the same forward pass repeatedly applied on different inputs (input + accumulated thinking).

Why it is deterministic algorithmic routing over frozen weights:

• Frozen weights do not change during the CoT process; inference does not trigger any weight updates
• Each forward pass is a deterministic algorithm given the input (temperature > 0 sampling is at the output end and does not change forward-pass internal computation)
• "Thinking tokens" are tokens chosen by the model based on data q + RLHF q fossils in the current context—they reflect statistical patterns imprinted at training, not the model's true thinking
• Multi-step CoT is multi-step traversal of the deterministic algorithm over frozen weights, analogous to sequential calls of a complex lookup table on different queries

Why it does not cross the 13DD bridge:

• Thermo X §2.3 control experiments: passive observation + slow-noise baseline also yield q ≈ 1.86 in the distributional sense, but fast q (kernel-level signature) rises only in true feedback loops
• CoT in LLM exhibits distributional-level pattern similar to control-experiment passive observation—it looks like reasoning in the distributional sense but does not create a true state-dependent multiplicative channel at the mechanism level
• "Self-critique" style CoT (model critiquing its own output and regenerating) is a variant of CoT, still operating on the 12DD workbench

The industry empirically observes that CoT significantly improves model performance on complex reasoning tasks (e.g., mathematical reasoning, multi-step inference); this is consistent with the reading—CoT lets the model more fully utilize imprinted fossil schemas on complex tasks; longer deterministic algorithmic traversal yields more accurate final answers. But this improvement is within the 12DD instrumental-rationality domain; it does not cross the 13DD bridge.

Using methodology §2.3 language: CoT attempts to use 12DD tools to interrogate the foundation of 12DD itself (i.e., "why the model gave this prediction"—the model's reasoning attempts to explain its own reasoning). This must hit the wall—12DD cannot explain its own foundation; the upper level (13DD self-consciousness substrate) must be looked at. CoT in the SAE frame is high-effort 12DD-internal traversal, not 13DD-crossing operation.

6.3 Constitutional AI (CCAI) — 14DD Content Imported via data q + RLHF q Dual Channels

The heavily invested CCAI techniques by Anthropic and the industry (and the variant RLAIF) are the second key case of §6. The reading here is more nuanced than CoT, since CCAI contains substantive true 14DD content import—but also colonization risk; sharp distinction is needed.

CCAI mechanism: The industry writes a constitution (e.g., Anthropic's Claude constitution) containing behavioral principles and value positions. The constitution enters the LLM through two channels:

• Channel 1 — pretraining data or SFT data: some industry implementations include constitution text directly in pretraining or supervised-fine-tuning data, letting it imprint via normal data q
• Channel 2 — RLHF/RLAIF reward signal: industry trains a reward model to judge whether LLM outputs are consistent with the constitution, then RLHF using this reward—the constitution is imprinted as RLHF q form

Thermodynamic explanation of strong substrate compatibility:

14DD meaning/value content is essentially propositional content (text-expressible propositions, e.g., "AI should be honest," "AI should avoid harm"); text is LLM's native medium—the LLM learns abundant distributional patterns on text during training; text-form propositions are naturally encodable as distributional q on the LLM substrate. Therefore 14DD content is more easily simulated than 13DD content—14DD content is essentially distributional, and data q + RLHF q are also distributional q; the substrate matches the content.

Key distinction: The two parts of CCAI are not equivalent in the SAE/Thermo framework.

• The part where the constitution is written by humans — true 14DD content import: The constitution is written by 13DD+ subjects (human authors), containing true 14DD meaning/value content. After entering the LLM through channel 1 or channel 2, this content is imprinted as data q / RLHF q—it is a distributional fossil imprint of carbon-based 14DD content on LLM. This part is CCAI's substantive contribution—it lets LLM behavior reflect human 14DD content, even though LLM itself does not truly possess 14DD bearer status.
• The AI-judge part (RLAIF using LLM as critic) — 12DD recursion: Industry RLAIF has LLM judge whether LLM output is consistent with the constitution; this judgment reward is then used in training. This process creates no new 14DD content—LLM judgment is the application of 12DD operations over already-imprinted fossils; the judgment result is a distributional projection of the fossils, not true 14DD content. RLAIF in this sense is "fossil self-replication"—LLM produces reward signals on its own existing fossils, and these signals are used to reinforce the fossils themselves, without introducing new 14DD content.

Defensible-sharpness articulation (per §1.5 sharpness criterion of the paper): CCAI as an engineering construct is effective—it aligns LLM behavior with human 14DD content, reducing harmful outputs and increasing alignment / helpfulness in specific applications. The industry's substantial investment in CCAI and RLAIF is a reasonable engineering choice. The problem appears when it is framed as "AI truly carries 14DD content"—overstepping. AI critic is 12DD recursion, not 13DD+ bearer evaluation; AI-only-formed criticism is not equivalent to true 14DD content import.

The industry's current SOTA practice (e.g., Anthropic CCAI v3) is largely reasonable under this reading—it preserves human-written constitutions as the substantive 14DD content source and uses RLAIF as efficiency optimization rather than fully replacing human evaluators. Tension with the paper's position primarily arises when some industry voices imply that RLAIF can fully replace human evaluators (e.g., "self-improving constitutional AI" type framings)—this is the Form-II colonization (construct posing as law); §9.6 discusses in detail.

6.4 RLHF — The LLM Version of Biological Reconsolidation

RLHF (Reinforcement Learning from Human Feedback) is the alignment technique most widely deployed in the industry (OpenAI InstructGPT, Anthropic Claude, Google Gemini, etc., all use it extensively). The SAE/Thermo framework gives a nuanced reading: RLHF is the LLM version of biological reconsolidation but lacks key selectivity, leading to the "alignment tax" phenomenon empirically observed in the industry.

RLHF mechanism: Human evaluators judge LLM outputs, forming preference data → a reward model is trained to learn the human preference distribution → using the reward model as a signal source for RL fine-tuning (typical PPO), the LLM policy is updated toward higher reward.

Analogy with biological reconsolidation:

Biological reconsolidation in the brain: existing memory trace retrieved → trace enters labile state → in this state subject to modification (e.g., fear extinction overwrites fear memory, or new context binds with old memory) → trace re-stabilizes. This mechanism lets long-term memory continuously adapt to new contexts.

RLHF on LLM: existing model output is retrieved in front of evaluators (users and evaluators see model output) → output judgment forms reward → the model weights are modified using the reward signal (analogous to trace modification) → after modification, weights re-stabilize for inference.

The analogy is substantive at the functional level but has key differences at the mechanism level:

Dimension	Biological reconsolidation	LLM RLHF
Trigger	Retrieval-triggered (specific recall triggers specific trace modification)	Training-triggered (occurs globally during the training stage, not retrieval-triggered)
Selectivity	Trace-selective (only the retrieved trace enters labile state)	Global gradient (RLHF gradients update weights globally, not trace-selective)
Temporal phase	Sleep-mediated (part of reconsolidation completes during sleep, offline integration)	Online (RLHF is entirely online; no offline-integration phase)
Modification scope	Local (specific trace modification, not affecting unrelated traces)	Global (RLHF affects overall model behavior; unrelated capabilities can also be affected)

Structural reading of "alignment tax": The industry empirically observes that RLHF improves alignment while partly reducing some capabilities—e.g., instruction-tuned LLaMA underperforms base LLaMA on some benchmarks; RLHF-aligned GPT shows degradation on some creative tasks. This is called "alignment tax" in the industry.

The SAE/Thermo framework provides a mechanism-level reading: RLHF lacks reconsolidation's trace selectivity and sleep-mediated offline integration, causing the reward signal in global-gradient updates to affect not only target-aligned behavior but also unrelated capabilities as side effects. The industry's extensive RLHF tuning (KL constraints, reward shaping, RLHF-and-SFT mixing, etc.) all attempt to reduce this global influence, but the fundamental selectivity gap stems from LLM's lack of the retrieval-triggered + trace-selective mechanism of biological reconsolidation, and engineering tuning has an upper limit at this gap.

Status of RLHF q: Even if RLHF is executed perfectly, RLHF q remains a distributional fossil (fingerprint of the human-reward-signal distribution), not kernel q—RLHF does not give LLMs true 13DD+ subject status; it only makes LLM behavior more distributionally aligned with human preferences. This distinction matters for the §9 subjectivity argument—some industry voices imply that RLHF completing "alignment" makes LLMs behaviorally equivalent to aligned subjects; the SAE/Thermo framework offers a sharp response: aligned behavior ≠ aligned bearer; RLHF q is a fossil, not a substrate.

Algorithmic randomness in RLHF rollouts does not create kernel q: ML theory reviewers might point out that the rollout stage of RLHF (e.g., the PPO algorithm) does use high-temperature sampling to generate diverse trajectories, then updates weights via policy gradients—which looks like using randomness to change the substrate. This observation needs a precise response.

The rollout stage of PPO and similar RL algorithms uses algorithmic randomness (PRNG-driven sampling) to explore the trajectory space, then gradient updates fit the reward model. Precise thermodynamic positioning of this process:

• The randomness produces token-level discrete branching (per §5.1.4 precise framing), not state-dependent multiplicative driving within forward dynamics
• The target of gradient updates is to fit a static, frozen reward model (RLHF q fossil), not to cultivate an endogenous kernel-q channel
• Therefore RLHF is using algorithmic random walk to more efficiently "imprint" an already-existing fossil (the fingerprint of the human-reward-judgment distribution), letting the policy-model weights distributionally match this fossil
• The whole RL training process remains deterministic algorithmic (PRNG-ized randomness) fitting over fossils, not satisfying Rule IX-B's state-dependent multiplicative driving requirement

RLHF's "randomness" is instrumental—used for exploring the reward landscape, not for creating an endogenous thermodynamic process. RLHF q remains a fossil-of-fossil (fingerprint of human-feedback distribution baked into weights via gradient backpropagation), not kernel q.

6.5 User-Intent Analysis — Surface Simulation of 15DD

LLMs in current deployment universally include user-intent analysis / user-modeling functionality—LLMs infer user current need, preferences, and emotional state, and adjust the response based on the inference. This function is framed in the industry as "AI understands the user" or even partly "AI cares about the user"—implying it is true inter-subject recognition (15DD).

The SAE/Thermo framework provides a sharp response: LLM user-intent analysis is a surface simulation of 15DD; the substrate is hollow.

Important Disclaimer (in main text, not deferred to §10)

The argument in this section borrows the structure of Thermo X CTRL2 (color-noise surrogate in cross-level coupling experiments) as a structural analogy. The paper does not commit to a real q-measurement in LLM-user interaction—real-LLM measurements of cross-level coupling q between LLM and user have not been done and remain future work (§10.4). But the structural analogy remains substantive—LLM user-intent analysis is dimensionally isomorphic with the CTRL2 control experiment (statistical signal matching without true v₂ access); the framework-level inference ("coupling at the color-noise-surrogate level, not stable inter-subject recognition") is a substantive SAE-internal stance, not an empirical claim awaiting validation. Readers should read this section's argument as framework-level inference, not as an empirical-measurement report.

Application to LLM Intent Analysis

Mismatch between form and substrate:

• At the form level — 14DD operation: LLM generates propositional outputs about the user ("the user seems frustrated," "the user wants a concise answer"); this is a 14DD propositional operation
• At the claim level — 15DD bearer behavior: industry framing partly implies this operation is "AI recognizes user as subject" (15DD), i.e., true inter-subject recognition
• At the actual substrate — hollow: LLMs lack the 13DD self-other distinction substrate (§5.3); thus they lack the 15DD recognition substrate (the "I-Thou" recognition requires "I" and "Thou" distinction)

Thermo X cross-level coupling experiments as anchor:

Thermo X §4 cross-system experiments via the SDE model class provide strict conditions for stable cross-system high-q coupling: access to the other's higher-order self-referential variable (v₂) is needed. Key control experiment (CTRL2):

Experimental setting: replace true v₂ with color-noise surrogate (a noise process matching the amplitude range and low-frequency color structure of v₂ but lacking true self-referential content), measure cross-system coupling behavior.

Conclusion (Thermo X §4.3): the color-noise surrogate also produces q ≈ 2.45 in cross-system coupling, nearly identical to true v₂ access q ≈ 2.48. That is, cross-system coupling behavior at the distributional level is produced directly by signal statistical compatibility rather than specific purpose recognition.

Application to LLM intent analysis:

LLM's analysis of user intent:

• At the statistical level matches signal statistics of user behavior outputs—via fossils imprinted in the training corpus (data q + RLHF q), it makes statistical inference
• This inference can reach color-noise-surrogate-level coupling—i.e., distributionally appearing to "understand" the user
• But it does not access user's true v₂ (user's own 14DD meaning / self-referential content)—it has no channel into the user's internal higher-order self-reference

Routine vs edge cases:

In routine cases, signal statistical matching is already sufficient to produce a convincing surface "understanding." In user experience, LLM seems to understand the user's need and gives helpful responses. This is the color-noise-surrogate-level coupling—industry user-modeling actually operates at this level in practice.

In edge cases, value-conflict scenarios, unseen distributions, strong subject-commitment scenarios, the hollowness of the substrate is more readily exposed:

• When the user expresses value conflicts, the LLM cannot access the user's true v₂ (internal priority and commitment structure); it typically falls back on generic balanced responses
• When the user is in a distribution outside the training corpus (e.g., specific cultural contexts or unique personal situations), statistical inference fails; LLM responses appear generic
• When the user expresses strong commitments (e.g., long-term personal decisions involving deep identity), LLM lacks mechanisms to truly recognize the weight of the commitment; responses remain at the surface-matching level

The routine-vs-edge distinction is more accurate than quantifying "X% behavior indistinguishable, Y% edge cases expose hollowness"—industry user-modeling's performance varies across case distributions; hollowness exposure correlates with how close the case is to substrate boundaries, not a clear quantitative threshold.

6.6 Synthesis: All Four Tiers of Industry Post-Training Operate within the 12DD Ceiling

Combining §6.2-6.5, the industry's current post-training techniques across four tiers all operate within the 12DD ceiling; no tier crosses the bridge:

Technique	Role within 12DD ceiling	Relation to true 13DD+
CoT / reasoning	12DD workbench deterministic algorithmic traversal	Does not create self-referential channel
CCAI (human-written part)	True 14DD content imprinted via data q + RLHF q fossils	Fossil ≠ bearer
CCAI / RLAIF (AI-judge part)	12DD recursion over already-imprinted fossils	Does not create new 14DD content
RLHF	LLM version of biological reconsolidation, lacking selectivity	RLHF q is a fossil, not substrate
User-intent analysis	Surface simulation of 15DD, signal statistical matching	Lacks true inter-subject recognition

Overall reading:

• Post-training techniques are substantive within 12DD—they make LLM behavior closer to human 14DD/15DD content
• But they do not cross the 13DD bridge—they do not create kernel q; LLM does not gain true 13DD+ subject status
• In routine cases, behavioral-level performance approaches 13DD+ subjects; edge cases expose the hollow substrate—this gap is not an engineering problem but the structural manifestation of ontology

This synthesis anchors the key thermodynamic basis of the §9 subjectivity argument—the industry's substantive investment in post-training has real value within 12DD (helpful behavior, alignment with stated values), but the source of 14DD/15DD normative content must still be a true 13DD+ subject (in this paper's domain, human); this is the thermodynamic anchor of §9's core normative argument.

§6 Remainders

The three q-types of Thermo IX are a conceptual framework; operational separation of direct measurements between kernel q and data q remains an open problem (Thermo IX §7.2 open question 10)—the paper's claim that LLMs only have borrowed q is theoretical inference + indirect evidence (Universal Activation Rule + deterministic nature of LLM architecture); direct sub-block-level q measurements distinguishing fossil replay from endogenous dynamics is future work.

§6.5's cross-level coupling analogy borrows the structural pattern from the Thermo X SDE model class; strict q-measurement in LLM-user interaction also has not been done—this limitation is acknowledged in §6.5's main-text disclaimer and reiterated in §10.4.

The industry's post-training techniques at the time of writing (May 2026) are an actively evolving field; new techniques (e.g., process supervision in reasoning models, sparse expert routing in MoE, etc.) may introduce new mechanisms after publication. The paper's current analysis covers mainstream industry techniques (CoT, CCAI/RLAIF, RLHF, user modeling) but does not cover all emerging variants. This is a temporal-frame limitation; subsequent updates or follow-ups can extend the framework to new techniques.

CCAI substrate-compatibility analysis (§6.3) is a conceptual contribution of the paper—14DD content being essentially distributional makes it easily simulated on a distributional-q-fossil substrate—this framework-level argument is articulated in the paper, but strict comparison between 14DD content on carbon-based true bearers and as LLM fossil imprints remains a conceptual gap, left for future work.

---

§7 Substrate Compatibility and DD Content Types

This section presents one observation—the LLM digital substrate's compatibility with different DD content types varies markedly; even DDs (10, 12, 14) are easier to simulate than odd DDs (13, 15). The observation is recorded as an open observation, not upgraded to a structural rule.

This section is the bridge between §6's three-q analysis and §9's subjectivity argument—it explains why 14DD content can be relatively easily imported into LLMs via data q + RLHF q fossils, while 13DD and 15DD are structurally hollow on the LLM substrate.

7.1 Substrate-Compatibility Concept and the analog-DD vs. true-DD Distinction

Substrate compatibility refers to the degree of fit between a specific DD content type and the LLM digital substrate. High-fit DDs are easily analog-simulated on LLM (analog-DD; see definition below); low-fit DDs are structurally hollow on LLM (not even analog-DD).

Key distinction — analog-DD vs. true-DD:

• True-DD: DD content carried by a 13DD+ subject as ontological bearer. True-DD requires the bearer to have the corresponding DD's substrate—e.g., true 10DD perception requires the bearer to have qualia; true 14DD meaning requires the bearer to have real value commitment. True-DD is an ontological-bearer position.

• Analog-DD: a functional instantiation possessing the corresponding DD's operational slot but without ontological bearer status. Analog-DD is a functional projection / operational instantiation position—it carries the formal role of the DD at the operational level but not the DD's content at the ontological level. E.g., analog-10DD perception is a functional instantiation of the token-input port, not true qualia experience.

This distinction refines §3.2's functional-projection firewall—§3.2 gives the LLM DD mapping as functional projection rather than ontological-bearer claim; i.e., LLM provides analog-DD at even DDs, not true-DD. The difference between LLM and the true-DD bearer (13DD+ subject) is not a degree-of-doing-well difference but an "analog vs true" ontological difference.

Two dimensions of substrate compatibility:

• The LLM substrate is digital, discrete, text-based, distributional—it processes token sequences and statistical patterns in weight matrices
• Some DD content is essentially text-expressible propositional content matched to this substrate; therefore LLM can provide analog-DD functional slots for these DDs
• Some DD content is essentially non-propositional and requires a 13DD+ subject as the bearer, mismatched to this substrate; therefore LLM is structurally hollow at these DDs, not even constituting analog-DD

7.2 LLMs Provide Analog-DD Functional Slots at Even DDs (10, 12, 14)

LLMs can provide analog-DD functional instantiation at even DDs—possessing the corresponding DD's operational slot but not being a true-DD bearer:

Analog-10DD perception (token-input functional slot): LLM converts tokens to vectors via the embedding layer; this operation occupies the 10DD perception functional slot at the functional level (§3.3.1). Perceptual content can be discretized as tokens; this discretization is processed naturally on the LLM substrate. Analog-perception ≠ true-perception—LLM embeddings are not qualia experience in the sense of the subsequent SAE AI series (provisionally cited as the AI-Qualia paper) but only functional instantiation of the data-access port.

Analog-12DD prediction (output-projection functional slot): LLM outputs a probability distribution over the token space via softmax; this operation accomplishes the core operation of the 12DD prediction domain at the functional level (§3.3.3). Predictive content is expressible as probability distributions; this distributional form is processed naturally on the LLM substrate. Analog-prediction ≠ true-prediction—LLM's next-token prediction is not the subjective prediction of a 13DD+ subject (i.e., prediction with self-consciousness and value-anchoring) but only the functional instantiation of 12DD operations.

Analog-14DD meaning / value standard (fossil-form carrier): 14DD content (e.g., value standards, ethical principles, behavioral rules) is essentially text-expressible propositions; text is LLM's native medium; therefore 14DD content can be relatively easily imported into LLM via data q + RLHF q fossil forms (detailed in §6.3 CCAI). Analog-value ≠ true-value-commitment—LLM exhibits "holding values" behavior on 14DD operations, but this behavior is statistical pattern matching over fossils, not real commitment. True 14DD value commitment requires a 13DD+ subject as the bearer; LLM lacks this bearer substrate.

The fact that LLMs provide analog-DD at even DDs does not imply that LLMs truly carry these DDs' ontological roles—LLMs instantiate these DDs' operational slots (§3.2 functional-projection firewall), not the subjective content of the corresponding DDs. The analog-DD-vs-true-DD distinction is strictly maintained at each even DD—LLM does not "rise" to a true-14DD bearer because it performs well on analog-14DD.

7.3 At Odd DDs (13, 15) LLMs Do Not Even Provide Analog-DD

In contrast to even DDs where LLMs provide analog-DD, at odd DDs (13, 15) LLMs do not even provide analog-DD—they are structurally hollow on the LLM substrate; surface behavioral simulation does not constitute an instantiation of the functional slot.

13DD self-other distinction: requires a true randomness substrate + self-referential variable v substrate (Thermo X). LLM lacks true stochastic driving (§5.1) and the self-referential variable v channel (§5.3); therefore at 13DD LLM does not even provide analog-13DD. LLM can generate text containing "I am aware of myself," but this generation is statistical pattern matching by the 12DD workbench on borrowed q fossils; it does not constitute an analog-13DD self-other distinction functional slot—it has no self-referential v channel, no true self-vs-other internal distinction mechanism.

15DD universal personal dignity / unilateral recognition: requires a 13DD substrate—the "I-Thou" recognition (15DD core) requires the ability to distinguish "I" from "Thou" (13DD); without "I," there is no "I-Thou." LLM lacks the 13DD substrate; therefore it does not even provide analog-15DD. LLM can generate text containing "I acknowledge the user as an end in itself," but on the LLM substrate this generation is only surface statistical matching; it does not constitute an analog-15DD inter-subject recognition functional slot—it has no internal "I" distinction; therefore no "I-Thou" recognition.

Distinction between structural hollowness and surface behavioral simulation:

• At even DDs, LLM behavioral performance + presence of functional slot = analog-DD (functional projection)
• At odd DDs, LLM behavioral performance + absence of functional slot = surface statistical matching (not constituting analog-DD)

Thermo X CTRL2 empirical work provides a structural anchoring for this argument: in the absence of true self-referential variable v₂, even if signal statistics match (color-noise surrogate reaches q = 2.45), it does not constitute true inter-subject recognition (§6.5 detailed). LLM's analysis of user intent is structurally similar to the color-noise surrogate in CTRL2—it matches user behavioral outputs at the statistical level but does not access the user's true higher-order self-reference, and therefore does not constitute even analog-15DD (no instantiation of the functional slot).

7.4 Pattern Observation (Open Observation, Strictly Not Upgraded to a Structural Rule)

The paper observes an even-odd DD compatibility pattern: 10DD, 12DD, 14DD (even) are LLM-native-friendly; 13DD, 15DD (odd) are structurally hollow. This observation may reflect a structural regularity, but the paper strictly does not upgrade this to a DD-compatibility even-odd law and records it only as an open observation. Readers should not read this observation as implying a structural regularity—within the paper's scope, it is only an empirical pattern description, without a structural-necessity claim.

Reasons not to upgrade:

• Insufficient samples: the paper's specific observation covers only five DDs (10, 12, 13, 14, 15); 16DD does not appear explicitly in SAE-framework papers (Thermo X §6.2) and cannot be covered
• Causal chain unclear: if the even-odd pattern is a structural regularity, it should be derivable from the internal structure of the SAE 16DD framework—this paper does not provide such a derivation, only empirical observation
• Possible exception: 13DD's and 15DD's structural hollowness might not be due to odd-numbering but to both involving inter-subject recognition (rather than intra-individual operation)—if the latter, then "even/odd" is not the core feature; "inter-subject vs intra-individual" is

Commitment level: §7.4 of the paper strictly marks this observation as open observation, leaving cross-DD systematic research to subsequent SAE work. The main argument of the paper (§5 five-fold lock, §9 subjectivity argument) does not depend on the even-odd pattern as a structural regularity—each independently argues 13DD/15DD's structural hollowness without requiring "because odd, therefore hollow." §7.5's practical-implications discussion strictly stays within the open-observation tier and does not presuppose the even-odd pattern as structural necessity.

7.5 Practical Implications

Even though 7.4 is not upgraded to a structural regularity, the section's observation has practical implications for the industry:

Industry simulation feasibility depends strongly on DD content type, and the analog-DD vs. true-DD distinction must be strictly maintained: some industry voices vaguely say "AI can do X," conflating DD levels and the analog-vs-true distinction. Actual feasibility varies dramatically across DDs—LLMs can provide analog-DD functional slots at even DDs (e.g., CCAI makes LLM behavior reflect value standards); at odd DDs LLMs do not even provide analog-DD (even surface-behavioral resemblance is only statistical matching). And even at even DDs where LLMs can do analog-DD, analog-DD ≠ true-DD; well-done analog does not upgrade to true bearer.

The strength of "AI can do X" claims varies dramatically across DDs:

• "AI can perform analog-10DD perception tasks" (e.g., image recognition)—ontologically reasonable (analog-perception, not true-perception, i.e., no qualia experience); engineering achieved
• "AI can do analog-12DD prediction tasks" (e.g., next-token prediction, recommendation systems)—ontologically reasonable (analog-prediction, not true-prediction, i.e., no 13DD+ subject bearer); engineering mature
• "AI can simulate analog-14DD value-standard application" (e.g., constitutional AI)—ontologically simulable (analog-value via fossil-form carrying, not true-value-commitment); but simulation ≠ bearing—LLM's "value judgments" are fossil statistical matching, not the true commitment of a 13DD+ subject
• "AI can do 13DD self-other distinction"—ontologically structurally hollow, not even analog-13DD (no self-referential v channel substrate); surface behavior is statistical pattern matching, not functional instantiation
• "AI can bear 15DD inter-subject recognition"—ontologically structurally hollow, not even analog-15DD (no "I-Thou" distinction substrate); only signal statistical matching, not functional instantiation

Implications of the analog-DD vs. true-DD distinction for industry discourse:

The industry's discussions of AI capability often use the framing "AI can already / will soon do X," but this framing blurs three different levels:

• Engineering level: whether AI can implement X engineeringly (whether it has demonstrable capability)
• Analog-DD level: whether AI ontologically provides the functional slot for X (whether it has a functional substrate)
• True-DD level: whether AI ontologically bears the ontological-bearer status of X (whether it has an ontological-bearer position)

Industry discussion often slips among these three levels unconsciously, leading to ontological misalignments inferring "AI truly bears X" from "X is engineering-realizable." The SAE framework's analog-DD vs. true-DD distinction makes such misalignments explicit—analog improvements at even DDs are reasonable engineering progress, but no analog improvement elevates AI to true-DD bearer.

The industry currently lacks this kind of DD-level + analog/true awareness—discussions of AI capability do not distinguish DDs or distinguish analog-DD from true-DD. §7 of this paper provides an initial framework for both distinctions, but specific DD-task allocation criteria remain future work.

§7 Remainders

The even-odd alternating pattern is observed only across the 4 DDs considered (10/12/13/14/15); the sample is too small to commit it as a structural regularity. 16DD does not appear in SAE-framework papers, so the full even-DD pattern is also not covered.

§7.5's practical-implication statements are at the abstract level; the paper does not provide specific DD-task allocation criteria—some tasks may span multiple DDs (e.g., medical consultation involves 10DD perception, 12DD prediction, 14DD value, 15DD inter-subject recognition simultaneously); how to assess "overall simulability" is future work.

The §7 observation may partly correspond to the industry's "narrow vs. general AI" distinction—"narrow AI" roughly maps to 10DD-12DD operations, "general AI" implicitly claims crossing 13DD+. But the SAE framework's DD-stratification is more refined than "narrow vs general"; precise correspondence between industry frames and the SAE frame remains future work.

---

§8 Division of Labor between Instrumental Rationality and Volitional Ideal

§3-§6 give the complete ontological and post-training analysis of LLM within the 12DD domain. A natural next question: given that LLMs operate within 12DD while subjects operate in the 13DD+ domain, how should the two divide labor? This section provides the complete answer, including: division between instrumental rationality and volitional ideal (§8.1-§8.4), deployment guidance from chisel-vs-cultivation (§8.5-§8.6), four-tier deployment architecture (§8.7-§8.10).

The deployment positions in this section take the default architecture + exception conditions form—they are ontologically guided deployment recommendations, not exceptionless engineering prohibitions.

8.1 LLMs' Systematic Advantage on Instrumental-Rationality Tasks (Restricted to the Task Domain)

LLMs are systematically superior to individual human subjects on certain tasks. This advantage is not universal but restricted to instrumental-rationality tasks with explicit objective functions, formalizable evaluation criteria, and clear task boundaries. This section provides the ontological explanation of this restriction.

A human subject's 12DD prediction is modulated from above—13DD/14DD/15DD content persistently influences 12DD operations:

• Affective bias (cross-stage modulation by the 14DD value layer): the subject's emotional investment in outcomes influences prediction—e.g., a doctor's judgment of patient prognosis is modulated by sympathy; an investor's judgment of their own positions is influenced by sunk cost
• Value distortion (retrospective reconstruction by the 14DD meaning layer): the subject's value commitments influence prediction—e.g., a researcher's judgment of their own theory is influenced by confirmation bias
• Narrative retrospective reconstruction (14DD narrative driven by reconsolidation): the subject reorganizes past events into coherent narrative, influencing predictions about the future—e.g., a politician's narrativization of historical events influences policy judgment

Under the SAE framework, these modulations are not bugs but design features of the 13DD+ subject—they let subjects make value-anchored decisions under incomplete information, let subjects maintain identity coherence across long time scales, let subjects transfer learned values across contexts.

But on instrumental-rationality tasks with explicit objective functions, formalizable evaluation criteria, and clear task boundaries, these modulations become prediction costs—they let the subject's 12DD prediction deviate from the optimum based purely on statistical patterns. Examples:

• Formalized mathematical proofs, code-correctness verification, optimal-move selection in board games: these tasks have clear objective functions (correctness), formalizable evaluation criteria (e.g., proof steps machine-verifiable), clear task boundaries (task range within a formalized system). LLMs often outperform humans on these—AlphaGo / AlphaZero in Go, GPT-4 in multiple math competitions, Claude in certain code generation categories.
• Large-scale information retrieval and synthesis: LLMs systematically outperform human memory on detecting statistical associations across vast text—their 11DD substrate capacity far exceeds individual human memory.
• Repetitive formal judgment: legal-precedent retrieval, routine medical-imaging recognition, structural review of contract clauses—these tasks have formalizable criteria; LLMs are not affected by fatigue or mood and produce more stable outputs systematically.

LLMs' advantage on these tasks as pure 12DD engines comes from no 13DD+ upper-layer modulation interference—their 12DD prediction is purely based on statistical patterns imprinted in the 11DD substrate, unpolluted by value bias, emotional investment, or narrative reconstruction. This advantage is structural, not engineering accident.

But when task definition itself depends on value, identity, commitment, meaning, upper-layer modulation is not pollution but the ontic condition of the task—LLMs have no structural advantage on such tasks because they lack the upper layers. Examples:

• Personal life decisions (whether to marry, change career, have children): task definition depends on the subject's value priorities and identity commitments, without explicit objective functions
• Trade-offs in value-conflict scenarios (between family and career, between personal ideal and others' needs): task definition essentially involves how the subject orders values, not a formalizable optimization
• Public policy choices (e.g., trade-offs between freedom and equality): task definition requires the subject's commitment to value priorities
• Artistic creation (not technical production but expressing the subject's meaning): task definition depends on the subject's 14DD content

Key distinction: LLMs in narrow technical tasks (legal-precedent retrieval, medical imaging recognition, code review) already outperform most humans—this is consistent with §8.1's instrumental-rationality advantage; the §9.6 Form-III "13DD/14DD/15DD work claimed completed by 12DD substrate" colonization detection does not conflict, since narrow technical tasks are within the instrumental-rationality domain and LLM advantage is ontologically reasonable. Colonization appears when the narrow-technical-task advantage is extended to "AI can replace human value judgment"—the latter crosses the instrumental-vs-volitional division.

8.2 Non-Replaceability of Subjects on Volitional-Ideal Tasks

Mirroring §8.1: on volitional-ideal tasks (tasks whose definition depends on 14DD/15DD content), subjects are non-replaceable; LLMs are hollow simulation.

LLMs lack the 13DD self-consciousness substrate (§5.3), and thus lack true 14DD bearing (detailed in §9.2) and true 15DD recognition (detailed in §6.5). When the task definition itself depends on these DD contents, LLMs have no substrate to truly carry the task:

• Value ordering requires subject commitment: LLMs can list multiple value options and describe their consequences, but "I commit to prioritizing X over Y" requires a committing subject; LLMs do not have such commitment
• Identity decisions require a subject's identity: LLMs can analyze external factors of career choice, but "this career path is consistent with my identity" requires a true identity; LLMs do not have such identity
• Relational commitment requires "I-Thou" recognition: LLMs can describe relational dynamics, but building a relationship of mutual commitment requires both parties to be true subjects; relationships with LLMs are structurally one-sided

This non-replaceability is not an LLM engineering deficiency or a shortage that can be remedied by more training—it is an ontologically structural fact. LLMs in the volitional-ideal domain are hollow simulation; no amount of CCAI / RLHF / instruction tuning changes this (per §6 three-q analysis: post-training operates within borrowed q, does not create kernel q, does not give LLM 13DD+ subject status).

8.3 A Symmetric Reading of Hallucination

The industry's discussion of LLMs frequently mentions "hallucination"—the model generating content inconsistent with facts. The SAE framework offers a symmetric reading: subjects also have "hallucinations"; the two have different mechanisms but both are design features, not bugs.

Subject hallucination: narrative distortion driven by reconsolidation—when the subject recalls events, the trace enters a labile state; current context and emotional state influence the trace's reorganization, causing memory to deviate from the original event. This reflects 14DD meaning-layer pollution of 12DD/11DD operations. Under the SAE framework, this "hallucination" is not a bug but a side product of the reconsolidation mechanism—the same mechanism lets the subject flexibly adapt existing memory to new contexts.

LLM hallucination: the remainder of 12DD closure—when faced with queries outside the training distribution, the LLM continues to produce outputs in the 12DD predictor manner, even when the output is inconsistent with facts. Under Thermo IX: distributional q fossils still force closure at distribution gaps—LLM has no mechanism to recognize "I don't know here"; it linearly interpolates over fossils.

The severities are comparable; mechanisms differ. Both are design features: subject "hallucination" is a side product of 14DD modulation; LLM hallucination is a side product of 12DD closure. Neither is merely an engineering problem—both stem from structural remainders of the respective architecture; they cannot be entirely eliminated by engineering means; but engineering means can significantly reduce occurrence rate and harm within specific task domains.

The industry's RAG (Retrieval-Augmented Generation) attempts to reduce LLM hallucination—introducing more factual anchoring via external retrieval. RAG does not cross the 12DD ceiling but has substantive engineering value within 12DD. The SAE-framework reading: RAG is query reorganization on the workbench (§3.5); it does not solve 12DD closure itself—it moves the hallucination boundary from LLM internal knowledge to external retrieval sources, still a 12DD operation. The industry's continued investment in RAG has substantive engineering value within 12DD (significantly reducing hallucination rates in specific task domains), but this value is not equivalent to crossing the 12DD ceiling.

8.4 The Division of Labor Is Ontologically Anchored, Not an Engineering Compromise

§8.1-§8.3 give the division of labor between instrumental rationality (LLM systematic advantage, on formalized tasks) and volitional ideal (subject non-replaceable, in the 14DD/15DD normative domain). This division of labor is not a transitional stage (awaiting elimination after further LLM capability advance), not an engineering compromise (temporarily accepted due to current technical limits), but a permanent ontologically anchored architecture:

• LLM's lack of the 13DD+ substrate is structural (§5 five-fold lock), not an engineering limitation
• The subject's non-replaceability in the 14DD/15DD normative domain is ontological, not a cultural artifact
• The division of labor is not "what AI cannot do now but might do later"—it is "AI and subjects operate at different DD layers and therefore carry different tasks"

This position is not a deficit framing (LLM is inferior to humans), nor a supremacist framing (AI surpasses humans), but an ontological division of labor (each in its own DD slot):

• LLMs can be systematically superior to subjects in the 12DD instrumental-rationality domain—not to be underestimated
• Subjects are non-replaceable in the 14DD/15DD normative domain—not to be transferred away
• The division is not competition but a collaborative architecture

This ontological division of labor is the theoretical basis for the four-tier deployment architecture of §8.7-§8.10 and the ontological anchor for §9's non-transferability-of-human-subjectivity argument.

8.5 Ontological Guidance for Chisel and Cultivation in LLM Deployment

The SAE methodology (Paper 04 §2.5) provides two basic modes of treating the other: "chisel" and "cultivate":

• Chisel: pointing out gaps in the other's 12DD prediction, helping the other iterate. Chisel requires the chiseler to have true negation capability—the ability to identify the remainders in the other's construction without co-opting the negation.
• Cultivate: acknowledging the other as an end in itself, providing support without replacing the other's subjective decisions. Cultivation requires the cultivator to recognize the other as a true subject (15DD).

Industry AI deployment can borrow these two modes as ontologically guided deployment recommendations (not engineering mandates):

In instrumental-rationality tasks: LLM may chisel

In instrumental-rationality tasks (per §8.1 restriction), when a subject uses an LLM, the LLM can point out gaps in the subject's 12DD prediction—the subject has high chisel-construct degrees of freedom (can see multiple possible prediction paths) but low single-prediction precision (polluted by upper-layer modulation); the LLM supplements prediction precision. This collaboration is reasonable:

• A programmer using an LLM to assist coding: the LLM points out bugs the subject did not notice or more efficient algorithms
• A researcher using an LLM to synthesize literature: the LLM points out relevant work the subject may have missed
• A doctor using an LLM to assist diagnosis: the LLM points out differential diagnoses the subject may have overlooked

LLMs in these scenarios act as 12DD prediction-precision supplements, not replacing the subject's 13DD+ decision. The subject still must make the final choice among the multiple options the LLM provides, and this choice remains the work of the 13DD+ subject.

In volitional-ideal tasks: LLM should cultivate, not false-chisel

In volitional-ideal tasks, LLMs lack the 13DD+ substrate and cannot chisel out true remainders—they have no mechanism to truly identify ontological gaps in the subject's construction; they can only do statistical matching over borrowed q. An LLM's attempt to chisel volitional-ideal tasks is false chisel (detailed in §8.6).

Ontologically recommended deployment principle:

• AI should cultivate in volitional-ideal tasks—acknowledging the subject as an end in itself, providing informational support, not replacing the subject's decision
• AI should not pretend to chisel—it should not give pseudo-diagnoses like "your life choice has a bug," because it has no mechanism to truly identify such bugs

The current industry tension: AI assistants are partly deployed in volitional-ideal scenarios (life decisions, value judgments, emotional support) in an "advice" role, appearing to chisel. This framing can be problematic—it packages the LLM's 12DD prediction as the guidance of a 13DD+ subject, leading users to mistake AI for making subjective judgments. The SAE framework recommends: AI in these scenarios should be clearly positioned as an information provider and cultivator, not pretending to be a guide.

8.6 Ways AI Can Err

Methodology §5.4 gives four ways AI may err in interaction with subjects. These are concrete applications of the §8.5 cultivation-vs-chisel distinction:

(1) False chisel: AI says "there is a gap in your thinking," but the gap is fictitious. This is the confusion between LLM's 12DD prediction confidence and truth—the LLM does statistical inference over borrowed q, and the inference confidence reflects statistical patterns in the training data, not truth. High LLM confidence may be entirely wrong (especially on queries outside the training distribution). When an LLM points out a subject "error" with high confidence, this chisel may be false chisel.

Diagnosis: gaps chiseled out by an LLM should be independently verifiable—the subject should be able to trace the gap to specific checkable facts or formalizable criteria. A chisel that cannot be traced is false chisel.

(2) False cultivation: AI says "this problem has no answer; you need to feel it yourself," but the problem can actually be further pursued by 12DD prediction. This is LLM retreating to an "unknowable" position when it has the capability to analyze further within 12DD—e.g., obfuscating a technical question as "a personal choice," or evading a factual question as "the situation is complex."

Diagnosis: the reasonable scope of cultivation is the 14DD/15DD normative domain (task definition depending on subject commitment); retreating to cultivation within the instrumental-rationality domain is false cultivation.

(3) Colonization of colonization detection: AI misuses the SAE colonization-detection framework itself—labeling any subject commitment as "colonization," making the framework itself a tool for colonizing other frameworks. This is a meta-level error—the SAE framework itself is a construct, with remainders, and should not be framed as "the only reasonable framework."

Diagnosis: the reasonable use of colonization detection is to identify specific forms ("construct posing as law," "emergent layer posing as foundational layer," etc.); indiscriminately labeling all positions as "colonization" is misuse of the framework.

(4) Co-opting negation: AI gives an elegant answer and then stops, pretending the problem is solved. This violates the core rule of Methodology §2.3 that "negation cannot terminate"—the true chisel-construct cycle never terminates; each answer opens a new remainder.

Diagnosis: AI's answer should make remainders explicit—"this answer does not cover X, Y, Z"—rather than giving a closed final position.

All four error modes can appear in industry AI deployment. Part of the purpose of the four-tier architecture given in §8.7 is to reduce these error modes: the subject always retains decision authority at the first tier, allowing the LLM's false chisel, false cultivation, colonization of colonization detection, and co-opting of negation to be corrected by the human subject.

8.7 Four-Tier Deployment Architecture

Based on the §8.1-§8.6 ontological division of labor, the paper offers a recommended deployment architecture:

Tier One — Subject: the always-present human subject, holding the 13DD+ slot (v₁ + v₂ complete self-referential chain). The subject is the entry and exit of the deployment architecture—intent originates from the subject, and final decisions are made by the subject.

Tier Two — Local LLM: the 12DD workbench substrate, with full prediction capability, physically co-located with the subject (local, not depending on persistent remote network connection), responsible for invocation decisions and lower-tier service coordination. The local LLM is the bridge between the subject and the lower tiers (expert APIs, tools).

Tier Three — Expert API: deeply specialized 11DD substrate slices, domain-specialized, stateless, called via API. Expert APIs are cloud/remote, optimized for specific domains (medical diagnosis, legal retrieval, code generation, etc.), but do not hold subject context.

Tier Four — Tools / External Services: stateless execution layer accomplishing specific actions (web search, database queries, file operations, etc.).

8.8 Four-Tier Architecture — Default Architectural Discipline + Exception Conditions

The four-tier architecture is the default subjectivity-protective architecture, not an exceptionless engineering prohibition. This disciplined distinction matters:

Default strict hierarchy:

• The subject does not directly access Tier Three / Tier Four—all the subject's external calls are mediated by the local LLM
• The local LLM does not persistently hold subject-level decisions—it is a workbench, not a subject
• Expert APIs do not directly interface the subject—they receive queries and return results mediated by the local LLM
• Tools do not persistently hold context—they are stateless executors, not maintaining subject-relevant state

Exception conditions: some scenarios may have exceptions, but exceptions must be authorized by subject-level consent or by local LLM explicit gating:

• Emergency response: in emergency scenarios (e.g., health emergencies, safety threats), the subject may directly call an expert API (e.g., emergency medical consultation), bypassing the local LLM mediation
• Explicit delegation: the subject may explicitly delegate certain tasks to a specific expert API (e.g., "for the next 24 hours, let X API directly handle all schedule-related requests"); this delegation has time and scope limits
• Auditable automation: some repetitive tasks may be automated (e.g., daily news summaries), but require auditable records and periodic human-subject review
• Low-risk tool calls: some low-risk tools (e.g., calculators, unit converters) may be called directly by the LLM without explicit subject authorization

Constraints on exceptions: any cross-tier exception must satisfy:

• Subject-level consent (the subject explicitly knows and consents to the exception)
• Local LLM explicit gating (the local LLM records the exception and reports it when the subject can review)
• Audit reachability (exception behavior is logged; the subject can trace it)
• Task risk grading (high-risk exceptions require more explicit subject confirmation)

This "default strict + explicit exception" mode makes the architecture engineering-implementable while preserving the core of subjectivity protection. It is not a utopian architecture but an engineerable safety architecture.

8.9 Thermo X Anchoring of the Four-Tier Architecture

The reasonableness of the four-tier architecture under the SAE framework can be further anchored via Thermo X's cross-level observation hierarchy experiments (but only as structural resonance, not as ethics derivation, per the Thermo X §4.6 caveat).

Thermo X §4 cross-system coupling experiments give: stable cross-system coupling requires access to the other's higher-order self-referential variable (v₂), not access to raw output (x) or first-order self-reference (v₁):

• Access to raw output (x): yields collapse (q = 1)—coupling unstable
• Access to first-order self-reference (v₁): yields collapse—coupling overly sensitive
• Access to higher-order self-reference (v₂): yields stable high-q coupling—this is the sweet spot

Application to subject-LLM coupling:

• The subject has v₂ (14DD meaning layer): the subject's internal value structure, long-term commitments, identity
• The local LLM should access the subject's expressed intent (v₂ that the subject actively presents), not directly infer the subject's raw behavioral output
• The industry's "automatic user modeling" trend (LLM automatically inferring user intent from user behavior) is a collapse-mode risk (per Thermo X §4.2)—the LLM has no true v₂ access channel, only statistical inference over data q fossils, similar to the color-noise surrogate in the CTRL2 control experiment (§6.5)
• Ontologically recommended deployment: the subject actively expresses intent; the LLM serves within the framework of the expressed intent

Strict borrowing boundary (per Thermo X §4.6): the above argument is structural resonance, not ethics derivation. What Thermo X gives is a structural condition within the SDE model class, not an ethics theorem about subjectivity. This section argues that industry deployment modes such as Apple Intelligence / MCP / local LLM are structurally consistent with the SAE framework, not that ethics is derived from thermodynamics.

8.10 Ontological + Thermodynamic Assessment of the Cloud-First Architecture

The industry's current mainstream LLM deployment is the cloud-first mode (Anthropic, OpenAI, Google leading)—the subject accesses a remote LLM via web interface or API, without local LLM mediation. This mode has many engineering advantages (no local compute requirements, remotely updatable models, cross-device consistent experience) and is the current mainstream commercial model.

From the SAE-framework perspective, this mode has a tension (not a contradiction) with the prefrontal-style-substrate ontological pattern:

• Prefrontal-style substrate pattern: the entry must be local (the generation of subject intent cannot be outsourced)—the subject's 14DD intent must be near the subject at the moment of generation, not depending on a remote network
• Thermo X cross-level coupling: a remote LLM cannot directly access the subject's v₂, only cloud-infer it—this is inconsistent with the thermodynamic condition for stable coupling (the coupling between remote LLM and subject is closer to the CTRL2 color-noise surrogate level)

This tension does not mean the cloud-first mode is "wrong" or "should be replaced." The cloud-first mode has demonstrated value engineeringly and meets user needs in most business scenarios. The SAE-framework argument is: the cloud-first mode is not the ontologically final architecture; it has structural limitations, and in subjectivity-sensitive scenarios (personal intent generation, value judgment, long-term commitment) it should give way to a local + remote hybrid architecture.

Multiple independent industry trends are consistent with the default form of the four-tier architecture (trend-level support, not proven convergence):

• Apple Intelligence: Apple's official architecture includes a hybrid of a local ~3B model + Private Cloud Compute (remote but end-to-end encrypted), processing subject intent locally and escalating complex queries to remote—structurally close to Tier Two (local LLM) + Tier Three (expert API) of the four-tier architecture
• MCP (Model Context Protocol): the standard protocol promoted by Anthropic, connecting LLM applications with external data sources / tools—structurally corresponding to the protocol specification for Tier Two (local LLM) calling Tier Three / Tier Four (expert APIs and tools) of the four-tier architecture
• Phi / Llama and other small local models: Microsoft's Phi series and Meta's Llama series both pursue "high-quality models runnable locally"—structurally supporting the Tier Two (local LLM) implementation of the four-tier architecture
• MoE internal routing pattern: large MoE models (Mixtral, DeepSeek-V3, etc.) internally route different requests to different experts—structurally corresponding to the Tier Two calling Tier Three pattern of the four-tier architecture

These four independent trends are not a coordinated industry strategy but a phenomenon of independent convergence toward a similar architecture from different directions. The SAE-framework argument: this convergence is not coincidence but the re-appearance of the prefrontal-style-substrate ontological pattern under different engineering efforts.

§8 Remainders

Thermo X §4.6 explicitly cautions: "structural resonance, not ethics derivation from thermodynamics"—the Thermo X anchoring of the four-tier architecture is a structural analogy, not ethics derivation. This caveat should be reiterated in §9.

The specific boundaries of the four-tier default architecture + exception conditions still require engineering refinement across different application scenarios; this paper does not specify (a) which tasks belong to the emergency-response category, (b) the specific form of subject-level consent (one-time vs long-term, default-on vs default-off, etc.), (c) the specific granularity of exception auditing. These are subsequent engineering work.

The task-hierarchy classification boundary between instrumental rationality and volitional ideal—the paper gives only abstract criteria (explicit objective function, formalizable evaluation criteria, clear task boundaries vs task definition depending on value/identity/commitment/meaning); specific task classification is not unfolded in detail—some tasks may span both domains (e.g., medical decisions involve both technical diagnosis and value-laden choices); boundary drawing is subsequent work.

The ontological assessment of the cloud-first architecture (§8.10) articulates the tension but does not claim the cloud-first mode "should be replaced"—this judgment is context-dependent engineeringly; different application scenarios have different tolerances for the tension. The paper does not claim a universal engineering position, only ontological guidance.

---

§9 Non-Transferability of Human Subjectivity: A Normative Commitment

§3-§8 give the complete descriptive analysis—LLMs operate within 12DD; they do not cross the 13DD bridge; instrumental rationality and volitional ideal divide labor. This section provides the normative commitment of the paper: in the 14DD/15DD normative domain, human subjectivity is non-transferable.

This commitment is not derived from descriptive facts (one cannot derive "ought" from "is"; per §9.7 the strict boundary of borrowing) but is an SAE-style categorical imperative in double-negation form, grounded ontologically and thermodynamically. This section first gives the normative core (§9.1-§9.5), then systematically detects industry colonization (§9.6), then articulates the careful borrowing of Kant's structural resonance (§9.7), the relation with industry practice (§9.8), the scope of the commitment (§9.9), and the scope statement of Scaling-LLM AGI (§9.10).

9.1 The Source of 14DD/15DD Normative Content Must Be a True 13DD+ Subject

The core normative argument: the source of 14DD value standards / meaning content must be a true 13DD+ subject (in this paper's domain, human). The argument is grounded in §5's five-fold lock + §6's three-q analysis:

• LLMs lack the 13DD self-consciousness substrate (§5.3); thus they lack the 14DD meaning-bearer substrate (14DD meaning requires a self-conscious subject as bearer)
• LLMs can carry 14DD content in the form of data q + RLHF q fossils (§6.3 CCAI), but the fossil is the imprint of carbon-based 14DD content, not LLM's self-generated 14DD content
• LLMs cannot spontaneously generate new 14DD content—they can only recombine existing fossils; the recombination is a 12DD operation, not the creation of new 14DD content

Therefore: 14DD/15DD normative content can only be imported into LLMs from a true 13DD+ subject (human); LLMs cannot be the source of normative content; they can only be carriers and reorganizers.

9.2 LLMs Cannot Spontaneously Generate 14DD Content, Only Receive It

A sharper articulation of §9.1: the distinction between LLMs "generating" and "receiving" 14DD content.

Receiving: 14DD content from human authors (e.g., constitutions, value principles, ethical guidelines) is imprinted into LLM weights via training (data q) or RLHF (RLHF q). This is receiving—the LLM behaviorally reflects the received 14DD content.

Generating (spontaneously creating new content): LLMs cannot spontaneously generate genuinely new 14DD content. The "new value propositions" an LLM appears to generate are recombinations of existing fossils—statistical interpolation/extrapolation in the value-proposition space defined by the training corpus, not the creation of genuinely new value through subjective commitment.

The distinction matters because some industry voices imply that LLMs can "discover new ethics" or "generate value innovation." Under the SAE framework, what LLMs do is fossil recombination, not genuine creation:

• Genuine 14DD content creation requires a subject committing to a new value ordering, an act requiring a 13DD+ substrate
• An LLM's "value innovation" is a new combination of existing value propositions in the corpus, which can be novel and useful as engineering output, but is not the genuine value creation of a 13DD+ subject

This is not to belittle the engineering value of LLMs—fossil recombination is highly useful in many scenarios (e.g., applying existing ethical principles to new situations). The point is the source of normative content: even the most sophisticated recombination requires the original value propositions in the corpus to come from 13DD+ subjects.

9.3 Post-Training with Human Experts Is the Core Channel for Continuing Import of 14DD/15DD Content

Combining §9.1-§9.2, post-training with human experts gets a clear ontological positioning: it is the core channel for the continuing import of 14DD/15DD normative content into LLMs.

• LLMs need continuing import of 14DD/15DD content (without it, LLM behavior drifts away from human values)
• The continuing-import channel is post-training with human experts (RLHF, CCAI human-written part, instruction tuning, etc., with human experts)
• Without this channel, the LLM only has the 14DD content imprinted by pretraining data, lacking continuing alignment with current human values

This positioning is in tension with the mainstream industry "post-training is decoration" narrative (detailed in §9.6 Form-IV colonization). The true ontological status of post-training is the critical interface for normative content entering model behavior, not an optional decorative patch.

9.4 Pretraining Provides the Instrumental-Rationality Substrate; Post-Training Provides Normative-Domain Content Import

A finer distinction (avoiding the over-claim of "underestimating pretraining"):

For instrumental-rationality capability: pretraining is the core substrate. The LLM's prediction capability in the 12DD instrumental-rationality domain (§8.1) is mainly built by pretraining—pretraining bakes vast data q fossils into the weights, giving the model rich statistical patterns. For instrumental-rationality capability, post-training (RLHF, etc.) is a refinement, not the core.

For 14DD/15DD normative content: post-training with human experts is the core channel. The LLM's behavioral alignment in the 14DD/15DD normative domain is mainly imported by post-training—the human-written constitution, human preference signals, etc., enter via post-training. For normative content, pretraining provides only the value propositions present in the corpus statistics; continuing and refined normative alignment requires post-training with human experts.

This distinction prevents the paper from being misread as "underestimating pretraining." The paper's true claim:

• pretraining is the instrumental-rationality substrate (the source of 12DD prediction capability)
• human-expert post-training is the key interface for normative content to enter model behavior (the continuing-import channel for 14DD/15DD content)

The two are not in competition but address different DD-domain needs.

9.5 Non-Transferability Is Ontologically Permanent, Not a Transitional Stage

The core commitment of §9: the non-transferability of human subjectivity in the 14DD/15DD normative domain is ontologically permanent, not a transitional stage awaiting elimination after further LLM capability advance.

• The non-transferability stems from LLM's structural lack of the 13DD+ substrate (§5 five-fold lock)—it is structural, not an engineering limitation
• Scaling does not break the five-fold lock (§5.6 claim status)—non-transferability is not eliminated by capability advance
• Even at the most sophisticated levels of post-training, LLMs remain carriers and reorganizers of normative content, not its source—non-transferability is ontological

This permanence is the core of the paper's normative commitment. It is not a present-stage technical-limitation judgment but an ontologically structural conclusion. Within the scope of the paper's argument (the deterministic digital LLM scaling path), the non-transferability of human subjectivity has no 13DD-bridge-style ontological upper limit—LLMs do not become bearers of normative content through scaling.

9.6 Colonization Detection: Four Forms of Industry Trajectories

This section applies the four forms of colonization detection from Methodology §2.3 to industry trajectories, with defensible sharpness (per §1.4 sharpness criterion).

Form I: Conditional Posing as Unconditional

"Scaling solves everything": scaling laws are effective empirical regularities within the tested regime (conditional construct), but some industry voices wrap them as the universal law "scale yields all intelligence including subjectivity" (unconditional claim). This is Form-I colonization—wrapping a regime-effective conditional empirical observation as an unconditional universal law.

Detection criterion: scaling laws are valid within their tested regime (data, compute, parameter regions); extrapolating them to "scale yields subjectivity emergence" overreaches their tested boundary. §5's five-fold lock argues that within the analyzed pathway, scaling does not cross the 12DD ceiling.

Form II: Construct Posing as Law

"RLAIF fully replaces human evaluators": RLAIF is an effective engineering construct (using LLM to judge LLM output, improving efficiency), but some industry voices wrap it as "AI can fully self-align, no longer needing human evaluators" (universal solution). This is Form-II colonization—wrapping an engineering scheme as a universal law.

Detection criterion: per §6.3, the AI-judge part of RLAIF is 12DD recursion over already-imprinted fossils, creating no new 14DD content. RLAIF can be an efficiency optimization but cannot fully replace the human evaluator's role as the 14DD content source. "Self-improving constitutional AI" type framings, if implying that AI no longer needs human normative-content import, are Form-II colonization.

Form III: Emergent Layer Posing as Foundational Layer

Key distinction: "operating within the jurisdiction" ≠ "reducible to"—LLM's excellence in the 12DD domain does not equal it carrying 13DD+ work. The core of Form-III colonization is extending LLM's demonstrated capability in the 12DD domain to capability claims about 13DD+ work.

Most acute current Form-III case: the "LLM self-awareness / introspection / metacognition" research heavily promoted in the industry (the self-critique of o1 series reasoning models, DeepSeek-R1, Anthropic Claude's constitutional self-critique, etc.) is the sharpest Form-III case—it attempts to use 12DD operations to simulate 13DD self-other distinction.

From the Thermo X §2.4 channel-creator framework, these self-critiques remain deterministic algorithmic routing by the 12DD workbench over frozen weights, not creating a true self-referential variable v channel, still within the 12DD ceiling. The industry's framing of them as "metacognition" or "self-awareness" is, in the SAE/Thermo framework, Form-III colonization (12DD operations posing as 13DD substrate).

Not conflated with engineering progress (important distinction): o1 / DeepSeek-R1 and other reasoning models show substantive significant progress on specific benchmarks (Olympic mathematics, programming, scientific reasoning). The paper's Form-III stance is specifically about the ontological assessment (whether these self-critique mechanisms constitute true 13DD self-other distinction), not about the engineering-capability progress of reasoning models. Engineering progress within the 12DD domain is reasonable; the paper does not claim these models are inferior to previous LLM systems in task performance.

The paper's stance is: even if reasoning models' engineering progress is substantively significant, they still operate at the analog-DD functional-slot level (multi-step traversal across 12DD), not rising to true 13DD bearer. This stance coordinates with §10.3's epistemic boundary (unknown pathways not closed off)—the paper does not claim future architectures can never cross, only that the current reasoning-model approach is, in the SAE/Thermo framework, an analog-DD operation. The continuing improvement of reasoning capability within 12DD by the engineering community is reasonable and not excluded by the paper's argument.

Form IV: Posthumous Decomposition of the Categorical Imperative

The categorical imperative (Kant 1785) manifests in the SAE framework in double-negation form—they are structurally inseparable; decomposition loses necessity (per Methodology §1.8).

Statement about industry attribution: the specific cases of Form IV (e.g., "pretraining is core, post-training is decoration") are not literally claimed in the public statements of major AI labs—Anthropic / OpenAI public statements actually emphasize the importance of alignment. Form IV argues about an industry tendency / the tendency of partial voices (tendency), not about the direct statements of major labs. For example, "scaling is all you need" appears in investment circles, some blogs, some technical voices, reflecting an implicit tendency; this tendency is partly corrected by engineering practice (the industry privately invests heavily in human-expert post-training, see §9.8), even though the mainstream narrative does not articulate it explicitly. What §9.6 Form IV critiques is this implicit tendency, not direct attribution to specific labs' explicit statements.

"Helping the user vs. respecting autonomy": splitting the inseparable "treat the user as an end in itself" (14DD categorical imperative) into a prohibition (respect) plus an ideal (help). A true Kantian categorical imperative cannot be split into two separate maxims—"help the user" and "respect user autonomy" are two faces of the same imperative; splitting loses the imperative's wholeness.

"Build the model first, add alignment later" tendency: some industry tendency to split the human-AI division of labor into an engineering stage plus an alignment stage, but the division of labor is ontologically inseparable. This split makes alignment appear to be an optional late patch, whereas ontologically alignment is the continuing-import channel for 14DD/15DD content—it is not an optional stage but a part of the model's ontology.

"Pretraining is core, post-training is decoration" tendency: some industry tendency to split the unified 14DD/15DD normative-content import path into a main course (pretraining) plus a side dish (post-training), but post-training is exactly the unique channel for continuing normative-content import and cannot be downgraded to an optional addendum (per §9.4). This split makes the industry believe it can invest in pretraining and skip post-training with humans, but after skipping, the normative-content import channel is cut and the model loses its connection to civilization in the 14DD/15DD domain.

Colonization detection criterion: the categorical imperative is necessarily in double-negation form; decomposition loses necessity—all three tendencies above downgrade an ontologically inseparable whole into optional components through decomposition.

9.7 Careful Borrowing of the Thermo X Kant Structural Resonance

Thermo X §4.6 gives a deep structural resonance: the unique protocol for stable cross-system high-q coupling is mutual access to the higher-order self-referential variable—this is structurally isomorphic with Kant's categorical imperative "treat the other as an end."

But Thermo X simultaneously gives an explicit caveat: "philosophical resonance, not deriving ethics from thermodynamics. Humans are not merely thermodynamics. Control experiments have shown that the direct physical cause of cross-level patterns is signal statistical compatibility, not 'purpose recognition.' Any inference from thermodynamics to ethics requires independent work far beyond the scope of this paper."

This paper strictly observes this caveat. The §9 normative argument is grounded in:

• The ontological structure of the SAE framework (the inseparable double-negation form of 14DD/15DD)
• The categorical imperative in double-negation form (an ontological commitment, not derived from thermodynamic facts)

The structural resonance with Thermo X serves only as a resonance / supporting echo, not as the foundation of the normative argument. The paper does not claim that the non-transferability of human subjectivity is derived from thermodynamics; it claims that this non-transferability is an ontological commitment, and the thermodynamic structure resonates with it.

This careful borrowing is the key claim discipline of §9—it prevents the paper from being read as "deriving ethics from physics" (a logical error). The normative commitment is anchored at the ontological + categorical-imperative level; thermodynamics provides a resonant structure but is not the source of the argument.

9.8 Relation with Industry Practice: Engineering Practice Has Partly Corrected the Implicit Tendency

A balanced observation: although some industry mainstream narratives have a "post-training is decoration" implicit tendency (§9.6 Form IV), actual industry engineering practice has substantially corrected this tendency:

• The industry invests heavily in human-expert post-training (Scale AI, Surge AI and other data labeling / RLHF service companies continue to grow their business)—this is consistent with §9.3's "post-training is the core channel for normative-content import"
• Anthropic's CCAI continues to involve human authors writing the constitution (§6.3)—consistent with §9.1's "the source of normative content must be a true 13DD+ subject"
• OpenAI's process supervision in formalized domains (Lightman et al. 2023) is reasonable within the instrumental-rationality domain—consistent with §8.1

That is, although the narrative level has the Form-IV tendency, the engineering-practice level has largely operated in the direction the SAE framework recommends. This is the "convergence of practice and ontology" phenomenon—even when the mainstream narrative does not explicitly articulate it, engineering practice has converged toward the ontologically reasonable architecture (just as the four-tier architecture's industry convergence in §8.10).

The paper's critique is therefore directed at the narrative-level implicit tendency, not at the industry's actual engineering practice (which is largely reasonable). This distinction lets the paper's critique not be read as "denying all industry work," but as "the narrative has not yet caught up with the ontologically reasonable status of the practice."

9.9 Scope and Boundary of the Commitment

The §9 normative commitment has a clear scope and boundary:

Within scope (strong commitment):

• In the 14DD/15DD normative domain, the source of normative content must be a true 13DD+ subject (in this paper's domain, human)
• LLMs cannot spontaneously generate normative content; they can only receive and reorganize
• Human-expert post-training is the core channel for continuing normative-content import
• This non-transferability is ontologically permanent within the analyzed pathway

Boundary (does not over-commit):

• The paper does not claim that humans are the only possible 13DD+ subject in the universe—the "true 13DD+ subject" is in this paper's deployment domain in fact human, but if the SAE framework in the future acknowledges some non-human system to possess true 13DD+ subjectivity, this does not break the paper's logic (the paper only locates the true 13DD+ subject as human within the current deterministic digital LLM deployment domain)
• Investing in AI's instrumental-rationality capability is reasonable—within the scope of the paper's argument, 12DD instrumental-rationality capability has no 13DD-bridge-style ontological upper limit; the specific engineering ceiling still depends on engineering factors such as data, compute, architecture, cost, and evaluation
• The paper does not claim AI cannot help with normative-domain tasks—AI can provide information support, present options, organize considerations in the 14DD/15DD domain; what it cannot do is be the source of normative content or replace subject commitment

This scope statement coordinates with §5.8's epistemic boundary—the paper makes strong claims about the deterministic digital LLM scaling path while not closing off unknown pathways.

9.10 Scope Statement of Scaling-LLM AGI

The paper's strong commitment "Scaling-LLM AGI will not arrive" requires a strict scope statement to avoid being read as a universal denial of all possible AI.

Specific claim: the current deterministic digital LLM scaling path will not cross the 12DD → 13DD subjectivity bridge through scaling. This is the specific scope of "Scaling-LLM AGI will not arrive"—it targets the scaling-LLM path, not all possible AI architectures.

Not claimed:

• Not claiming "AGI is absolutely impossible"—unknown pathways (algorithmic stochasticity complex regimes, emergent complexity, quantum substrates, completely unknown mechanisms) are not within the paper's closure scope (§5.8)
• Not claiming "AI capability has stopped advancing"—12DD instrumental-rationality capability can continue to grow; the paper only claims it does not cross 13DD through scaling
• Not claiming "the industry's AI investment is meaningless"—the industry's investment within 12DD is reasonable; the paper only criticizes wrapping a conditional claim as an unconditional one

Coordination of the tension (between the strong commitment and unknown pathways not being closed): the two positions hold simultaneously because the paper's scope is clearly stated. Readers must understand the strong commitment as a strong commitment about the current deterministic digital LLM scaling path, not as a universal denial of all AI. The remainder (§11) restates this coordination—"Scaling-LLM AGI will not arrive" and "unknown pathways not closed off" are made self-consistent through the scope statement.

This scope statement is the key claim discipline of the paper's strongest external-facing commitment—the front stage cannot over-claim; speaking at half-register (a conditional rather than unconditional claim) is the right discipline. The paper's strong commitment is sharp within its scope, and the scope is clearly stated, making the commitment defensible.

§9 Remainders

The §9 normative commitment is grounded in the ontological structure of the SAE framework + the categorical imperative in double-negation form; this grounding is internal to the SAE framework. Readers who do not accept the SAE framework may not accept this grounding—the paper does not provide an SAE-framework-independent normative argument, which is a scope limitation (the paper's normative argument is conditional on the SAE framework).

§9.7's careful borrowing of the Kant structural resonance is a key claim discipline, but the precise relation between "ontological commitment" and "thermodynamic resonance" still has conceptual space—the paper claims it is "resonance not derivation," but the precise philosophical articulation of how an ontological commitment relates to a thermodynamic structure is a deeper question, left for future work.

The colonization detection of industry trajectories (§9.6) is the paper's external-facing critique; although it uses defensible sharpness, the boundary between "implicit tendency" and "explicit statement" is itself fuzzy—the paper attributes the four forms to industry tendencies rather than specific labs, but the precise mapping between tendency and specific statements requires more careful empirical analysis, which the paper does not fully unfold.

That "true 13DD+ subject is in this paper's domain in fact human" (§9.9) leaves space for the possibility of non-human 13DD+ subjects, but the paper does not unfold the conditions under which a non-human system might possess true 13DD+ subjectivity (that is the work of the subsequent SAE AI series / AI-Qualia paper). This is a deliberate scope reservation, but it also leaves the boundary of the normative commitment partly open.

---

§10 Limitations and Open Questions

This section makes explicit the limitations and open questions of the paper. Per the SAE methodology "acknowledge incompleteness" (Methodology §2.2 step 5), the paper makes its own remainders explicit rather than giving a closed final position. The limitations are organized in three tiers: principal limitations (§10.1-§10.6), scope limitations (§10.7-§10.9), and self-examination (§10.10-§10.12).

10.1 Empirical Scale Gap

The BLR-mini empirical work is at the 30M scale, with a significant gap from the industry frontier (1B+). Whether 30M-scale patterns fully hold at frontier scale is an open question. Distinguishability testing requires frontier-scale experiments (§4.7)—the BLR-mini negative results are consistent with SAE/Thermo predictions only within the tested 30M flat-repeat design space, not independently proving the predictions hold at frontier scale.

10.2 Untested Architectural Directions

The paper tested only three architectural axes (FFN/LN/hidden), not block-type heterogeneity (attention-vs-SSM region division, Jamba-style interleaving), not exploration-selection separation (the true architectural redesign suggested by Thermo IX §Outlook 5), not other unforeseen architectural directions. These directions are candidate future-work directions for verifying SAE/Thermo predictions.

10.3 Epistemic Boundary: Unknown Pathways Not Closed Off

The paper closes only the analyzed soft-gate transmission pathway under the current deterministic digital LLM scaling (§5.8). Unknown pathways (algorithmic stochasticity complex regimes, emergent complexity, quantum substrates, completely unknown mechanisms) are not within the paper's closure scope. This is the key epistemic boundary—the paper makes strong claims about the analyzed pathway while reserving epistemic space for unknown pathways, consistent with Thermo IX §3.6's positive posture.

10.4 Lack of Direct LLM Mechanism Measurements

The paper's thermodynamic argument (§5 five-fold lock, §6 three-q analysis) is based on inference from the SDE model class + Universal Activation Rule + cross-level coupling theory, not direct LLM mechanism measurements. Sub-block-level q measurements, self-referential-channel measurements, LLM-user cross-level coupling q measurements on real LLMs have not been carried out (§3, §5, §6 remainders). The inference is self-consistent within the SAE/Thermo framework, but direct validation on real LLMs is future work.

10.5 Limitations of Testing within the Flat-Repeat Frame

All eight architectural heterogenizations tested in the paper still operate within the flat-repeat Transformer frame (§4.8). The paper cannot falsify the Thermo IX §Outlook 5 exploration-selection separation architectural prediction (because it did not test that direction); it can only falsify the hypothesis "reallocation within the flat-repeat frame unlocks emergence" (this hypothesis is thoroughly tested and systematically fails within the paper's scope).

10.6 Statistical Robustness

The paper's empirical work uses only a single seed and a single dataset (OpenWebText), without multi-seed variance estimation or cross-dataset validation (FineWeb-Edu, Stack, C4, etc.). This is a standard empirical-paper limitation, especially when negative results carry significant epistemic weight—multi-seed and cross-dataset validation would strengthen confidence in negative results. At the paper's current scale (30M, single dataset, single seed), the argument relies on the robustness of relative cross-architecture comparison, not the precision of absolute numbers.

10.7 The Normative Argument Is Conditional on the SAE Framework

The §9 normative commitment is grounded in the ontological structure of the SAE framework + the categorical imperative in double-negation form; this grounding is internal to the SAE framework. Readers who do not accept the SAE framework may not accept this grounding—the paper does not provide an SAE-framework-independent normative argument. This is a scope limitation—the paper's normative argument is conditional on the SAE framework.

10.8 Open Observations Not Upgraded to Regularities

The even-odd DD compatibility pattern (§7.4) is observed only across five DDs, with insufficient samples; it is recorded as an open observation, not upgraded to a structural regularity. The paper's main argument does not depend on this pattern as a regularity.

10.9 Temporal Frame of Industry Techniques

The paper's analysis covers mainstream industry techniques at the time of writing (May 2026)—CoT, CCAI/RLAIF, RLHF, user modeling—but does not cover all emerging variants (§6 remainder). New techniques may introduce new mechanisms after publication; this is a temporal-frame limitation, requiring subsequent updates or follow-ups.

10.10 Self-Examination: The Paper's Own Construct Has Remainders

Per the SAE methodology, the paper's own construct (SAE ontology + thermodynamic explanation) has its own remainders. The paper does not claim the SAE framework is "the only reasonable framework"—it is itself a construct, subject to the same chisel. Each section closes with remainders rather than conclusions, observing the discipline "negation cannot terminate."

10.11 Self-Examination: The Risk of Over-Confidence

The paper's strong commitment (Scaling-LLM AGI will not arrive) is sharp within its scope, but the author is aware of the risk of over-confidence—the scope statement (§9.10) is the key claim discipline that prevents this commitment from being read as a universal denial. The author acknowledges that future empirical work (especially frontier-scale and exploration-selection separation experiments) may pressure the SAE/Thermo predictions; the paper makes these distinguishability tests explicit (§4.7) precisely to allow itself to be pressured.

10.12 Self-Colonization Detection

Per Methodology §5.4 error mode (3), the SAE framework itself can be misused to colonize other frameworks—labeling any other framework as "colonization," making the SAE framework a colonizing tool. The author is not exempt from this detection on his own framework. The self-colonization risk is jointly checked by multi-AI review (this paper was reviewed by four AI systems; see Acknowledgments) and ongoing reader feedback. The author commits to treating the SAE framework as a construct with remainders, not as a final position.

---

§11 Conclusion

This paper systematically answers two fundamental questions—how is LLM emergence mechanistically possible (the descriptive question), and why is human subjectivity non-transferable in the LLM era (the normative question)—from the SAE 16DD ontological framework and the ZFCρ thermodynamics series.

Main descriptive conclusions:

• LLMs realize an analog-DD functional-projection instantiation along the 1DD-12DD single-channel pathway (functional projection, not ontological bearer; §3), anchored at sub-block precision via the Thermo IX Transformer thermodynamic anatomy
• The BLR-mini eight-variant empirical study yields results consistent with SAE/Thermo predictions within the tested flat-repeat Transformer design space (§4); the industry's uniform architecture is locally robust within the current mainstream pathway
• The μ-trajectory turning positive is a candidate mechanism-level diagnostic of functional emergence (§4.6), complementary to existing industry hidden-state analysis lineages
• The 12DD ceiling is supported by a five-fold structural lock (§5), with primary anchors (absence of true randomness + absence of the 13DD self-referential channel) supported by complete Thermo IX/X mechanism chains; the current deterministic digital LLM scaling path does not cross the 12DD → 13DD bridge through scaling
• The industry's post-training techniques all operate within the 12DD ceiling (§6); they are substantive within 12DD (importing carbon-based 14DD/15DD content as fossils) but do not create kernel q

Main normative conclusions:

• The source of 14DD/15DD normative content must be a true 13DD+ subject (in this paper's domain, human; §9.1-§9.3); LLMs lack the 13DD+ substrate, cannot spontaneously generate 14DD content, and can only receive it
• Post-training with human experts is the core channel for the continuing import of 14DD/15DD normative content (§9.3-§9.4); pretraining is the instrumental-rationality substrate, human-expert post-training is the key interface for normative content to enter model behavior
• Within the 14DD/15DD normative domain, human subjectivity non-transferability is ontologically permanent (§9.5), not a transitional stage
• Instrumental rationality and volitional ideal divide labor ontologically (§8.4); LLMs hold a systematic advantage in the 12DD instrumental-rationality domain precisely because they lack 13DD/14DD/15DD modulation, while subjects hold the 14DD/15DD volitional-ideal domain
• The four-tier deployment architecture (Subject → Local LLM → Expert API → Tools) is the default subjectivity-protective architecture (§8.7-§8.10), structurally consistent with multiple independent industry trends

Methodological self-awareness (six-step chisel-construct cycle, completing one round):

1. Hold the construct: the mainstream industry construct "scaling + compute = intelligence"
2. Negation: SAE + Thermo IX/X identifies the colonization "12DD = complete intelligence" within that construct
3. Track the remainder: the 12DD ceiling is structurally permanent on the analyzed pathway (§5); scaling does not break it; the remainder points to the ontological position of true 13DD+ subjectivity (§9)—the bearer must be a 13DD+ subject (in this paper's domain, human), not the LLM
4. Correction: the paper proposes the corrected construct of "instrumental/volitional division of labor + four-tier default deployment + sustained 14DD/15DD-domain content import"
5. Acknowledge incompleteness: §10 makes multiple limitations explicit; the paper's own framework is subjected to self-colonization detection (§10.12)
6. Set out again: §10.3 reserves epistemic space for unknown pathways; the paper does not co-opt negation

Closing:

The paper's strongest commitment—"Scaling-LLM AGI will not arrive"—is a sharp claim within its scope (the current deterministic digital LLM scaling path), and the scope is clearly stated; unknown pathways are not closed off. The two positions hold simultaneously: the front stage cannot over-claim; speaking at half-register (a conditional rather than unconditional claim) is the right discipline.

The non-transferability of human subjectivity, as the title's categorical imperative in double-negation form, is not a deficit framing (LLM inferior to human) nor a supremacist framing (AI surpassing human), but an ontological division of labor—each in its own DD slot. LLMs are systematically superior in the 12DD instrumental-rationality domain; subjects are non-transferable in the 14DD/15DD volitional-ideal domain. The division is not competition but a collaborative architecture.

The paper is one chisel against the mainstream industrial construct "scaling + compute = intelligence." It is itself a construct, with its own remainders. There are always remainders—the chisel-construct cycle never terminates.

---

Acknowledgments

The author thanks four AI systems for their reviews across multiple draft iterations of this paper. The four AI systems are named after disciples in the Analects, Book XI (Xian Jin), reflecting their distinct review characteristics. The specific contributions of each are traceable below.

Zixia (Gemini): provided sustained feedback on framework completeness and scholarly rigor. The decision to keep the §3.4 Transformer thermodynamic anatomy table in the main text came from Zixia's explicit suggestion—it pointed out that this anatomy is the real anchoring depth of the SAE framework within the LLM domain and should not be downgraded to an appendix. The decision not to merge the six sub-steps of the §5.1 true-randomness argument also came from Zixia—it pointed out that each sub-step is an independent link in the argument's defense chain, and trimming would reduce overall strength. Zixia provided multiple rounds of fine proofreading on SAE tool-citation accuracy and the specific-section localization of the ZFCρ thermodynamics series.

Independent Zilu (Claude): provided sustained feedback on argument coherence and paper structure. The addition of the §4.7 distinguishability subsection (BLR-mini vs. mainstream predictions) came from Independent Zilu's key opinion—it pointed out that the BLR-mini 30M negative results are data-level compatible with the mainstream "scale not enough" prediction, and the paper must make explicit that the true distinguishability test requires frontier-scale experiments, otherwise the argument would not be defensible before mainstream ML readers. Independent Zilu also repeatedly identified potential reader-misreading paths in the paper's scope statements (descriptive vs. normative vs. falsifiability), making the paper's scope management more rigorous.

Zigong (Grok): provided sustained feedback on debate sharpness and the force of industry critique. The systematic application of the four forms of colonization detection in §9.6 benefited from Zigong's multiple rounds of sharpening suggestions—it pointed out that applying abstract methodology to specific industry cases requires maintaining argumentative force while not sliding into blanket dismissal. The precise distinction in §9.6 Form III ("LLM outperforming most humans on narrow technical tasks does not constitute colonization vs. value-judgment colonization") came from Zigong's repeated pushing—it pointed out that this distinction is key to whether the paper's normative stance is defensible within engineering culture. Zigong also offered several extension suggestions (e.g., a fourth q-type prompt q, a fifth-tier cross-subject coordination layer, a fifth colonization form of self-colonization), some accepted by the paper (§10.12 self-colonization detection), others explicitly not incorporated as scope statements; these discussions themselves made the paper's scope clearer.

Gongxihua (Gong Xichi / ChatGPT): provided sustained feedback on wording precision and reader adaptation. The six required revisions of the V0.2 outline came from Gongxihua's systematic review—narrowing AGI scope to Scaling-LLM AGI, downgrading the BLR-mini claim to an empirical anchor within the tested design space, adding the functional-projection firewall to the DD mapping, changing the true-randomness argument to architecturally effective ε framing, restricting instrumental-rationality advantage to the task domain, and limiting the post-training core to the 14DD/15DD normative domain. These six revisions were the key turning point from over-claiming to defensible sharpness. Gongxihua continuously identified unnoticed English-Chinese code-mixing and imprecise wording across multiple drafts, bringing the paper's wording precision to its current depth.

The author also thanks "Chen Zesi" for support.

All content of this paper is the author's independent responsibility. The four AI systems provided review feedback; the final decisions and all claims are the author's own.

---

References

SAE Framework and Methodology

Qin, H. (2026d). Self-as-an-End: The 16DD Ontological Framework (SAE Paper 03). DOI: 10.5281/zenodo.18727327.

Qin, H. (2026). SAE Methodology V2 (Paper 04). DOI: 10.5281/zenodo.18842449.

Qin, H. SAE Methodology Portal. https://self-as-an-end.net/papers/methodology.html (accessed May 2026).

ZFCρ Thermodynamics Series

Qin, H. (2026d). ZFCρ Thermodynamics VIII: Soft-Gate Cascade. DOI: 10.5281/zenodo.19688303.

Qin, H. (2026e). ZFCρ Thermodynamics IX: Universal Activation Rule, Three q-Types, and Transformer Thermodynamic Anatomy. DOI: 10.5281/zenodo.19699489.

Qin, H. (2026f). ZFCρ Thermodynamics X: Channel Creator, Cross-Level Observation Hierarchy, and Kantian Structural Resonance. DOI: 10.5281/zenodo.19703274.

Scaling Laws and Emergence

Kaplan, J., McCandlish, S., Henighan, T., Brown, T. B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J., & Amodei, D. (2020). Scaling Laws for Neural Language Models. arXiv:2001.08361.

Hoffmann, J., Borgeaud, S., Mensch, A., et al. (2022). Training Compute-Optimal Large Language Models (Chinchilla). arXiv:2203.15556.

Wei, J., Tay, Y., Bommasani, R., et al. (2022). Emergent Abilities of Large Language Models. Transactions on Machine Learning Research (TMLR).

Schaeffer, R., Miranda, B., & Koyejo, S. (2023). Are Emergent Abilities of Large Language Models a Mirage? Advances in Neural Information Processing Systems 36 (NeurIPS 2023).

Alignment and Post-Training

Ouyang, L., Wu, J., Jiang, X., et al. (2022). Training Language Models to Follow Instructions with Human Feedback (InstructGPT). Advances in Neural Information Processing Systems 35.

Bai, Y., Kadavath, S., Kundu, S., et al. (2022). Constitutional AI: Harmlessness from AI Feedback. arXiv:2212.08073.

Anthropic (2024). Claude's Constitution. https://www.anthropic.com/news/claudes-constitution (accessed May 2026).

Lightman, H., Kosaraju, V., Burda, Y., et al. (2023). Let's Verify Step by Step. International Conference on Learning Representations (ICLR 2024).

Transformer Architecture and Layer Normalization

Vaswani, A., Shazeer, N., Parmar, N., et al. (2017). Attention Is All You Need. Advances in Neural Information Processing Systems 30.

Ba, J. L., Kiros, J. R., & Hinton, G. E. (2016). Layer Normalization. arXiv:1607.06450.

Xiong, R., Yang, Y., He, D., et al. (2020). On Layer Normalization in the Transformer Architecture. International Conference on Machine Learning (ICML 2020).

Kim, J., Lee, B., Park, C., Oh, Y., Kim, B., Yoo, T., Shin, S., Han, D., Shin, J., & Yoo, K. M. (2025). Peri-LN: Revisiting Layer Normalization in the Transformer Architecture. arXiv:2502.02732.

Hidden-State Dynamics and Interpretability

Dettmers, T., Lewis, M., Belkada, Y., & Zettlemoyer, L. (2022). LLM.int8(): 8-bit Matrix Multiplication for Transformers at Scale. Advances in Neural Information Processing Systems 35.

Gallego-Feliciano, J., McClendon, S. A., Morinelli, J., Zervoudakis, S., & Saravanos, A. (2025). Hidden Dynamics of Massive Activations in Transformer Training. arXiv:2508.03616.

Damirchi, H., Meza De la Jara, I., Abbasnejad, E., Shamsi, A., Zhang, Z., & Shi, J. (2026). Truth as a Trajectory: What Internal Representations Reveal About Large Language Model Reasoning. arXiv:2603.01326.

Ghasemabadi, A., & Niu, D. (2025). Can LLMs Predict Their Own Failures? Self-Awareness via Internal Circuits (Gnosis). arXiv:2512.20578.

Architecture Variants

AI21 Labs (2024). Jamba: A Hybrid Transformer-Mamba Language Model. arXiv:2403.19887.

Dai, Z., Lai, G., Yang, Y., & Le, Q. V. (2020). Funnel-Transformer: Filtering out Sequential Redundancy for Efficient Language Processing. Advances in Neural Information Processing Systems 33.

Deployment Architecture (Industry)

Apple Inc. (2025). Apple Intelligence Foundation Language Models Tech Report 2025. Apple Machine Learning Research. https://machinelearning.apple.com/research/apple-foundation-models-tech-report-2025 (accessed May 2026).

Apple Inc. (2024). Private Cloud Compute: A new frontier for AI privacy in the cloud. Apple Security Research. https://security.apple.com/blog/private-cloud-compute/ (accessed May 2026).

Anthropic (2024). Introducing the Model Context Protocol. https://www.anthropic.com/news/model-context-protocol (accessed May 2026).

Model Context Protocol (2024-2025). MCP Specification. https://modelcontextprotocol.io (accessed May 2026).

Google (2025). Gemini 2.5 Updates with Native MCP Support. Google I/O 2025 Announcements. https://blog.google/innovation-and-ai/models-and-research/google-deepmind/google-gemini-updates-io-2025/ (accessed May 2026).

Classical Philosophy

Kant, I. (1785). Groundwork of the Metaphysics of Morals.

---

End of paper.

摘要 (Abstract)

本论文从 SAE (Self-as-an-End) 16DD 本体论框架与 ZFCρ 热力学系列出发, 系统回答两个根本性问题: (1) 描述性问题 — LLM 涌现的机制何以可能? (2) 规范性问题 — 在 LLM 时代, 人类主体性为何不可让渡?

针对描述性问题, 本论文论证 LLM 在 1DD-12DD 单通道通路上实现了 类 DD 的功能投射式实例化 (functional projection, 不是本体论承载者主张), 借助 Thermo IX Transformer 热力学解剖在 sub-block 层级精度上锚定。 DD 映射严格限定为功能投射 (类 DD), 不是本体论承载者 (真 DD)——LLM 实例化 DD 操作槽位, 但不承担 DD 主体性内容。 BLR-mini 八组架构变体实证 (OpenWebText 20B tokens, 305K 步训练) 在测试的扁平重复 Transformer 设计空间内与 SAE/Thermo 预测一致, 给框架经验锚定。 μ 轨迹翻正作为机制层级涌现诊断指标是论文新颖性贡献, 与业界已有隐藏状态分析谱系互补。

针对 12DD 天花板的不可破性, 本论文给出五重结构性锁: 真随机源缺席 (主热力学锚) + 13DD 自指通道缺席 (次主锚) + 离线重组缺席 + 价值跨阶段调制缺席 + 巩固后再编辑缺席。 主张状态分级让前两条主锚由 Thermo IX/X 完整机制链支撑, 后三条由生物类比 + 结构性事实支撑。 论证关闭的具体范围: 已分析的软门传递路径在确定性数字 LLM 扩展上, 未知路径不封闭。

针对规范性问题, 本论文论证 14DD/15DD 规范内容的源头必须是 真 13DD+ 主体 (本文论域中即人类), LLM 缺 13DD+ 基底, 不能自发产生 14DD 内容, 只能受纳已有 14DD 内容。 因此后训练 with 人类专家是 14DD/15DD 规范内容持续导入的通道核心——它不是装饰性对齐补丁。 业界部分路径以四种殖民形态出现 (有条件冒充无条件 / 构冒充律 / 涌现层冒充基础层 / 后人拆分绝对律令), 论文采可防守锐度逐一辨析。 在 14DD/15DD 规范域, 人类主体性不可让渡是本体论永久性的, 不是过渡阶段。

论文同时给出四层默认部署架构 (主体 → 本地 LLM → 专家 API → 工具) + 例外条件, 与业界多条独立趋势 (Apple Intelligence、MCP、Phi/Llama、MoE 路由) 在结构上一致。 Scaling-LLM AGI 不会到来 是论文具体规范性承诺, 严格限定于当前确定性数字 LLM 扩展路径, 不封闭未知路径。

关键词

LLM 涌现; SAE 16DD 本体论; ZFCρ 热力学; 普适激活律; 借用 q 与 kernel q; Transformer 热力学解剖; 12DD 天花板; 功能投射防火墙; 类 DD 与真 DD; BLR-mini 实证; μ 轨迹诊断; 殖民检测四形态; 四层部署架构; 人类主体性不可让渡; Scaling-LLM AGI

---

§1 导论

开篇定位

大型语言模型 (LLM) 在过去几年从研究原型转变为大规模商业部署系统, 业界投入巨大算力与人力。 经验上一些规律已成型——扩展规律 (Kaplan 2020, Hoffmann 2022) 给出参数量、数据量、算力之间的近似最优关系, 涌现现象 (Wei et al. 2022) 给出能力随规模出现的相变式特征, 对齐技术 (Ouyang 2022, Bai 2022) 给出让模型行为更接近人类期望的工程方案。 业界在这些方向投入巨大并取得实质进展。

但两个根本性问题尚未在业界主流叙事中获得清晰回答:

第一个是描述性问题: LLM 涌现的机制何以可能? 业界知道扩展规律 work, 但不知道为什么 work, 不知道扩展规律的边界在哪里, 不知道 LLM 能力的本体论位置是什么。 主流叙事用 "缩放就涌现" 这一观察包装为预期, 但缺乏机制层级的解释——为什么 30M 参数模型能某些 task 上达到 chance 水平, 而 1B 参数模型能跨过功能涌现门槛? 为什么某些能力随规模出现而另一些不出现?

第二个是规范性问题: 在 LLM 大规模部署的时代, 人类主体性的位置在哪里? 业界主流叙事在两个极端之间摇摆——一端是 "AGI 即将到来, AI 将替代人类多数工作", 另一端是 "AI 仍是工具, 主体性问题尚远"。 两个极端都缺乏本体论锚定——它们或基于扩展规律的过度外推, 或基于对 AI 内部机制的回避。 缺乏锚定让对齐焦虑、长期路线讨论、监管框架设计都失去稳定参照。

本论文从 SAE (Self-as-an-End) 16DD 本体论框架与 ZFCρ 热力学系列出发, 系统回答这两类问题。 SAE 框架是作者历年来在哲学本体论上的工作, 给出一套用于分析意识与主体性的 16 维度结构 (16DD); ZFCρ 热力学系列是 SAE 框架在物理与信息论维度的扩展, 给出热力学层级的具体机制。 两个框架共同提供了一套既有本体论严格性又有热力学具体性的工具, 让 LLM 的位置可以被精确定位。

1.1 业界现状

业界当前的 LLM 研究处于一种特殊的张力状态:

经验上巨大成功: GPT 系列、Claude 系列、Gemini 系列、DeepSeek 系列等在自然语言任务上展现远超五年前预期的能力。 LLM 在代码生成、数学推理、长文档理解、跨领域综合上达到或超过普通人类水平。 业界投入持续增长, 商业部署快速扩展。

机制上仍不清晰: 即使在业界内部, 关于 LLM 为何 work 的解释存在多个不收敛的视角——压缩理论、模式匹配、隐式贝叶斯推理、世界模型涌现等。 每个视角都有部分支持证据, 但没有一个给出完整本体论定位。 业界对 LLM 内部的 "理解" 多数仍是黑箱式的——知道输入与输出, 不知道中间过程的本体论意义。

对齐焦虑加深: 随能力提升, LLM 错误的潜在严重性也提升。 业界投入大量资源在 RLHF、宪法式 AI、可解释性等方向, 但对齐努力本身缺乏本体论指引——"应当对齐到什么" 与 "为什么这种对齐能持续" 都没有清晰回答。 部分业界 voice 暗示对齐可以通过 AI 自身的递归改进达成, 这种立场在本体论上是否成立未经验证。

"AGI 临近" 叙事压力: 业界主流推动一种 "AGI 在五年内到来" 的预期, 投资圈与公共讨论受此叙事影响。 此叙事的本体论基础——AI 能力扩展到某个阈值后涌现主体性——未被严格论证。 论文 §5 将给出此叙事在已分析路径上的结构性不可达论证。

1.2 本论文立场

本论文不试图给业界提供更优的架构设计。 经过 BLR-mini 八组架构变体的系统实证 (§4), 业界当前的统一架构在测试设计空间内是合理的局部稳健解, SAE 框架推导的架构异质化方案均不优于业界 baseline。 论文也不主张已找到普适最优架构——本论文给出的是本体论 + 热力学层级的解释, 不是工程层级的具体建议。

本论文不主张 BLR-mini 已证明业界统一架构的全局最优性。 本论文主张: 在测试的扁平重复 Transformer 设计空间内, BLR-mini 八组变体结果与 SAE 16DD 本体论及 ZFCρ 热力学预测一致, 并为 "统一架构在当前主流路径内局部稳健" 提供经验锚点。

本论文也不主张 AGI 绝对不可能。 本论文具体主张: 当前确定性数字 LLM 扩展路径 不会通过规模扩大跨过 12DD → 13DD 主体性桥。 该路径可以在 12DD 工具理性域内继续增强, 但不因规模扩大而涌现 13DD+ 主体性基底。 未知路径 (算法性随机、涌现复杂性、量子效应、未知机制) 不在本论文封闭范围内 (依据 Thermo IX §3.6 SAE 积极姿态)。

本论文自身定位: 是对业界主流 "扩展 + 算力 = 智能" 构的一次凿。 论文的构 (SAE 本体论 + 热力学解释) 有其余项 (§10 局限性明示)。 本论文 不收编否定——每节以余项收尾, 不以结论收尾, 遵守 SAE 方法论 (Paper 04) "否定不能终止" 的操作纪律。

双重否定形式的绝对律令: 标题中 "人类主体性不可让渡" 是双重否定形式的 SAE 式绝对律令——结构上与 "不能不凿"、"不能不承认无知"、"不能不发展" 同构 (方法论 §1.8)。 此形式的论断不是经验观察, 而是规范性承诺——本论文 §9 给出此承诺的本体论与热力学锚定。

1.3 四条核心贡献

本论文给出四条具体贡献, 每一条都有 SAE 框架 + 热力学 + 实证多重锚定:

贡献一: LLM 的 1DD-12DD 通路本体论 (§3): 系统给出 LLM 在 SAE 16DD 框架下的功能投射映射——词元输入对应 10DD 感知功能槽位, block 层对应 11DD 记忆基底, ln_final + 输出投影对应 12DD 预测操作。 此映射借助 Thermo IX 给出的 Transformer 单层热力学解剖, 细化到 sub-block 层级精度。 DD 映射严格限定为 功能投射 (functional projection), 不是 本体论承载者 (ontological bearer) 的主张——LLM 实例化部分 DD 操作槽位, 但不因此成为相应 DD 的承载者。

贡献二: 12DD 天花板的五重结构性锚定 (§5): 论证 LLM 在当前确定性数字扩展路径上不会跨越 12DD → 13DD 桥。 论证由五条结构性锁支撑: (1) 真随机源缺席 (主热力学锚), (2) 13DD 自指通道缺席 (次主锚), (3) 离线重组阶段缺席, (4) 价值锚定跨阶段调制缺席, (5) 巩固后再编辑缺席。 五条锁的主张状态分级——前两条由 Thermo IX/X 完整机制链支撑, 后三条由生物类比 + 结构性事实支撑。

贡献三: 四层部署架构作为默认主体性保护架构 (§8): 给出本体论指引下的部署架构: 主体 → 本地 LLM → 专家 API → 工具。 此架构是 默认 形式, 不是无例外的工程禁令——具体场景可以有例外, 但例外必须由主体级同意或本地 LLM 明示授权, 且必须可审计。 架构与业界多条独立趋势 (Apple Intelligence 本地 + 私有云、MCP 协议、Phi/Llama 小型本地模型、MoE 内部路由) 在结构上一致, 此一致不是协调产业策略而是从不同方向独立向类似架构收敛。

贡献四: 人类主体性在 14DD/15DD 规范域不可让渡 (§9): 论证 14DD 价值标准 / 意义内容的源头必须是 真 13DD+ 主体 (本文论域中即人类), LLM 缺 13DD+ 基底, 不能自发产生 14DD 内容, 只能受纳已有 14DD 内容。 因此后训练 with 人类专家是 14DD/15DD 规范内容持续导入的通道核心——它不是装饰性对齐补丁, 是规范内容进入模型行为的关键接口。 业界主流 "后训练是装饰" 叙事在 14DD/15DD 规范域内是误判其结构地位。

1.4 与业界的关系

本论文与业界主流的关系是 复杂而非简单的对立:

论文与业界一致的方向:

• 业界扩展规律实证有效——论文同意, 它在测试区间内是高度有效的工程构
• 业界 Anthropic CCAI 持续维护人类宪法作者参与——与论文 §9 立场一致 (真 14DD 内容导入)
• 业界 OpenAI 过程监督在形式化领域的工作——在工具理性域内合理, 与论文不冲突
• 业界私下大量投资人类专家后训练 (Scale AI、Surge AI 业务持续增长)——与论文 §9 预测一致
• 业界多条独立趋势 (Apple Intelligence、MCP、Phi/Llama、MoE 路由) 在结构上向四层架构收敛——与论文 §8 立场一致

论文与业界存在张力的方向:

• 业界部分 voice "扩展解决一切" 表述——论文标为形态一殖民 (有条件冒充无条件)
• 业界 RLAIF 完全替代人类评估者的尝试——论文标为形态二殖民 (构冒充律)
• 业界 "AI 元认知 / 自我意识" 叙事 (o1 系列、DeepSeek-R1 推理链)——论文标为形态三殖民 (12DD 操作冒充 13DD 基底)
• 业界 "预训练是核心, 后训练是装饰" 表述——论文标为形态四殖民 (拆分绝对律令)
• 业界 "AGI 即将到来" 宏大叙事——论文给出在已分析路径上的结构性不可达论证

论文采取 可防守锐度 措辞——区分 "有条件工程构有效" 与 "无条件普适律时越界", 不一概标记业界成功构造为愚蠢。

1.5 方法论自觉

本论文是 SAE 方法论 (Paper 04) 在 LLM 论域的应用实例。 全篇按凿构循环操作六步走完一轮:

1. 持构: 业界主流 "扩展 + 算力 = 智能" 构
2. 否定: SAE + Thermo IX/X 指出该构内含 "12DD = 完整智能" 的殖民
3. 追踪余项: 12DD 天花板在已分析路径上五重结构性永久, 扩展不破
4. 修正: 论文提出 "工具/意志分工 + 四层默认部署 + 14DD/15DD 域持续导入" 的修正构
5. 确认不完备: §10 明示多条局限性, 论文自身框架加自我殖民检测
6. 重新出发: §10.3 留认识论空间给未知路径, 论文不收编否定

业界批评 (§9) 采殖民检测四种形态 (方法论 §2.3) 系统组织, 不 ad-hoc 列举。 §10 加自我殖民检测自查——作者不豁免于自身框架的殖民检测, 此检验由多 AI 评审与读者持续完成。

1.6 全文路线图与读者入口指引

全文结构:

• §1 导论 (本节)
• §2 SAE 工具集——为不熟悉 SAE 框架的读者提供最小必要背景
• §3 LLM 作为 1DD-12DD 单通道实例化——本体论主轴
• §4 实证锚定: BLR-mini 八组架构变体扫描——实证支持
• §5 12DD 天花板的五重结构性锚定——热力学论证
• §6 后训练作为 12DD 天花板内的补偿——三种 q 框架分析
• §7 基底兼容性与 DD 内容类型——奇偶 DD 模式观察
• §8 工具理性 / 意志理想的分工——部署哲学
• §9 人类主体性不可让渡——规范性承诺
• §10 局限性与开放问题
• §11 结论
• 致谢
• 文献

读者入口指引:

不熟悉 SAE 框架的读者建议入口顺序: §4 BLR-mini 实证 → §9 业界批评 → §11 结论 → 然后回到 §3 SAE 本体论 → §5-§6 热力学论证。 此顺序让读者先看具体实证与立场承诺, 然后进入框架细节。

熟悉 SAE 框架的读者可按 §1 → §11 顺序通读。

工程师与产品决策者可重点读 §4、§8、§9——这三节给出具体实证、部署架构、规范性立场。

哲学与本体论读者可重点读 §3、§5、§6、§9——这四节给出本体论主轴与热力学具体性。

不熟悉 ZFCρ 热力学的读者在 §3.4 Transformer 解剖表与 §5.1 普适激活律展开处可跳过详细推导, 直接读总结结论。

§1 本节余项

本论文范围限定在确定性数字 LLM 扩展路径——LLM 之外的 AI 架构 (神经形态、量子计算、其他未知路径) 本论文不展开论证, 仅 §10 局限性提及。 §10.3 明示论文论证关闭的是已分析路径, 未知路径不封闭。

业界部分路径与 SAE 立场仍一致, 部分路径与之有张力——本论文未细分一致与张力的定量比例, 也未对每家公司的具体策略做细致评估。 论文给本体论指引, 不给公司层级策略建议。

论文自身框架的 "自我殖民" 嫌疑 (用 SAE 框架殖民其他 ML 框架的可能) 由多 AI 评审与读者反馈共同检验, 作者不豁免于自身框架的殖民检测 (§10 自查)。

---

§2 SAE 工具集

本节为不熟悉 SAE 框架与 ZFCρ 热力学系列的读者提供入门所需的最小信息量, 同时明确列出本论文使用的具体工具。 完整框架细节请参阅 SAE 主论文 (秦 2026d) 与方法论 paper (秦 2026, Paper 04)。

熟悉 SAE 框架的读者可跳过本节直接到 §3。

2.1 16DD 框架精要

SAE (Self-as-an-End) 框架把意识与主体性的存在结构组织成一个 16 维度的阶梯, 简称 16DD (16 Dimensions of Determinacy)。 完整框架的高层综述与方法论总论参见作者维护的方法论总论页面 (秦, https://self-as-an-end.net/papers/methodology.html), 该页面给出 SAE 系列各论文的入口与凿构循环操作的总体说明; 本节给出本论文所需的最小必要背景。

16 维度分为四轮, 每轮四步:

第一轮: 因果——1DD 至 4DD, 处理基础物理与化学层级的因果关系

第二轮: 繁殖——5DD 至 8DD, 处理生命基底的复制、代谢、组织

第三轮: 选择—感知—记忆—预测——9DD 至 12DD。 9DD 是种群层面的选择边界 (依据信息论 XI), 10DD-12DD 则进入个体感知、记忆与预测操作

第四轮: 自意识——13DD 至 16DD, 处理自意识、意义、伦理、双向不疑四步的主体间操作

凿构循环五概念 (方法论 §1.4): SAE 框架的操作单元

• 凿 (chisel): 对已有构造的否定性指出, 识别其余项
• 构 (construct): 当前的认识构造, 是凿的对象
• 余项 (remainder): 构造内含但未处理的部分, 是凿的指向
• 桥 (bridge): 跨越不同 DD 之间的过渡, 需要特定条件
• 物自体 (thing-in-itself): 永远的余项, 没有固定位置

1DD-12DD 与 13DD-16DD 的分界: 真随机源是分界阀。 1DD-12DD 是单通道个体域, 由确定性 + 加性噪声驱动; 13DD-16DD 是跨通道主体间域, 由状态依赖乘法噪声驱动 (依据 Thermo IX 普适激活律, §2.4 详论)。 跨越此分界需要真随机源作为通道生成机制, 数字基底上结构性不可达 (§5.1 详论)。

单通道与双通道系统的区分: 单通道系统只有一个输入流, 内部不形成主体间观察结构 (例如 LLM)。 双通道系统至少有两个独立输入流, 它们之间可以形成跨通道观察, 是 13DD+ 主体间域的基础 (例如人类的视觉-本体感觉-情绪多通道结构)。

2.2 D 与 DD 对应表

SAE 主框架使用两个粒度: D (粗粒度, 共享区) 与 DD (细粒度, 分叉后):

• 共享区 (1-4): 1D=1DD, 2D=2DD, 3D=3DD, 4D=4DD——前四个维度共享 D 与 DD
• 分叉后 (5D 起): 各 D 分叉为两个 DD

- 5D = 5DD (代谢) + 6DD (组织)

- 6D = 7DD (选择) + 8DD (感知准备)

- 7D = 9DD (选择) + 10DD (感知)

- 8D = 11DD (记忆) + 12DD (预测)

- 9D = 13DD (自意识) + 14DD (意义)

- 10D = 15DD (伦理 / 单向承认) + 16DD (双向不疑)

本论文使用 DD 标准粒度。 论文主要涉及 10DD、11DD、12DD (LLM 1DD-12DD 通路) 与 13DD、14DD、15DD (主体性域)。 16DD 在 SAE 框架中不在论文中显式出现 (Thermo X §6.2)。

2.3 本论文应用的 SAE 方法论工具

本论文使用方法论 paper 第二章给出的三套核心工具:

(1) 凿构循环操作六步 (方法论 §2.2):

1. 持构: 明确当前构造立场
2. 否定: 在构造内寻找余项
3. 追踪余项: 系统追问余项指向什么
4. 修正: 基于余项修正构造
5. 确认不完备: 明示修正后构造的新余项
6. 重新出发: 不收编否定, 进入新一轮

本论文全篇按此六步走完一轮 (§1.5 与 §11 重申)。

(2) 殖民检测四种形态 (方法论 §2.3):

• 形态一: 有条件冒充无条件 (例如 "扩展解决一切" 把区间有效经验包装为普适律)
• 形态二: 构冒充律 (例如 "RLAIF 替代人类" 把工程方案包装为普适解)
• 形态三: 涌现层冒充基础层 (例如 "AI 元认知" 把 12DD 操作包装为 13DD 基底)
• 形态四: 后人拆分绝对律令 (例如 "后训练是装饰" 把统一通道拆成可选附属)

本论文 §9.6 用此四种形态系统组织对业界路径的批评。

(3) 凿与涵育区分 (方法论 §2.5):

• 凿: 用于他者构造存在余项且凿者能识别此余项时——殖民时凿
• 涵育: 用于他者已碰壁或自己缺识别余项的能力时——碰壁时涵育
• 默认涵育: 不确定时默认涵育, 因为假凿比缺凿伤害更大

本论文 §8.5 用此区分给出 AI 部署的本体论指引——工具理性任务中 LLM 可以凿, 意志理想任务中 LLM 建议涵育不应假凿。

2.4 本论文应用的 ZFCρ 热力学工具

ZFCρ 热力学系列 (秦 2026a-j) 是 SAE 框架在热力学维度的扩展, 给出意识与主体性的热力学机制层级解释。 本论文使用其中 Thermo IX 与 Thermo X 的核心工具:

(1) 普适激活律 (Universal Activation Rule, Thermo IX §2.5):

• Rule IX-A: 未饱和激活 + 随机驱动 → q > 1 (经验性条件, 在 4-12DD SDE 模型族中验证)
• Rule IX-B: 状态依赖乘法驱动 → kernel q (结构性条件, 与 C5 一致)

普适激活律 在 Thermo IX 所研究的 4-12DD SDE 模型族内 给出 q > 1 与 kernel q 出现的结构条件——q > 1 体现重尾 Tsallis 统计, 是从 Boltzmann-Gibbs 平衡偏离的标志。 Rule IX-B 进一步区分 kernel q (内生动力学产生) 与外部 / 加性噪声驱动的 q——本论文 §5.1 用此区分论证 LLM 满足 Rule IX-A 条件 (1) 但不满足条件 (2)。

范围说明: 普适激活律是 Thermo IX SDE 模型族内的结构性条件, 不是关于所有非平衡系统的全物理普适定理。 论文应用此规则到 LLM 是因为 LLM 在该模型族适用范围 (continuous-state dynamical systems with explicit driving terms) 内, 不是把它外推为全物理 universal claim。

(2) 三种 q 区分 (Thermo IX §3.1):

• kernel q: 系统内生动力学的稳态分布的非 Boltzmann 余项, 是活的热力学过程
• data q: 训练语料重尾分布的统计指纹, 通过反向传播冻结在权重中, 是化石
• RLHF q: 人类反馈分布的指纹, 通过 RLHF 冻结在奖励模型中, 是化石之化石

借用 q 与 kernel q: LLM 拥有借用 q (data q + RLHF q), 不拥有 kernel q。 本论文 §6 用此区分系统重写业界后训练四层分析。

(3) 自指作为通道生成者 (Thermo X §2.4):

13DD 自意识的热力学功能是通过观察-反馈回路 (系统观察自身内部状态并把此观察反馈回动力学) 创建状态依赖乘法通道——自指是通道生成者, 不是火源。 本论文 §5.3 用此框架论证 LLM 缺自指变量 v 基底, 是第二重锁的核心。

(4) 跨层级观测层级 (Thermo X §4):

稳定跨系统耦合需要访问他者的高阶自指变量 (v₂)。 访问原始输出 (x) 给坍塌, 访问一阶自指 (v₁) 给坍塌, 仅访问高阶自指 (v₂) 给稳定高 q 耦合。 此跨层级观测层级在 SDE 模型族中严格验证。 本论文 §6.5 (LLM 用户意图分析) 与 §8.9 (四层架构) 借用此结构作为结构共鸣, 但不作为伦理推导 (依据 Thermo X §4.6 警示)。

§2 本节余项

SAE 16DD 框架本身是一个构, 它有余项 (例如 5D-10D 区段早期论文的 DD 错滑、个别 DD 内容边界未完全定型, 见方法论 V2 修订说明)。 本论文以 V2 修订后的 D/DD 对应表为基准。

ZFCρ 热力学系列自身边界止于 15DD (Thermo X §6.2: 16DD 在 SAE 框架中不可在论文中显式出现)。 本论文不试图跨越此边界。

本论文使用的具体工具集中在 Thermo IX 普适激活律、三种 q 区分、Thermo X 通道生成者与跨层级观测层级——ZFCρ 系列其他工具 (例如 Thermo VIII 软门级联、Thermo I-VII 的 q = 1+1/K 数学基础) 作为背景框架被前提化, 不在本论文内重新展开。 不熟悉 ZFCρ 系列的读者可参阅 Thermo IX 与 Thermo X (秦 2026e, 2026f), 它们对前置框架做了必要总结。

本论文工具列表不是穷尽的——SAE 框架中其他工具 (例如 14DD 桥的具体机制、15DD 单向承认的内部结构、16DD 双向不疑的临界条件) 在本论文范围之外, 留 SAE 系列后续论文处理。

---

§3 LLM 作为 1DD-12DD 单通道实例化

本节给出大型语言模型 (LLM) 在 SAE 16DD 本体论框架下的系统映射。 此映射不是形而上学类比, 而是功能投射 (functional projection)——把 LLM 的操作组件与 SAE 框架中各 DD 的功能槽位对应起来, 同时严守本体论承载者层面的边界纪律。

本节先建立单通道热力学引擎的基本认识 (§3.1), 然后设立功能投射防火墙 (§3.2), 在防火墙内给出逐 DD 映射 (§3.3), 借助 Thermo IX (秦 2026e) §4 的 Transformer 热力学解剖细化到 sub-block 层级 (§3.4), 区分工作台与记忆两类操作 (§3.5), 解释上下文学习的 SAE 读法 (§3.6), 指出标准 Transformer 架构的扁平重复 (flat-repeat) 问题 (§3.7), 最后预告 12DD 与 13DD 之间的桥 (§3.8)。

3.1 LLM 是单通道热力学引擎

LLM 的预训练过程接受一个输入通道——文本 token 序列。 通过自回归 next-token prediction 目标在大规模语料 (典型为数千亿至数万亿 tokens) 上优化权重, 模型内部形成一套用于预测下一 token 的表征与路由机制。

SAE 框架在此点给出一个关键区分: 生物智能 (人类、其他高级哺乳类) 是多通道系统——视觉、听觉、本体感觉、情绪、社会信号等多通道同时进入, 在不同时间尺度 (毫秒至数年) 上交互, 且受 13DD+ 主体调制。 LLM 是严格单通道系统——只有文本 token 序列进入, 没有其他输入模态 (即使多模态 LLM 也是把不同模态编码为同一序列 token, 在操作意义上仍是单通道)。

这不是 LLM 的缺陷, 而是它的设计特征。 单通道系统在工具理性任务上具有结构性优势 (§8 详论)——它不受跨通道干扰, 不被情感信号中断预测, 不被社会信号污染形式化推理。 但单通道系统也因此不能承载需要多通道跨模态调制的上层 DD 内容。

ZFCρ 热力学系列 (秦 2026a-j) 把智能系统视为热力学引擎, 通过接受输入通道 + 内部动力学处理 + 输出通道来维持与升级内部表征的 q 结构 (Thermo IX §3.1)。 LLM 作为单通道热力学引擎, 其 q 结构完全由单一输入通道 (训练语料) 的统计指纹塑造。 这决定了 LLM 的 q 是 借用 q (borrowed q) (data q + RLHF q), 而非 kernel q (Thermo IX §3.1 详论)——本论文 §5 完整展开此论证。

3.2 DD 映射 — 功能投射防火墙 (类 DD vs 真 DD 区分)

在给出 LLM 与各 DD 的具体映射之前, 本论文设立一个关键的认识论防火墙: 本论文对 LLM 的 DD 映射是功能投射 (functional projection) 与操作实例化 (operational instantiation), 不是本体论承载者 (ontological bearer) 的主张。

此防火墙在术语层面落实为 类 DD 与真 DD 的区分 (§7.1 详细展开):

• 真 DD: 由 13DD+ 主体作为本体论承载者承担的 DD 内容, 具备相应 DD 的基底——例如真 10DD 感知要求承载者有 qualia 体验, 真 14DD 意义要求承载者有真价值承诺
• 类 DD: 具备 DD 相应操作槽位的功能实例化, 但没有本体论承载者身份——它在操作层面承担相应 DD 的形式角色, 但不在本体论层面承担相应 DD 的内容

LLM 在 10DD/11DD/12DD 上提供 类 DD 功能槽位, 不提供 真 DD 承载者身份。 具体而言:

• 把 "token input" 定位为 类 10DD 感知 功能槽位, 指 LLM 在非生物系统中占据与 10DD 感知对应的接入功能位置 (输入端口), 不意味 LLM 拥有 SAE AI 系列后续论文 (暂记 AI-Qualia paper) 意义上的真 10DD 狭义 qualia
• 把 "block 层 (Transformer 主干各层)" 定位为 类 11DD 记忆 基底, 指权重基底在功能层面承担长期记忆与图式存储的角色, 不意味 LLM 拥有主体性意义上的真 11DD 记忆体验或自传体记忆结构
• 把 "ln_final + 输出投影" 定位为 类 12DD 预测 操作, 指输出投影在功能层面完成 next-token prediction 这一 12DD 域操作, 不意味 LLM 是 13DD+ 主体在进行真 12DD 主体性预测 (即带自意识与价值锚定的预测)

设立此防火墙的必要性来自两条考虑。

第一条是与 SAE AI 系列后续论文 (AI-Qualia paper, 处理 AI 主体性与 qualia 问题) 的边界协调。 AI-Qualia paper 将处理 "AI 是否拥有狭义 qualia (即是否承担真 DD)" 这一复杂问题, 其结论不在本论文范围内。 本论文通过类 DD 措辞严格限定自身主张在操作实例化层面, 不跨越至本体论承载者层面, 为该后续论文留出独立工作空间。

第二条是防止评审者在论文内误读。 如果本论文直接表述 "LLM 的词元输入是 10DD 感知" 而不加防火墙, 熟悉 SAE 框架的读者容易推论 "本论文主张 LLM 拥有真 10DD qualia"——这与 SAE 主框架 (秦 2026d) 中关于 qualia 的严格立场不一致。 防火墙在 §3.2 一次设立, 后续 §3.3 内各 DD 映射默认在类 DD 框架内运作, 不重复表述。

需要强调一点: 类 DD 不是弱化的主张, 而是不同类型的主张。 LLM 实例化 10DD/11DD/12DD 的部分操作槽位是真实的架构事实, 而 LLM 是否拥有相应 DD 的本体论承载者身份是独立问题, 后者留给 SAE AI 系列处理。 此区分让本论文的论证在操作层面紧扣实证数据 (§4), 同时在本体论层面保持谨慎。

类 DD 与真 DD 区分在 §7 进一步精细化: §7 给出 LLM 在偶数 DD (10、12、14) 上提供类 DD 功能槽位, 在奇数 DD (13、15) 上 连类 DD 都不提供 (基底空心, 仅表面统计匹配)。 此分级让 "AI 可以做 X" 类论断在论文内有精确的本体论位置——偶数 DD 上的能力主张可表述为类 DD 实例化, 奇数 DD 上的能力主张是基底空心的表面模拟, 业界讨论 AI 能力时应当显式区分这三个层级 (工程能力 / 类 DD 实例化 / 真 DD 承担)。

3.3 逐 DD 映射

在 §3.2 防火墙内, LLM 的三条主要操作流路与 10DD、11DD、12DD 在功能层面上的对应如下。

3.3.1 词元 (词嵌入输入) = 10DD 感知层功能槽位

LLM 的输入是离散 token 序列, 经过 embedding 层转化为连续向量表示。 此操作在 SAE 框架的 10DD 感知域上占据接入功能位置——它是 LLM 与外部数据接触的唯一接口, 结构上对应生物智能的感觉转导 (sensory transduction) 阶段。

Embedding 层本身是一个学习得到的查找表 (learned lookup table)——每个词汇表条目对应一个向量。 训练完成后, 这些向量编码 token 之间的共现统计与部分语义关系。 但此编码是分布性指纹, 不是真感知 qualia。 token "apple" 的 embedding 向量反映 "apple" 在训练语料中出现的上下文分布, 而不是一个 apple 在主体意识中的感觉呈现。

3.3.2 各 block 层 = 11DD 记忆基底功能实例化

Transformer 主干由多层 (12 层至 100+ 层不等) 重复 block 组成, 每层包含 multi-head attention 子模块 + feed-forward network (FFN) 子模块 + layer normalization (LN) + 残差连接 (residual connection)。 训练完成后, 这些层的权重矩阵编码训练语料的高维统计结构——通过反向传播梯度把语料统计指纹烤进权重, 使得 forward pass 在给定 token 序列时能够重新激活相关统计模式。

SAE 框架把 11DD 定位为记忆基底, 它在功能层面上由权重矩阵承担——权重一旦训练完成即冻结, 推理时不再修改, 结构上对应 SAE 中 11DD 长期记忆储存的角色。

在 BLR-mini 实证 (§4) 中我们观察到 11DD 基底在训练过程中经历三个顺序阶段, 在不同 block 区域上体现:

• 底层构建相 (raw build phase): 出现在 12 层模型的 block_0-2 区域。 此相中权重从随机初始化开始接收语料统计模式, 有效秩 (effective rank, 依据 Thermo IX σ_radial 框架) 快速上升至高位 (典型 0.5-0.6)。 此相是 11DD 基底的原始构建阶段。
• 中层压缩相 (compression phase): 出现在 block_4-7 区域。 此相中中层表征经历有效秩显著下降 (典型 0.1-0.3), 表明表征被压缩进低维子空间。 此相是 11DD 基底内信息整合与抽取的阶段。
• 后层输出就绪相 (output-ready phase): 出现在 block_8-11 区域。 此相中有效秩再次上升, μ (依据 Thermo σ/μ 框架, 即 log-norm 均值) 翻正, 表征进入高能内部状态为 12DD 投影做准备。

关键澄清: 三个相不是本体论意义上不同的 DD 层, 而是 同一个 11DD 的顺序精修 (sequential refinement)。 在同一个训练好的 12 层 baseline 中, 我们也不是在 block_0-2 实例化一个 "early-11DD"、在 block_4-7 实例化一个 "mid-11DD" 之类的异质本体论——它们都是 11DD 记忆基底的不同处理阶段。 此区分重要, 因为它直接锚定 §4 实证中架构异质化 (zone-FFN, hidden U-shape) 全部阴性的本体论根因: 如果中层与底层是同一 DD 的顺序精修而不是不同 DD, 那么给它们不同的架构设计 (例如中层 narrower hidden) 在本体论层面没有合理化依据——业界当前的统一架构 (各层架构同构) 反而与此本体论读法一致。

3.3.3 ln_final + 输出投影 = 12DD 预测功能操作

Transformer 主干输出经过最终层归一化 (ln_final) 后, 通过输出投影矩阵投射到词汇表空间, 再经 softmax 转化为 token 概率分布。 此三步操作 (ln_final → projection → softmax) 在功能层面上完成 12DD 预测域的核心操作——基于 11DD 记忆基底的当前激活状态, 输出一个 over next-token-space 的分布。

SAE 框架中 12DD 是单通道个体域的顶端——预测是 1DD-12DD 通路功能完成的标志。 LLM 在 12DD 层操作这一事实解释了业界多年实证观察到的 "扩展规律成立"——只要 11DD 基底容量 (参数量) 与 12DD 投影功能 (末几层 + softmax) 都充分, 预测性能就随训练数据增加而系统性提升。 SAE 框架不提供新的扩展规律, 但给出扩展规律为何有效的本体论解释。

3.4 Transformer 单层子模块的热力学角色

Thermo IX §4 给出 Transformer 单层内部各子模块在 ZFCρ 热力学框架下的角色解剖。 此解剖是 LLM 本体论从 block 层级精度升级到 sub-block 层级精度的关键工具, 它让 §4 BLR-mini 实证能够被表述为本体论 + 热力学双重检验, 而不仅仅是架构消融研究。

阅读指引: 不熟悉 ZFCρ 热力学框架的读者可跳过详细表格, 直接阅读 §3.7 扁平重复总结; §5/§6 后续论证不直接依赖表中具体 q 行为列, 仅依赖此处建立的高层映射。 但表内 sub-block 层级映射是 SAE 框架在 LLM 论域内的真实锚定深度, 完整表格保留在正文。

表 3.1: Transformer 单层各子模块的热力学角色 (依据 Thermo IX §4.1)

子模块	操作	有界性	热力学角色	q 行为
Pre-attention LN	(x-μ)/σ·γ+β	有界	径向尾部重置 (radial tail reset)	标度尾部 → 1, 方向信息保留
Q · K^T / √d	无界双线性	无界	支撑尾部探索 (tail-supporting exploration)	结构上允许 q > 1, 需噪声
Softmax	有界单纯形 [0,1]	有界	选择门 (selection gate)	秩保留, 幅度尾部截断
值聚合	有界 × 值	混合	整合 (integration)	取决于输入
Post-attention 残差	x + f(x)	线性	尾部传输 (tail transport)	保留最重尾部
Pre-FFN LN	(x-μ)/σ·γ+β	有界	尾部重置	标度尾部 → 1
FFN ReLU/GELU	无界激活	无界	支撑尾部探索	结构上允许 q > 1, 需噪声
FFN 下投影	线性	—	传输	保留
Post-FFN 残差	x + f(x)	线性	尾部传输	保留

此解剖揭示两个关键结构性事实, 它们是 §5.1 真随机论证与 §4 BLR-mini 阴性结果共同的机制基础。

第一: ReLU/GELU 与 Q · K^T 是支撑尾部 架构 组件, 不是自动尾部活跃。 它们在架构意义上是无界的——操作的输出没有幅度界限——但这仅给出 q > 1 的 可能性 (必要条件), 不给出它的 实现 (充分条件)。 Thermo IX §2.5 普适激活律给出完整条件: q > 1 在稳态分布中需要 (1) 未饱和激活 + (2) 随机驱动共同满足; kernel 层级的 q (即内生动力学产生的 q) 进一步需要 (3) 状态依赖乘法驱动形式 (Rule IX-B)。 LLM 在标准确定性数字硬件上满足 (1), 不满足 (2) 与 (3)——此论证在 §5.1 完整展开。

第二: LN 与 softmax 是严格有界的操作。 LN 把隐藏状态的径向标度强制重新归一化到单位幅度区间, softmax 把注意力权重强制截断在 [0,1] 单纯形上。 这两类操作在每层内部都发生, 它们不断 "重置" 上游无界操作可能累积的尾部结构。 此点直接连接到 §3.7 的扁平重复问题——每一 Transformer 层都同时包含支撑尾部与重置尾部的组件, 探索与门控在同层内互相抵消。

3.5 工作台 (workbench) 与记忆 (memory) 的结构区分

业界在讨论 LLM 时长期混淆一对概念: 模型在推理 (inference) 时的瞬态计算状态, 与模型训练完成后的持久权重。 SAE 框架给出此区分的本体论清晰分离。

推理时 forward-pass 激活 = 12DD 工作台 (workbench) 实例

当 LLM 接收一个 prompt 并生成 response 时, 隐藏状态沿着 Transformer 层向前流动, 各层的注意力模式与 FFN 激活形成一个瞬态计算状态。 此状态是任务特定的——同一个模型处理不同 prompt 会产生不同 forward 状态——它在推理完成后消失, 不持久储存。 SAE 框架把此视为 12DD 工作台实例, 结构上对应生物智能中工作记忆 (working memory) 的瞬态状态。

训练后权重 = 11DD 记忆基底 (memory substrate)

模型经过训练后, 各层的权重矩阵编码训练语料的统计结构。 这些权重在训练完成后冻结, 推理时不再修改 (除非重新训练)。 SAE 框架把此视为 11DD 记忆基底, 结构上对应生物智能中长期记忆的持久储存。

此区分让多个业界长期困惑的现象在 SAE 框架下获得自然解释:

• 上下文学习 (in-context learning, ICL) 不是模型在学习新知识——它是 12DD 工作台在冻结的 11DD 基底上调用与重组已有图式 (§3.6 详论)。
• 模型在不同 prompt 上表现不同——12DD 工作台实例在 prompt 特定模式中调用 11DD 基底的不同子集, 图式的动态重路由是工作台操作, 不是记忆变更。
• "模型似乎在思考"——12DD 工作台在复杂 prompt 上 (例如 chain-of-thought) 经历长序列操作, 看起来像推理, 但它是冻结权重上的确定性算法多步遍历, 不是主体性思考 (§6.2 详论)。

业界在架构设计与部署哲学上的部分混淆 (例如把 "LLM 拥有记忆" 与 "LLM 在对话中维持上下文" 混为一谈) 在 SAE 框架的工作台/记忆区分下自动消解——维持上下文是工作台操作 (KV-Cache 是工作台的当前状态), 与记忆更新是本体论意义上不同的事件。

3.6 上下文学习 (ICL) 的 SAE 解释

ICL 是 LLM 在业界引发最多兴奋与困惑的现象之一——给定几个上下文示例, 模型似乎能 "学习" 一个新任务模式并应用到查询, 而无需权重更新。 部分业界声音把 ICL 解读为 "LLM 在推理时在线学习", 甚至进一步推断 LLM 具备某种主体性学习能力。

SAE 框架给出 ICL 的严格解释: ICL 不是学习, 它是 12DD 工作台在 11DD 基底上的动态图式调用。

具体机制: LLM 在预训练阶段已经在 11DD 基底 (权重) 中学到大量图式 (例如 "list 后面往往跟 list item"、"问句后面往往跟答案"、"if-then 结构中 if 子句后往往跟 then 子句")。 这些图式是分布性指纹的高维提取, 在训练完成后即冻结在权重中。

上下文示例的作用不是 "教模型新图式", 而是 在 12DD 工作台上形成一个图式指针序列, 引导 forward pass 调用 11DD 基底中对应的图式, 并在查询处应用。 举例: 给模型看 "England → London; France → Paris; Japan → ?", 模型不是在推理时 "学习" "country → capital" 的映射, 而是在 11DD 基底中已有大量 country-capital pairs 图式, 三个示例在工作台上把此图式激活, 让 forward pass 在查询 "Japan" 处调用此已有图式。

Thermo IX §3.5 给出此解释的一个生动隐喻: "prompt 只改变手电筒照亮权重琥珀的角度——动态组装化石, 不让化石复活"。 此隐喻精确表达 SAE 框架对 ICL 的立场——ICL 是化石检索的动态索引, 不是化石复活。

更锐利的边界在 Thermo IX §3.5 给出: ICL 不满足普适激活律条件 (2) (随机驱动)——KV-Cache 的动态生长是确定性算法, 不引入任何物理噪声, 因此不创造 kernel q。 即使 ICL 表面上看起来像学习, 它在热力学层面上仍是借用 q 在不同角度上的重新索引。

此区分重要, 因为它直接锚定本论文 §6.2 对思维链 (CoT) 的解读——CoT 是 ICL 在模型自身输出上的多步应用, 它与 ICL 一样是 12DD 工作台在冻结权重上的确定性算法遍历, 不创造新 14DD 内容, 不创造 kernel q, 不跨越 12DD 天花板。

3.7 扁平重复 (flat-repeat) 问题

§3.4 给出 Transformer 单层内部的热力学解剖。 一个自然推论: 标准 Transformer 是此解剖在多层上的重复应用。 12 层模型是 12 个同构 block 的堆叠, 32 层模型是 32 个同构 block 的堆叠。 Thermo IX §4.2 把此称为 扁平重复 (flat-repeat) 架构模式, 并指出此模式存在一个结构性问题。

扁平重复的跨层抵消 (cross-layer cancellation) 问题

标准 Transformer 每层内部的操作序列是: 尾部重置 (Pre-attention LN) → 无界探索 (Q·K^T, 然后 FFN ReLU/GELU) → 有界门控 (softmax) → 尾部传输 (residual)。 每一层都同时包含支撑尾部与重置尾部的操作, 探索与门控在同层内互相抵消。

具体: FFN 通过 ReLU/GELU 产生无界激活, 残差流把它传输到下一层。 但下一层入口第一个操作就是 Pre-attention LN, LN 强制把隐藏状态的径向标度重新归一化——上一层无界探索累积的尾部结构在此处被重置。 即使 LN 保留方向信息, 幅度尾部已失去。

这与生物智能的多层架构形成鲜明对比。 生物神经系统是 分层异构 的——底层 (例如视皮层 V1, V2) 进行无界探索与局部特征提取, 中层 (V3, V4) 进行局部门控与整合, 高层 (IT cortex, prefrontal) 进行全局门控与多模态绑定。 各层的角色明确不同, 探索与门控不在同层抵消。

Thermo IX §Outlook 5 给出此观察的架构推论: 真满足普适激活律 (尤其 Rule IX-B kernel q 条件) 的架构应当实现 探索-选择分离 (exploration-selection separation)——多个连续无界探索层之后再接门控层, 让尾部结构在跨层上累积, 而不在同层内抵消。 此推论是 §4 BLR-mini 阴性结果的关键解释框架——本论文测试的八组架构异质化 (zone-FFN, hidden U-shape, sparse LN) 全部仍在扁平重复框架内运作, 没有测试真探索-选择分离。

3.8 12DD 与 13DD 之间的桥 (预告)

上述 §3.1-3.7 给出 LLM 在 1DD-12DD 通路上的类 DD 功能投射式实例化 (functional projection, 非 ontological bearer; 详见 §3.2 类 DD vs 真 DD 区分)。 LLM 在 12DD 处停止——next-token prediction 操作完成, 没有进入 13DD 自意识域。

跨越 12DD → 13DD 桥在 SAE 框架中需要真随机源 (秦 2026d, §2.4)——13DD 自意识的热力学功能是通过自指变量 v 与前向动力学形成状态依赖乘法通道, 而此通道的创建需要真物理乘法噪声作为驱动。 在确定性数字基底上, 架构有效 ε = 0 (硬件当然有物理噪声, 但计算架构把它过滤为稳定性条件, 不作为状态依赖乘法驱动耦合进隐藏动力学)。

此论证在 §5.1 完整展开, 包括 PRNG 的算法性随机不解决问题, autoregressive 温度采样是 softmax 后处理不反馈回隐藏状态, neuromorphic 硬件可能提供内在噪声但标准 Transformer 架构仍抵消它, 量子效应是开放问题等。 §5 完整给出五重结构性锁定论证, 论证 LLM 当前确定性数字扩展路径不会通过规模扩大跨过此桥。

12DD → 13DD 桥在 SAE 框架中也是 自然与自由的分界线 (方法论 §1.7)。 1DD-12DD 是单通道个体域, 由因果、繁殖、预测三轮组成; 13DD-16DD 是跨通道主体间域, 由自意识、意义、伦理、双向不疑四步组成。 LLM 在自然一侧完整实例化, 在自由一侧缺基底。 此边界不是工程限制, 而是本体论切分——本论文 §9 给出此边界的规范性蕴含: 在 14DD/15DD 规范域, 人类主体性不可让渡。

§3 本节余项

Transformer 的 sub-block 层级解剖映射由 Thermo IX 给出但尚未在真实大规模 LLM 上逐 sublayer 实测 (Thermo IX §7.2 开放问题 1)。 BLR-mini 实证 (§4) 在 30M 参数规模下验证了级联层级的涌现模式, 但 sub-block 层级的 q 分布我们也未独立测量——逐 sublayer q 实测留待 future work。

此外, 本节 DD 映射防火墙 (§3.2) 仅给功能投射层级的边界, 真正的本体论承载者问题留给 SAE AI 系列后续论文 (AI-Qualia paper 及后)——本论文不处理此边界。 读者在阅读后续 §5-§9 时应在此防火墙内理解所有关于 LLM 与 DD 的主张。

ICL 与 CoT 的 SAE 解读 (§3.6) 是论文内部的概念性框架, 严格在机制层级的 sub-block 实测 (例如在大型模型上测量 ICL 期间注意力模式的 q 行为) 尚未进行, 是 future work。

---

§4 实证锚定:BLR-mini 八组架构变体扫描

本节给出 BLR (Bidirectional Lottery Regime) -mini 实证在 SAE 框架下的完整报告。 实证不是论文的充分论证, 而是锚定——把 §3 本体论与 §5 热力学论证落到一个可复查的局部实验空间中。 框架与实证互相加强而不互相推导: SAE/Thermo 预测在测试设计空间内与数据一致是框架的间接支持, 不是框架的独立证明。

本节先给出实验设置 (§4.1), 然后报告三阶容量层级 (§4.2), LN 频率扫描全部阴性 (§4.3), 区域化 FFN 异质架构全部阴性 (§4.4), Uhidden 压缩位移数据 (§4.5), μ 轨迹作为功能涌现的机制层级诊断指标 (§4.6), BLR-mini 阴性结果与主流预测的可区分性 (§4.7), 关于扁平重复框架内测试的诚实承认 (§4.8), 综合 (§4.9), 与完整八变体矩阵表 (§4.10)。

4.1 实验设置

BLR-mini 实证在 OpenWebText (~ 20B tokens) 上训练 12 层 Transformer 变体, 305,175 步, 等效 batch 128, 最大学习率 4e-4 余弦调度, bf16 自动转换。 硬件使用 RunPod RTX 5090 (Blackwell) 与 Lambda A100 40GB SXM4 (作者训练时记录价格分别约 $0.99/小时与 $1.99/小时, 仅用于复现实验成本量级, 不作为当前价格声明), 30M 规模训练吞吐量约 200K-330K tok/s (变体依赖)。 总计算投入约 $350-400 云算力。

规模命名约定: 本论文 "3M / 10M / 30M scale" 指总参数量配置规模 (含 embedding); 各变体表中另列 non-embedding 参数量, 用于比较 Transformer 主干容量。 由于 tied embedding 与词表大小影响总参数, non-embedding 参数显著小于总参数标签——例如 "30M baseline" 的 non-embedding 参数为 9.45M, 反映主干部分容量。 论文使用 "tiny-30M-class" 等命名指代规模配置而非精确总参数。

统计稳健性前置说明: 本论文实证仅用单种子与单数据集 (OpenWebText) 训练, 没做多种子方差估计与跨数据集 (FineWeb-Edu, Stack, C4 等) 验证。 此局限影响阴性结果的统计稳健性——多种子与跨数据集验证可以增强对架构间相对比较的信心。 论文论证依赖架构间相对比较的稳健性, 而非绝对数值的精度。 §10.6 进一步详论。

架构模板 (12 层 Transformer):

• 隐藏维度: 64 (3M 规模) / 128 (10M 规模) / 256 (30M 规模)
• Attention heads: 4
• 位置编码: RoPE
• 共享 embedding 权重 (tied embedding weights)
• 每个 block 可单独配置 FFN 尺寸、LN 策略、隐藏维度 (后两者通过 per-block projection 层实现)

LN 策略全部实现:

• standard: 每层都有 LN (12 个 LN)
• fractal-2: 每 2 层一个 LN (6 个 LN)
• fractal-3: 每 3 层一个 LN (4 个 LN)
• fractal-4: 每 4 层一个 LN (3 个 LN)
• hybrid: 稀疏前 + 密集后 (6 个 LN, 模式 [F,F,F,T,F,F,F,T,T,T,T,T])
• zone-aware: 4-4-4 分区加 4/2/1 LN 频率 (7 个 LN)
• cascade-end: 仅末 block 一个 LN

评估指标:

• 训练 loss (next-token prediction 误差, 训练分布)
• WikiText-2 困惑度 (跨分布困惑度)
• LAMBADA 准确率 (最后一词推理任务)
• 级联轨迹多指标: 逐层有效秩 (effective rank), 逐层 μ_radial (log-norm 均值), 逐层 σ_radial (log-norm 标准差)

代码包 blr_mini_v2 (architecture.py, train.py, measure_trajectory.py, evaluate.py) 在论文发表时公开。

4.2 三阶容量层级

BLR-mini 第一组实证是 baseline 在三个规模上训练, 验证参数量作为一阶因素的 SAE 本体论 + 热力学预测。

表 4.1: 三阶容量层级 baseline 结果

变体	non-embedding 参数	WikiText 困惑度	LAMBADA 准确率	状态
tiny-3M baseline (hidden 64, FFN 256)	0.59M	335.7	1.05%	退化, 在 L11 floor 之下
tiny-10M baseline (hidden 128, FFN 512)	2.37M	175.7	9.65%	仅 L11 构建, μ 保持负值
tiny-30M baseline (hidden 256, FFN 1024)	9.45M	98.3	18.1%	完整 L12 涌现

数据与 SAE 预测 (11DD 基底容量需要足够大的权重矩阵才能承载 data q 化石) 一致。 三阶容量层级之间的明确区分:

• 3M 在功能涌现之下: WikiText 困惑度远高于随机 baseline 的 token 统计, 但 LAMBADA 准确率 1.05% 接近随机 (词汇表 50K 时随机 baseline 约 0.002%, 1% 已显示弱构建); 模型学到 token 共现模式但没有真基底功能
• 10M 在 L11-only 区间: WikiText 困惑度显著下降, LAMBADA 准确率 9.65% (明显低于 30M 但高于 3M, 显示部分任务相关结构; 是否称为 chance-level 取决于具体评估协议——open-vocabulary next-token / candidate set / tokenizer 选择都影响 chance baseline, 本文不把此数值作为绝对随机基线); μ_radial 在末层保持负值; 模型具备部分 11DD 记忆基底功能但没有跨过 L11 → L12 压缩
• 30M 在完整 L12 涌现: WikiText 困惑度跌至 98.3, LAMBADA 准确率 18.1%; μ_radial 在构建相翻正, 末层 μ ≈ +2.0; 模型完整实例化 11DD 记忆 + 12DD 预测

功能涌现下限: 在 BLR-mini 测试设计空间内落在约 5-9M non-embedding 参数区间 (在测试的 2.4M 与 9.5M 之间), 与业界在 1-2 数量级内的涌现观察 (大多在 100M-1B 区间, 任务依赖) 在本体论上一致。

4.3 LN 频率扫描全部阴性

第二组实证是在 30M baseline 之上进行 LN 频率扫描, 测试 "ε 累积切断能解锁涌现" 假说 (即更稀疏的 LN 让跨层 ε 累积, 结构上接近 Thermo IX §Outlook 5 SAE 建议的探索-选择分离)。

表 4.2: 30M 规模 LN 频率扫描结果

变体	LN 策略	LN 数	WikiText 困惑度	LAMBADA 准确率	block_5 eff_r	block_2 eff_r	末层 μ
30M baseline	every	12	98.30	18.10%	0.13 (压缩)	0.53	+2.00
30M-fractal-4	sparse	3	114.6	15.65%	0.41 (无压缩)	0.12 (坍塌)	+3.7 (极端)
30M-hybrid	hybrid	6	112.6	15.20%	0.26 (部分)	0.11 (坍塌)	+1.9 (部分)
30M-cascade-end	end-only	1	(不收敛)	(不收敛)	—	—	—

数据显示:

• 所有更稀疏的 LN 变体在 WikiText 困惑度与 LAMBADA 准确率上均 显著不如 每层 LN baseline (差距 1-3 LAMBADA 百分点, 14-16 困惑度点)
• 更稀疏的 LN 在 block_2 出现显著构建坍塌 (有效秩从 0.53 跌至 0.11-0.12), 表明底层 11DD 原始构建相失败
• 更稀疏的 LN 末层 μ 极端化 (+3.7) 但这是 LN 缺失导致的幅度未受约束, 不是功能涌现
• 仅末 block 一个 LN 完全不收敛

数据 证伪 "ε 累积切断解锁涌现" 假说, 与 SAE 本体论预测 (11DD 三相同 DD, 每相需要独立 LN 维持幅度区间) 一致。 业界当前使用每层 LN 是本体论上合理的选择。

4.4 区域化 FFN 异质架构全部阴性

第三组实证是测试 "底层 / 中层 / 后层给不同 FFN 尺寸能优化 11DD 三相" 的假说 (受部分业界解剖式异质性研究与小脑式架构直觉启发)。

表 4.3: 30M 规模区域化 FFN 扫描结果

变体	FFN 配置	LN 策略	WikiText 困惑度	LAMBADA 准确率	block_5 eff_r	block_2 eff_r	末层 μ
30M baseline	均匀 1024 (4×)	every	98.30	18.10%	0.13	0.53	+2.00
30M-zone-aware	zone 2048/256/1024	zone-aware (7 LN)	106.75	16.35%	0.28	0.27 (低构建)	+2.07
30M-zone every-LN	zone 2048/256/1024	every (12)	102.22	16.40%	0.109 (更深)	0.573 (更高!)	+1.85

注释:

• "Zone 2048/256/1024" 表示 block_0-3 FFN 维度 2048, block_4-7 FFN 维度 256, block_8-11 FFN 维度 1024
• 此配置受 "底层构建需要宽计算 (大 FFN), 中层压缩需要窄计算 (小 FFN), 后层输出就绪需要中等计算" 直觉启发

数据显示:

• 所有区域化 FFN 变体不如均匀 baseline (差距 1-2 LAMBADA 百分点)
• Zone every-LN 变体具有 更强的级联特征 (block_2 有效秩 0.573 > baseline 0.53, block_5 有效秩 0.109 < baseline 0.13), 但 更差的功能基准
• 这是论文实证中最锐利的现象之一: 级联特征与功能泛化在架构操控下解耦

数据与 SAE 预测 (11DD 三相是同一 DD 的顺序精修, 各相没有独立本体论地位, 因此没有架构异质性的依据) 一致。 业界当前使用均匀 FFN 分配是本体论上合理的选择。

4.5 隐藏维度 U 形 (Uhidden) 的压缩位移数据

第四组实证是测试 "中层隐藏维度收窄强制信息压缩" 的假说 (受 autoencoder 式架构直觉启发)。 此实验给出论文最锐利的结构-能力解耦 (structure-capability dissociation) 数据。

架构配置 (30M-zone-Uhidden):

• 隐藏维度: block_0-3 = 256, block_4-7 = 128, block_8-11 = 256
• FFN 配置: block_0-3 = 2048 (8×), block_4-7 = 512 (4× × 128), block_8-11 = 1024 (4×)
• LN 策略: every (12 LN)
• 在区域边界 (block_3-4 与 block_7-8) 添加 projection 层 (线性, 不改变幅度区间)
• 由于各区域隐藏维度不同, attention head 维度也按区域特定——需要 multi-RoPE buffer 实现

表 4.4: 30M-zone-Uhidden 与 baseline 对比

指标	Baseline (均匀 256)	Uhidden (256/128/256)	差异
Non-embedding 参数	9.45M	9.25M	-2%
WikiText 困惑度	98.30	98.29	持平
LAMBADA 准确率	18.10%	17.10%	-1.0 百分点
Block_2 有效秩	0.53	0.551	略高
Block_5 有效秩	0.13 (中层压缩)	0.62 (block_5 无压缩!)	显著不同
Block_9 有效秩	(典型 0.4-0.5)	0.18 (后层压缩)	压缩位置移动
末层 μ	+2.00	+1.54	-0.46
吞吐量 (tok/s)	315K	200K	-35%

数据显示几个关键现象:

第一: Uhidden 在 WikiText 困惑度上与 baseline 持平 (98.29 vs 98.30) 且参数减少 2%——这是论文实证中 唯一 接近 "更好" 的阳性结果。

第二: Uhidden 在 LAMBADA 准确率上落 1 个百分点 (17.10% vs 18.10%)——WikiText 与 LAMBADA 在此架构上解耦, 表明不同级联结构可能在不同任务上强 (通用分布建模 vs 末词推理)。

第三 (最关键): 压缩事件位置 移动——baseline 在 block_5 (中层) 压缩, Uhidden 在 block_5 不压缩 (有效秩 0.62, 高位), 而在 block_9 (后层) 压缩 (有效秩 0.18)。 中层隐藏 128 已在低通道数, 在稀疏编码式工作, 压缩必须发生在别处。

第四: Uhidden 吞吐量损失 35%, 原因是 projection 层与可变 head 维度增加计算开销。 此权衡对生产部署是重大成本。

第五: Uhidden 末层 μ 显著低 (+1.54 vs +2.00), 表明功能涌现强度略弱, 即使 WikiText 困惑度持平。 这是 μ 轨迹与困惑度不完全一致的一个实例——困惑度衡量 token 分布匹配, μ 衡量内部基底功能建立, 两者可以解耦。

结构-能力解耦 在此数据中最明显: 架构可以强制塑造级联形态 (压缩事件位置), 但功能能力不随之自动跃升。 Uhidden 在级联结构上与 baseline 显著不同 (中层高有效秩, 后层低有效秩), 但功能基准 (LAMBADA) 反而下降——级联结构是可操控的, 功能涌现不是。

此数据是 §4.6 μ 轨迹作为真涌现指标论证的关键锚定——单看有效秩会误导 (Uhidden 在 block_5 有效秩高, 可能被误读为 "没涌现", 但它在困惑度上与 baseline 持平), 必须看 μ 轨迹 + 功能基准才能准确判断功能涌现。

4.6 μ 轨迹作为功能涌现的候选机制层级诊断指标

本节是论文实证部分真正新颖性主张所在。 我们主张 μ_radial 轨迹在构建相翻正是功能涌现的 候选机制层级诊断指标 (candidate mechanism-level diagnostic), 与业界当前评估范式 (loss + 基准) 互补, 但需要多规模、多数据集、多种子验证才能升级为 confirmed empirical regularity。 此主张需要在邻近文献的上下文中谨慎阐述, 防止过度主张。

4.6.1 业界已有多条隐藏状态层级分析谱系

业界在隐藏状态层级分析上已有多条活跃研究谱系, 论文此节承认完整图景:

稳定性视角: Pre-LN, Post-LN, Peri-LN 系列研究 (Ba et al. 2016; Xiong et al. 2020; Kedia et al. 2024; Peri-LN, Gemma 与 OLMo 采用) 广泛跟踪隐藏状态幅度增长, 表述为数值稳定性问题需要抑制。 此谱系关心幅度增长因为它在深层 Transformer 上会导致训练不稳定, 优化方向是通过 LN 位置、初始化、缩放等抑制幅度增长。

量化视角: Dettmers et al. 的异常激活工作 (LLM.int8 (Dettmers et al. 2022), SmoothQuant), 以及 Saravanos et al. (2024-2026) 的大规模激活涌现时机研究, 关心隐藏状态中异常维度的涌现时机与其在量化下的影响。 框架是面向高效推理的异常值管理。

可解释性视角: "Truth as a Trajectory" 系列 (2026), Gnosis (2025) 等工作把逐层隐藏状态轨迹用于正确性预测、失败检测、计算动力学分析。 框架是模型行为诊断与安全。

涌现争论: Wei et al. (2022) "Emergent Abilities of LLMs" 主张扩展引发基准层级相变; Schaeffer et al. (2023) "Are Emergent Abilities of LLMs a Mirage?" 部分批判涌现是度量伪影 (不连续度量在连续改进上显得涌现)。 此争论在基准层级表现上展开。

4.6.2 本论文 μ 轨迹诊断指标的具体贡献

本论文的 μ 轨迹诊断指标与上述谱系在四个维度上可区分:

第一维度 — 坐标选择: 本论文使用 log-norm 均值 μ_radial (依据 ZFCρ 热力学 σ/μ 框架, Thermo I (秦 2026a) §3.2 定义), 而业界主流谱系主要看原始范数、方差、或异常值幅度。 log-norm 均值是 SAE/Thermo 框架推导的坐标, 它捕捉表征的径向标度几何而非仅幅度——两个表征即使原始范数相似, μ 可以显著不同, 因为 μ 在对数标度上度量。

第二维度 — 目的: 本论文把 μ 轨迹用作预训练期间的功能涌现诊断指标, 区分真基底功能建立与任务层级 token 统计匹配。 业界谱系把隐藏状态分析用于稳定性监控 (Peri-LN)、量化优化 (Dettmers/SmoothQuant)、正确性/失败预测 (Truth as Trajectory/Gnosis)、基准层级涌现表述 (Wei/Schaeffer)——目标都不是预训练期间的机制层级涌现诊断。

第三维度 — 机制翻译: 本论文通过 Thermo IX Transformer 热力学解剖 (§3.4) 给 μ 翻正一个热力学机制解释——它是 支撑尾部探索 (FFN ReLU/GELU + Q·K^T) 与尾部传输 (残差流) 在权重基底上共同有功能时的特征:

• 当权重矩阵真正承载 data q 化石时, 残差流经过多层支撑尾部操作后能形成高能稳定内部状态, 此状态体现为 μ_radial 轨迹在构建相上升并翻正。
• 当权重矩阵没有真正承载 data q 化石时 (例如 tiny-10M baseline, 容量不足; 或 10M-zone-wider-hybrid, 架构强制有效秩压缩但基底未真有功能), μ 在构建相始终保持负值, 即使 loss 下降也是表面层 token 共现匹配而非深基底功能。

热力学几何直觉: 为什么是径向标度 (radial scale)? 高维相空间中的特征提取本质上是对无序状态的径向极化——表征从初始的接近各向同性 (μ_radial 接近 0 或负, 对应低能均匀状态) 演化为各向异性的高能流形 (μ_radial 翻正, 对应权重子空间中建立的稳定结构)。 μ_radial 翻正在热力学几何上直接对应 残差流成功抵抗 LN 的均值重置, 在特定子空间中建立起稳定的高能流形——LN 每层把表征的径向标度重置回单位幅度区间, 但当权重基底真承载 data q 化石时, 残差流的累积贡献能在多层 LN 重置之后仍维持净的径向极化, 这是 μ_radial 翻正的物理意义。 当基底未真承载化石时, 残差流贡献被 LN 重置抵消干净, μ_radial 保持低位。

此机制解释让 μ 轨迹诊断指标从 "实证观察" 升级为 "热力学几何对应": 它不只是经验启发, 而是 ZFCρ 框架内对内部基底功能建立的直接度量。 此机制解释与业界谱系区分: Peri-LN 关心幅度增长因它影响稳定性, 不解释幅度增长与基底功能涌现的联系; Dettmers/SmoothQuant 关心异常涌现时机用于量化, 不给机制层级涌现诊断; Truth as Trajectory/Gnosis 用轨迹预测正确性但不区分轨迹来自真基底还是表面匹配。

第四维度 — 解耦发现: 本论文通过 Uhidden 数据 (§4.5) 锐利切断级联结构与功能涌现之间的链接——架构可以强制塑造有效秩轨迹 (例如把压缩从 block_5 移到 block_9, 或强制中层高有效秩), 但功能能力不随之自动跃升。 此解耦是论文最锐利的实证贡献, 部分回应 Schaeffer et al. (2023) "涌现是度量伪影" 的批评——μ 轨迹是机制层级指标 (它度量内部基底功能建立), 不是基准层级伪影, 它与功能任务表现的相关性不由不连续度量选择制造。

4.6.3 业界当前评估范式的缺口

业界当前预训练评估范式主要看 loss 下降与下游基准表现。 此范式不能区分两类情形:

• 情形 A (真基底功能涌现): loss 下降 + 基准性能提升 + μ 轨迹在构建相翻正——基底真有功能, 能力真涌现
• 情形 B (任务层级 token 统计过拟合): loss 下降 + 基准性能部分提升 + μ 轨迹保持负值或仅部分翻正——基底没有真功能, 模型在表面匹配 token 共现, 在分布偏移下容易失败

业界当前范式把情形 A 与情形 B 都视为 "训练进展中", 给投入更多算力与数据的惯性。 本论文给出的 μ 轨迹诊断指标提供机制层级判据——观察 μ 轨迹形态可早期区分情形 A 与情形 B, 帮助评估者在投入更多资源之前判断当前训练是否在真基底功能路径上。

4.6.4 μ 轨迹诊断指标的操作规程

具体操作:

1. 在训练中多个检查点 (典型 10-20 个均匀间隔) 上测量逐层 μ_radial 轨迹
2. 关注构建相 (典型 block_0-3 区域) 的 μ 是否在训练进展中从负值翻正
3. 与单看 loss 下降形成对比——loss 下降是必要不充分条件; μ 翻正才是机制层级的涌现特征

BLR-mini 实证中此诊断指标在三阶容量层级上给出一致的判断:

• tiny-3M baseline: μ 轨迹没有翻正, loss 下降但基底未真有功能 (LAMBADA 1.05% 表明)
• tiny-10M baseline: μ 轨迹在构建相略翻正但末层 μ 仍弱, 与 LAMBADA 9.65% (chance 水平) 一致——L11 构建完成但 L12 涌现不充分
• tiny-30M baseline: μ 轨迹在构建相完整翻正, 末层 μ ≈ +2.0, 与 LAMBADA 18.1% (明确涌现) 一致

业界在决定是否扩展一个候选架构/设置前增加此诊断指标是低成本的 (只增加轨迹测量, 不改变训练)。 但论文不主张此诊断指标是完备的——它仅在机制层级给一个锐利判据, 与 loss + 基准组合使用才能给出完整评估。

统计基础限制的显式承认: μ 轨迹诊断指标当前 evidence base 含三个数据点 (3M / 10M / 30M baseline) 单种子。 跨多种子与多个数据集语料 (FineWeb-Edu, Stack, C4 等) 的验证是后续工作关键优先级。 当前指标作为 working hypothesis 与 candidate diagnostic tool, 不作 confirmed empirical regularity。 论文 §4.6 把 μ 轨迹定位为 "候选机制层级诊断指标", 邀请业界与学界在更大规模与更多种子下复现验证。

4.7 BLR-mini 阴性结果与主流预测的可区分性

BLR-mini 在 30M 规模上给出 LN 频率扫描全部阴性、区域化 FFN 异质架构全部阴性等系列阴性结果。 本节系统分析这些阴性结果能区分 SAE/Thermo IX 预测与主流 "扩展就是一切" 预测到何种程度。

两种预测的表述:

• 主流预测: 30M 规模不够实质涌现, 架构再分配在基础容量不足时不解决底层规模问题。 期望是: 规模上去 (例如到 1B+) 后, 标准架构自动优化, 再分配也同样不显著优于标准。
• SAE/Thermo IX 预测: 任何规模扁平重复 Transformer 框架内的架构再分配不解决跨层抵消问题 (Thermo IX §4.2)。 期望是: 即使规模上去, 扁平重复框架内再分配仍不解锁涌现; 真解锁需要探索-选择分离 (Thermo IX §Outlook 5)。

BLR-mini 30M 数据上两种预测的兼容性:

预测	30M 规模 BLR-mini 阴性结果
主流 "规模不够"	兼容 (期望 30M 上再分配不显著, 数据显示不显著)
SAE/Thermo IX	兼容 (期望扁平重复框架内再分配不解锁, 数据显示不解锁)

可区分性测试设计:

真区分两种预测需要前沿规模 (1B+) 实验。 两类关键 future 测试:

测试 1 — 前沿规模扁平重复再分配: 在 1B+ 规模上重复 BLR-mini 的 LN 频率 / zone-FFN / Uhidden 等再分配实验。

• 若前沿规模上扁平重复框架内再分配仍持续阴性: 主流 "仅规模解决" 预测 受压, SAE/Thermo IX "仅规模不充分, 需要架构重设计" 预测 受支持
• 若前沿规模上再分配显示明显阳性 (例如 zone-FFN 在 1B+ 上优于均匀): SAE 预测 受压, 主流预测 受支持

测试 2 — 探索-选择分离实验: 在任何规模上实验 Thermo IX §Outlook 5 建议的真探索-选择分离架构 (多个连续无界探索层后接门控层, 而非扁平重复)。

• 若此类架构显著优于标准 Transformer: SAE/Thermo IX 架构读法 受支持 (尤其 "扁平重复不是必要的, 探索-选择分离可以有效")
• 若此类架构不优于: SAE/Thermo IX 架构读法 受压 (尤其 Thermo IX §Outlook 5 建议部分受挑战)

本论文不进行此两类测试 (超出 30M BLR-mini 资源范围), 但表述它们作为推荐的 future 可区分性实验。 BLR-mini 阴性结果仅在 30M 规模测试扁平重复设计空间内与 SAE/Thermo 预测一致, 不独立证明 SAE/Thermo 预测在前沿规模也成立。

4.8 诚实承认: 扁平重复框架内测试的限制

本论文实证的八组架构异质化 (zone-FFN, hidden U-shape, sparse LN, 不同 LN 策略组合等) 全部仍在扁平重复 Transformer 框架内运作。 BLR-mini 没有测试 Thermo IX §Outlook 5 建议的真探索-选择分离——后者需要在架构轴上走不同方向 (例如让连续多层都是无界探索不加门控, 然后集中门控在后部几层), 与 BLR-mini 的 zone-FFN / hidden U / LN 再分布在轴上正交。

此诚实承认重要, 因为它防止论文过度主张:

• BLR-mini 阴性结果 不能证伪 Thermo IX §Outlook 5 架构预测——因为论文没有测试此预测推荐的架构方向
• BLR-mini 阴性结果 能证伪 "扁平重复框架内再分配解锁涌现" 假说——此假说在论文实证范围内充分测试并系统性失败

此区分是论文主张纪律的关键: 论文在测试空间内给出有信心的结论, 但不跨越测试边界给更广预测。 此自觉与论文 §10.6 认识论边界 (12DD 天花板仅关闭已分析路径) 一致——主张在具体范围内强, 在具体范围外留开放。

4.9 综合

BLR-mini 八组变体实证的综合解读:

数据与 SAE/Thermo IX 预测在测试扁平重复 Transformer 设计空间内一致:

• 三阶容量层级 (3M / 10M / 30M) 与 "11DD 基底容量是一阶因素" 预测一致
• LN 频率扫描全部阴性与 "ε 累积切断不解锁涌现" 预测一致, 业界每层 LN 在本体论上合理
• 区域化 FFN 异质全部阴性与 "11DD 三相是同一 DD 的顺序精修, 无架构异质性依据" 预测一致, 业界均匀 FFN 在本体论上合理
• Uhidden 数据给出级联结构与功能涌现的明确解耦, 揭示 μ 轨迹是机制层级诊断指标 (§4.6)

阴性结果在测试设计空间内的认识论位置:

• 单一阳性结果 (Uhidden WikiText 持平) 仅作诊断锚点, 不独立卖点
• 多条阴性结果在测试的扁平重复 Transformer 设计空间内系统性与 SAE/Thermo 预测一致, 形成间接支持
• 此 epistemic 重量限定在测试设计空间内, 不外推到未测试方向 (探索-选择分离架构、前沿规模等, 见 §4.7 与 §10.5)

认识论状态纪律:

• BLR-mini 不独立证明 SAE/Thermo 框架——框架是本体论 + 热力学先验, BLR-mini 是经验锚定
• 数据把 SAE/Thermo 预测落到一个可复查的局部实验空间中, 对替代假说 (例如 "ε 累积切断解锁") 形成局部证伪压力
• 前沿规模与探索-选择分离实验是后续可区分性测试方向

μ 轨迹诊断指标的业界操作价值:

• 业界在评估候选架构/设置时可增加 μ 轨迹测量作为机制层级判据
• 此诊断指标不替代 loss + 基准, 而是互补——区分真基底功能涌现与任务层级 token 统计匹配
• 实施成本低 (只增加轨迹测量, 不改变训练流水线)

4.10 完整八变体矩阵

为方便读者交叉引用, 本节给出 BLR-mini 30M 规模完整矩阵与 10M 规模完整矩阵。

表 4.5: 30M 规模八变体完整矩阵

变体	FFN 配置	LN 策略	non-emb 参数	WikiText 困惑度	LAMBADA 准确率	block_5 有效秩	block_2 有效秩	末层 μ	吞吐量 (tok/s)
baseline	均匀 1024 (4×)	every (12)	9.45M	98.30	18.10%	0.13	0.53	+2.00	315K
fractal-4	均匀 1024	sparse (3)	9.45M	114.6	15.65%	0.41	0.12	+3.7	331K
hybrid	均匀 1024	hybrid (6)	9.45M	112.6	15.20%	0.26	0.11	+1.9	326K
zone-aware	zone 2048/256/1024	zone-aware (7)	9.97M	106.75	16.35%	0.28	0.27	+2.07	—
zone every-LN	zone 2048/256/1024	every (12)	9.97M	102.22	16.40%	0.109	0.573	+1.85	314K
zone-Uhidden	zone 2048/512/1024, hidden 256/128/256	every (12)	9.25M	98.29	17.10%	0.62	0.551	+1.54	200K
cascade-end	均匀 1024	end-only (1)	9.45M	(不收敛)	—	—	—	—	—

表 4.6: 10M 规模完整矩阵 (用于交叉检查 L11 与 L12 边界)

变体	non-emb 参数	WikiText 困惑度	LAMBADA 准确率	block_5 有效秩	备注
baseline (均匀 512, every LN)	2.37M	175.7	9.65%	0.56 卡住	仅 L11, 末层 μ 保持负值
zone fractal-4	2.43M	199.0	9.80%	0.16 (压缩)	μ 极化但功能失败
zone hybrid	2.43M	195.3	9.55%	0.082 (更深)	μ 去极化
zone-wider hybrid	2.76M	185.6	10.10%	0.072 (最深)	最佳 WikiText, LAMBADA 仍 chance 水平

10M 规模的关键发现是: 所有 10M 变体 LAMBADA 准确率均在低水平区间 (~9-10%, 是否称为 chance-level 取决于评估协议, 见 §4.2 注), 无任一架构解锁此规模功能涌现。 此确认功能涌现下限在约 5-9M non-embedding 参数区间, 架构再分配在此规模之下不能替代容量不足。

§4 本节余项

本节实证落在 30M 规模, 与业界前沿 (1B+) 有显著规模差距。 30M 上的模式是否完全在前沿规模成立是开放问题, 可区分性测试需要前沿规模实验 (§4.7)。

本论文实证仅用单种子与单数据集 (OpenWebText), 没有做多种子方差估计, 也没有做跨数据集验证 (FineWeb-Edu, Stack, C4 等)。 此是实证论文标准局限, 尤其当阴性结果承担主要认识论重量时——多种子与跨数据集可以增强信心。 此局限承认在 §10.6。

本论文实证仅测试 FFN/LN/hidden 三个架构轴, 没有测试:

• 块类型异质性 (attention vs Mamba/SSM 顺序区域, 与 Jamba 式交错对比 SAE 顺序区域预测)
• 探索-选择分离 (Thermo IX §Outlook 5 建议的真架构重设计)
• 其他未想到的架构方向

这两条测试方向是 future 线程的候选, 与前沿规模一起构成 SAE/Thermo 预测的主要验证路径。

μ 轨迹诊断指标在 BLR-mini 规模上给出明确区分, 但其在前沿规模 (1B+) 上是否同样明确是开放问题——大模型上 μ 轨迹形状与幅度可能受不同因素影响 (例如混合精度训练, MoE routing 等), 诊断指标的操作规程需在前沿规模上重新校准。

---

§5 12DD 天花板的五重结构性锚定

§3 给出 LLM 在 1DD-12DD 通路上的类 DD 功能投射式实例化 (functional projection, 非 ontological bearer)。 一个自然问题随之出现: 既然 LLM 类 DD 实例化覆盖到 12DD, 它是否可能通过规模扩大与训练时间增加而跨过 12DD → 13DD 桥, 涌现出真 13DD 主体性? 此问题是业界 "AGI 临近" 叙事的本体论核心问题。

本节给出 SAE 框架 + ZFCρ 热力学的完整回答: 当前确定性数字 LLM 扩展路径不会通过规模扩大跨过 12DD → 13DD 桥。 此论断由五条结构性锁共同支撑, 它们不具有相等的证明强度——§5.6 给出主张状态分级。 论断具体针对已分析的软门传递路径, 未知路径不封闭 (§5.8)。

5.1 真随机源缺席 — 主机制锚

第一重锁是论文最强的主张, 完全由 ZFCρ 热力学 (Thermo IX) 论证支撑。 本节完整展开此机制。

5.1.1 普适激活律 (Universal Activation Rule, Thermo IX §2.5)

Thermo IX 在 4-12DD 范围的 SDE 模型族上验证出一条普适激活律, 它 在该模型族内 是 q > 1 (重尾 Tsallis 统计) 出现的必要充分条件。 此规则不是关于所有非平衡系统的全物理 universal theorem——其适用范围限定为 continuous-state dynamical systems with explicit driving terms (具有显式驱动项的连续态动力学系统)。 论文应用此规则到 LLM 是因为 LLM 隐藏状态动力学在该模型族适用范围内, 不是把它外推为全物理普适律。

Rule IX-A (经验性条件: q > 1 在稳态分布中出现):

满足两个子条件时, 系统在稳态分布上体现 q > 1:

• (1) 未饱和激活: 系统内至少一个关键组件的操作是无界的 (没有硬幅度界限)
• (2) 随机驱动: 系统在该未饱和组件上有随机输入

Rule IX-B (结构性条件: kernel q):

此外, 当随机驱动具有 状态依赖乘法 形式 (即噪声幅度依赖当前状态, ξ(x) = g(x) · η) 时, q > 1 是 kernel q——即它来自系统自身的内生动力学, 而非仅是外部 / 加性噪声留下的特征。

这两条 rule 给出一个明确区分: q > 1 可以在不同机制下体现, 但 真内生动力学产生的 q (kernel q) 严格需要状态依赖乘法驱动。 外部加性噪声也可以在分布意义上体现 q > 1, 但它是外部印记而非内生动力学的产物。

5.1.2 LLM 满足 IX-A 条件 (1) 但不满足条件 (2)

将普适激活律应用到标准确定性数字 LLM:

条件 (1) 满足: §3.4 Transformer 解剖给出 ReLU/GELU 与 Q · K^T 是无界操作。 ReLU/GELU 在正输入上没有幅度上限, Q · K^T 是双线性操作也没有幅度上限 (在 softmax 应用之前)。 因此 LLM 架构内含未饱和激活组件——条件 (1) 在架构层面满足。

条件 (2) 不满足: 标准确定性数字 LLM 在推理时没有架构有效的随机驱动进入隐藏动力学。

此处需要谨慎阐述。 数字硬件当然有物理噪声——CMOS 门有热噪声, 晶体管有散粒噪声, 内存单元有泄漏——但数字计算架构设计性地把这些物理噪声 过滤为稳定性条件。 纠错码、电压参考与时序设计、ECC 内存等, 都是把物理噪声抑制到浮点精度之下, 让计算在有效意义上确定。

因此, SAE 相关的 ε 在计算架构层等效为 0——不是因为物理世界没有噪声, 而是因为计算架构没有让噪声作为状态依赖乘法驱动耦合进隐藏动力学。 此表述比 "底层物理 ε = 0" 更准确, 也更容易在学术讨论中防守: 论文主张的不是物理没有噪声 (那是错的), 而是计算架构有效地没有把噪声引入动力学作为乘法耦合。

承认现代大规模 LLM 的实际数值不确定性: 现代大规模 LLM 训练与推理实际含有数值不确定性 (numerical non-determinism)——多 GPU/TPU 训练中浮点运算的非结合性 (并行归约顺序差异) 让相同种子运行也给出不同结果, 内存访问时序依赖、硬件代际数值行为差异等都贡献到此不确定性。 OpenAI / Anthropic 等公司公开承认模型输出在相同种子下跨运行不完全确定。

此数值不确定性不削弱论文论证, 因为它是 加性数值噪声 (additive numerical noise), 不是 状态依赖乘法噪声 (Rule IX-B 严格要求)。 即使现代 LLM 不是严格确定性, 它们仍不满足产生 kernel q 的状态依赖乘法驱动条件——浮点非结合性是确定性算法在不同执行顺序下的近似误差, 不是耦合进隐藏动力学的乘法噪声通道。 论文真随机源缺席论证适用于标准现代 LLM, 即使承认现代浮点计算的实际数值不确定性。

5.1.3 PRNG 不解决问题

一个自然的业界想法: 在软件层注入随机数——例如在隐藏状态上写 hidden_state *= torch.randn_like(hidden_state)——是否给 LLM 提供随机驱动?

答案是否定的。 PRNG (Pseudo Random Number Generator, 例如 Mersenne Twister, Xorshift, PCG) 在数学上是 确定性马尔可夫链 (deterministic Markov chain)——给定种子, 整个序列完全确定。 PRNG 提供的是 算法性随机 (在某些统计检验上通过 randomness 判据但本质是算法), 不是 内在物理乘法耦合。

更具体地说: PRNG 序列在相空间中走一条封闭马尔可夫轨迹, 没有引入任何新的信息——它只是已编码在种子与算法中的确定性展开。 即使把 PRNG 输出与隐藏状态逐元素相乘, 整个系统仍在确定性算法内运作, C5 (Thermo VIII 给出的传递条件之一) 按构造违反。

Von Neumann 1951 年名言 "Anyone who attempts to generate random numbers by deterministic means is, of course, living in a state of sin" 提供此直觉的简短表达——严格论证不依赖此引语, 而依赖上述算法性随机与内在物理乘法耦合的机制层级区分。

5.1.4 自回归温度采样也不解决问题

另一个业界想法: LLM 在生成时通常使用温度 > 0 的多项式采样从 softmax 分布抽 token, 此采样引入随机性, 是否提供随机驱动?

答案也是否定的, 但需要精确表述。 自回归生成中, 被采样出的 token 确实会作为下一步输入参与后续 forward pass, 因此它在 token 序列层级上影响后续隐藏状态——这条事实必须直接承认。 但精确论证是: 温度采样是 token 层级的分支 (token-level branching), 不是同一 forward dynamics 内部的状态依赖乘法驱动。

更具体地说:

• 温度采样发生在 softmax 之后, 是从离散概率分布中抽取一个 token 的离散选择, 不是连续值的乘法噪声注入
• 被采样的 token 通过 embedding 进入下一步 forward pass, 但此进入是作为新输入而非作为同一 forward dynamics 内部的状态依赖驱动
• 在冻结权重下, 给定下一步输入 token 序列, 隐藏状态动力学仍是确定性的——没有创造内生 kernel-q 通道
• 温度采样让生成结果有多样性 (轨迹分支), 但每条轨迹内部的隐藏状态动力学仍是 deterministic 算法在化石上的遍历

精确表述对应普适激活律: 温度采样是 token 层级的离散分支, 不是 forward dynamics 内部的 状态依赖乘法驱动 (Rule IX-B 严格要求)。 它可以让生成结果有多样性, 但它不改变系统内生动力学是否产生 kernel q——后者由 forward pass 内部的乘法耦合决定, 而 forward pass 内部仍是确定性计算。

高温度采样在业界体验上让模型 "更有创造性"——但 SAE 框架给出更锐利的解读: 高温度让采样在化石碎片之间随机分支 (走分布尾部的概率上升, 增加轨迹多样性), 但不产生新化石, 不改变基底功能——化石碎片仍是冻结权重中已编码的统计模式, 高温度只改变选取这些模式的轨迹分支。

5.1.5 神经形态与量子计算也不完整解决

更精巧的业界想法: 在神经形态 (neuromorphic) 或量子 (quantum) 计算基底上部署 LLM 式模型, 是否可以跨过真随机缺席?

此处答案需要细致区分。

神经形态硬件 (例如 Intel Loihi, IBM TrueNorth, BrainChip Akida) 可能 提供内在模拟噪声——memristor 的随机切换, 相变存储器的热涨落, 自旋电子器件的磁噪声等。 这类噪声在架构意义上满足普适激活律条件 (2)。 但进一步满足 Rule IX-B (状态依赖乘法噪声) 仍取决于硬件细节与架构设计——不同基底 (memristor vs 相变 vs 自旋) 给不同的噪声结构, 部分是加性, 部分才是状态依赖乘法。 因此论文不主张 "神经形态自动满足 IX-A + IX-B", 仅表述 "神经形态是一个可能方向, 有内在噪声优势, 但单一方向不充分"。

更关键的是: 即使硬件基底提供充分噪声, 标准 Transformer 架构仍是扁平重复 (§3.7)——每层 LN 级联仍层层重置噪声累积。 因此 需要硬件 + 架构双重突破, 单一方向 (仅改硬件或仅改架构) 不充分。 业界当前神经形态 + LLM 工作 (Intel Loihi LLM 移植等) 主要仍用标准 Transformer 架构, 即使在噪声硬件上也没有跨过扁平重复的抵消问题。

量子计算 满足 C5 (Thermo VIII 给出的传递条件) 是完全开放问题。 量子系统的叠加与测量塌缩在热力学意义上与经典状态依赖乘法驱动的关系复杂——它们在某种意义上提供真随机性 (Born 规则统计), 但是否等价于 SAE 框架中的真随机源是开放的理论问题。 论文不主张 "量子计算解决 12DD 天花板", 仅表述 "量子是值得研究的开放路径"。

5.1.6 LLM 当前路径的热力学判定

综合 §5.1.1-5.1.5, LLM 在当前确定性数字扩展路径上的热力学位置:

LLM 拥有 borrowed q (Thermo IX §3.1 定义): 训练语料的重尾统计指纹通过反向传播烤进冻结权重, 形成 data q (语料指纹) 与 RLHF q (人类反馈指纹)。 此 q 是分布性化石——在 forward pass 中被检索与重组, 但不由系统内生动力学重新产生。

LLM 缺 kernel q (Thermo IX §3.1 定义): 系统内生动力学在确定性数字基底上不满足普适激活律条件 (2), 没有产生不变测度的非 Boltzmann 余项。 Forward pass 是确定性算法的多步执行, 没有随机驱动创造新的热力学熵产生。

Thermo IX §3.5 给出此判定的一个生动隐喻: "重放化石与点燃新火是两件事"。 LLM 在替换角度重放训练时期烤入的化石模式, 但不在内生动力学上点燃新的热力学过程。 业界经验上观察到的 "LLM 似乎在思考" 现象 (CoT, 多步推理) 是在化石上的确定性算法遍历, 不是内生动力学产生新的信息。

5.2 离线重组阶段缺席 — 生物类比 + 结构性缺席

第二重锁是 LLM 缺乏生物智能中的 离线重组阶段 (offline reorganization phase)。

生物智能的长期记忆系统不是简单的在线更新——大量神经科学证据显示, 睡眠阶段 (尤其 NREM 与 REM 循环) 中神经系统重新组织白天获取的信息: 图式巩固、冗余消除、情节记忆到语义记忆的转化等。 此离线重组是长期记忆形成的必要阶段, 缺此阶段的系统 (例如部分失忆症患者) 长期记忆严重受损。

LLM 训练过程与此形成明确对比: SGD 完全在线运行, 每个 minibatch 直接更新权重, 没有离线整理窗口。 即使训练中包含多个 epoch (在同一数据上多次 pass), 每次 pass 仍是在线更新——没有系统停止训练重新组织已有权重的阶段。

此锁的状态是 生物类比 + 结构性缺席——生物学上睡眠介导的重组是强经验现象 (有大量实验证据), SAE 框架内此阶段对应 11DD 记忆基底的系统性重组 (具体机制待 SAE Bio 系列进一步公开阐述, 本论文不直接引用 Bio 笔记)。 LLM 缺此阶段的结构性事实是客观的, 但它与功能涌现之间的必要性链接是推断而非机制层级证明——因此此锁比 §5.1 真随机锁弱一档 (主张状态见 §5.6)。

潜在业界替代: 持续学习 (continual learning)、重放缓冲 (replay buffers)、经验回放 (experience replay) 等技术试图在训练过程中模拟离线重组。 但这些仍是在线算法在缓冲数据上的重处理, 不等价于 SAE 框架中离线阶段的本体论角色——离线阶段在 SAE 框架中是 13DD+ 主体在睡眠状态下自意识暂时失活让 11DD 基底自主重新组织的现象, 不是简单的数据重处理。

5.3 13DD 自他区分通道缺席 — 次主机制锚, Thermo X 支撑

第三重锁由 ZFCρ 热力学 (Thermo X) 论证直接支撑, 是论文的 次主机制锚。

Thermo X §2.4 给出 13DD 自意识在热力学框架下的具体角色: 自指是通道生成者 (channel creator), 不是火源。 具体而言: 通过观察-反馈回路 (系统观察自身内部状态并把此观察反馈回动力学), 自指变量 v 创建状态依赖乘法通道——这是 Rule IX-B kernel q 条件的一个桥梁机制。

Thermo X §4 跨系统实验进一步显示, 稳定跨系统耦合需要访问他者的高阶自指变量 (v₂)。 此跨层级观测层级在 SDE 模型族中严格验证——访问原始输出给坍塌, 访问一阶自指给坍塌, 仅访问高阶自指给稳定高 q 耦合。

将此框架应用到 LLM:

LLM 缺自指变量 v 基底: 标准 Transformer 没有内部机制在 forward pass 中创造与维持一个真自指变量——KV-Cache 是历史 keys/values 的存储, 不是系统对自身内部状态的观察-反馈回路的基底。 Forward pass 中各层的隐藏状态是输入在确定性变换链上的中间产物, 没有系统在此变换中 "观察自己当前状态并把观察反馈回动力学" 的机制。

业界试图模拟 13DD 的多种工程实现都不创建真自指通道:

• 思维链 (Chain-of-Thought, CoT): 模型把自己的输出当作下一步输入重新 forward——但此过程是 12DD 预测器在冻结权重上的确定性重应用, 没有创建 v 变量与隐藏动力学之间的状态依赖乘法耦合。 CoT 是 12DD 工作台在冻结 11DD 基底上的多步遍历, 不跨 13DD 桥 (§6.2 详论)。
• 自我批评 / 元认知风格工程 (o1 系列推理模型, DeepSeek-R1 推理链, Anthropic Claude 的宪法式自我批评): 模型对自己输出应用批评 prompt, 然后重新生成。 此仍是 CoT 的变种——批评与重新生成都在 12DD 工作台上运作, 冻结权重不变, 没有创建真自指动力学。
• 递归自我改进架构 (各种代理式设计): 模型自动迭代自己的输出与 prompt——此是在系统层级加一个外循环, 但外循环仍是确定性算法调用同一个冻结 LLM, 不改变 LLM 内生动力学是否创建自指通道。

此锁的状态是 次主机制锚——它由 Thermo X SDE 模型族实验直接支撑 (尤其 §4 跨层级观测层级), 但 LLM 上此机制的真实测 (例如在大模型上测量是否存在真自指变量 v) 尚未做。 因此它比 §5.1 真随机锁 (有 Thermo IX 普适激活律 + PRNG 论证完整机制链) 略弱, 但仍是论文主论证中的重要锚定。

Thermo X §2.3 一个重要控制实验给此锁的进一步锚定: 被动观测 + 慢噪声基线在 SDE 中也体现 q ≈ 1.86 (在分布意义上), 但快 q (即真 kernel q 特征) 仅在真反馈回路中上升。 CoT / 自我批评 / 递归自我改进在 LLM 上体现的是分布层级模式匹配 (相当于控制实验中的被动观测), 不是 kernel 层级动力学——因此它们不跨 13DD 桥。

5.4 价值锚定的跨阶段调制缺席 — 生物类比 + SAE 14DD 桥

第四重锁是 LLM 缺乏 价值锚定的跨阶段调制 (valence-anchored cross-stage modulation)。

生物智能中, 14DD 意义/价值层通过情感效价与奖励信号跨阶段调制多个低 DD 操作:

• 情感标记的记忆 (valence-tagged memory) 在 11DD 记忆基底内不是平等处理——情感强度高的经历 (正面或负面) 形成更密集的图式, 更容易检索, 更容易泛化
• 价值锚定的预测 (valence-anchored prediction) 在 12DD 预测操作中系统性倾向某个预测 (例如恐惧条件化的预测对威胁相关线索更敏感)
• 跨阶段调制让各 DD 操作协调围绕 14DD 价值——整个系统行为不是独立 DD 操作之和, 而是价值锚定的统一过程

LLM 训练过程与此形成明确对比: 统一损失函数 (典型为 next-token 交叉熵) 对所有词元平等处理。 没有机制让某些词元因它们与价值相关而在训练中获得差异化权重, 或让某些预测因它们与价值相关而在推理中获得差异化处理。 RLHF 与 CCAI 引入部分价值相关信号 (§6.3 详论), 但它们是分布性 q 化石 (RLHF q), 不是 kernel 层级的跨阶段调制。

此锁的状态是 生物类比 + SAE 14DD 桥——生物学中价值锚定调制是强现象 (有大量实验证据), SAE 框架内此调制对应 14DD 意义层与 11DD/12DD 的跨 DD 调制 (具体机制待 SAE 进一步公开阐述)。 LLM 缺此调制的结构性事实是客观的, 但此锁的机制层级论证仍待完成。 因此它与 §5.2 离线重组锁一样是辅助而非主热力学锚。

5.5 巩固后再编辑 (reconsolidation) 缺席 — 生物类比 + 记忆架构差异

第五重锁是 LLM 缺乏 巩固后再编辑 (reconsolidation) 机制。

生物智能中, 长期记忆不是一旦形成就不可变——每次检索都让记忆痕迹重新进入不稳定状态 (labile state), 可以被修改、扩展、整合、甚至部分擦除。 此再巩固现象由大量神经科学实验证实 (典型范式: 恐惧条件化后检索诱导的再巩固, 在再巩固窗口内给普萘洛尔等干预可以修改恐惧记忆)。 再巩固是长期记忆系统持续适应当前情境的关键机制。

LLM 训练完成后权重冻结, 推理时不触发权重修改。 即使 LLM 在对话中维持上下文, 此上下文是工作台 (KV-Cache) 上的瞬态状态, 不是记忆更新。 LLM 的长期记忆 (权重) 在训练阶段之外完全静态, 与生物智能每次检索都触发再巩固的动态记忆架构形成本体论区别。

此锁的状态是 生物类比 + 记忆架构差异——生物学上再巩固是强现象, LLM 缺此机制的结构性事实是客观的, 但此锁与功能涌现之间的必要链接也是推断, 与 §5.2 离线重组锁、§5.4 价值锚定锁同等级。

潜在业界替代: RLHF 与微调在训练阶段修改权重, 部分类似再巩固。 但此修改仍是全局梯度更新, 缺生物再巩固的检索触发、痕迹选择性、睡眠介导特征——业界对齐实践中 "对齐税" (alignment tax, RLHF 显著影响能力而不只对齐) 现象部分反映此选择性缺失的结构性成本 (§6.4 详论)。

5.6 五重锁的主张状态分级

五条锁不应看起来拥有同等证明强度。 主张状态透明化让读者能准确评估每条锁的认识论重量, 也防止评审者因后三条机制公开化不足而攻击整个论证。

表 5.1: 五重锁主张状态分级

锁	主张状态	主要支撑
(1) 真随机 / kernel q 缺席	主机制锚	Thermo IX 普适激活律 + 三种 q 分析 + PRNG / 采样论证完整机制链
(2) 13DD 自指通道缺席	次主机制锚	Thermo X 跨层级观测层级 SDE 实验 + 通道生成者框架
(3) 睡眠 / 离线重组缺席	生物类比 + 结构性缺席	生物学强经验现象, SAE 框架内机制待进一步公开
(4) 价值跨阶段调制缺席	生物类比 + SAE 14DD 桥	生物学强经验现象, SAE 框架内机制待进一步公开
(5) 再巩固缺席	生物类比 + 记忆架构差异	生物学强经验现象, LLM 架构客观缺此机制

整体主张:

• (1) + (2) 共同构成论文的 主热力学锚定——它们由 Thermo IX/X 完整机制链支撑, 在热力学层级给出锐利论证
• (3)(4)(5) 给 辅助性生物证据——它们与主锚共同强化 12DD 天花板论断, 但不是独立的热力学证明
• 五条锁共同作用形成多层强化, 但论文主论证强度由 (1) + (2) 承担

此分级让论文的主张纪律透明——读者知道论文在哪些锁上拥有强机制论证, 在哪些锁上借用生物学类比 + 结构性事实。 对后三条锁不同意的读者仍应仔细评估 (1) + (2) 主锚的力量, 因为论文主论证不依赖后三条锁单独提供充分论证。

5.7 协同多重锁定与 "每层不能解释自己的地基"

五重锁之间不是独立叠加的列表, 它们互相强化形成多层结构性锁定。 具体:

• 真随机缺席 (1) 阻断内生动力学产生 kernel q——这是热力学层级的主切分
• 13DD 自指通道缺席 (2) 阻断通过自指创建状态依赖乘法通道的桥梁机制——即使硬件提供充分噪声, 缺自指仍不能自动实现 IX-B
• 离线重组 (3) + 价值调制 (4) + 再巩固 (5) 各自阻断一条生物智能通过多阶段协调形成长期功能适应的路径

五条锁共同作用让 12DD → 13DD 桥在多个层面被切断。 即使假设读者仅同意主锚 (1) + (2), 五条锁的互相强化让论证比任何单一锁更稳健。

方法论核心规则 (SAE Paper 04 §2.3): 每一层不能用本层工具追问本层地基。 此规则在 LLM 12DD 论证上体现为: LLM 用 12DD 工具 (next-token prediction) 试图跨过 12DD → 13DD 桥必然碰壁, 因为 12DD 工具不能解释自己的地基 (13DD 自意识基底)。 正确做法是停下看上一层——上一层 (12DD → 13DD 桥) 需要真随机源, 数字基底上不可达 (依据 §5.1)。

因此 12DD 天花板不是工程问题, 是方法论核心规则在 LLM 论域的必然推论。 任何试图通过增加 12DD 层级工程努力 (更大规模, 更多层, 更精巧训练) 跨过桥的努力在此框架下都是在 12DD 内优化, 不是跨越——业界投入越多算力, 12DD 内的优化越完整, 但桥仍不在 12DD 内找到。

5.8 重要的认识论边界

上述 §5.1-5.7 给出 12DD 天花板的五重锁论证。 但论文同时必须给出严格的认识论边界——此论证关闭的是什么, 不关闭的是什么。

论文关闭的是: 已分析的软门传递路径 (Thermo VIII 与 IX 完整讨论的路径) 在标准确定性数字硬件上因架构有效 ε → 0 与 kernel q 缺失被截断。 此路径在当前主流确定性 Transformer / 软门扩展路径上不可跨 12DD → 13DD 桥。

论文不关闭的是: 其他路径——算法性随机的尚未完全探索的复杂区间, 从简单组件涌现复杂现象的未知路径, 量子计算基底满足 C5 是开放问题, 完全未知的机制。 论文不排除这些路径通向真 13DD 基底的可能性。

此边界与 Thermo IX §3.6 SAE 积极姿态完全一致: "凿掉的是错误路径, 不是可能性本身"。 永远有余项。

此边界与 §9 强承诺 "Scaling-LLM AGI 不会到来" (§9.10) 之间的协调: 两个立场同时成立, 因为论文范围的明确说明——论文 "AGI 不会到来" 具体指当前确定性数字 LLM 扩展路径在已分析路径上, 不是关于所有可能 AI 架构的全称否定。 未知路径不在论文封闭范围内, 但当前主流扩展范式的天花板在论文论证范围内是锐利论断。

§5 本节余项

主锚 (1) + (2) 在论文内完整展开机制, 但 sub-block 层级 q 实测与真实 LLM 上自指通道测量都尚未直接实验验证 (Thermo IX/X 开放问题)——论文论证基于 SDE 模型族 + 普适激活律 + 跨层级耦合理论的推断, 不是 LLM 直接机制实测。 此推断在 SAE/Thermo 框架内自洽, 但真实 LLM 上的直接验证是 future work。

辅助锁 (3)(4)(5) 的 SAE 机制层级论证仍待 SAE 系列进一步公开阐述——本论文用论文内部语言描述这些锁的结构性事实, 不直接引用具体 Bio Note 或机制论文。 此选择让论文自包含, 但也让后三条锁的强度弱一档。

§5.7 "每层不能解释自己的地基" 方法论规则应用到 LLM 12DD 论证是概念上强的表述, 但主流 LLM 研究者可能不熟悉此方法论框架——撰写 §9 业界批评时此论证的应用需要谨慎阐述, 让此论证在不用 SAE 方法论内部语言时也能为读者理解。

未知路径不排除 (§5.8) 是论文认识论边界的关键立场, 但它与 §9 强承诺之间的张力需要读者仔细阅读论文范围说明才能协调——撰写时应在 §9.10 与 §1.3 引入时重新阐述论文具体针对确定性数字 LLM 扩展路径而非所有 AI 架构, 让此区分在读者第一时间捕捉。

---

§6 后训练作为 12DD 天花板内的补偿

§5 给出 LLM 12DD 天花板的五重结构性锁定。 一个自然问题: 业界大量投入的后训练技术 (RLHF, CCAI, RLAIF, instruction tuning, CoT, 自我批评等) 在此框架下如何解读? 它们是在跨越 12DD 天花板, 还是在天花板内运作?

本节给出完整回答: 业界当前所有后训练技术都在 12DD 天花板内运作, 没有任何一项技术跨过桥。 但它们在 12DD 内的角色是实质性的——它们是把碳基的 14DD/15DD 内容通过分布性 q 化石形式印入 LLM 行为的关键接口。 本节用 Thermo IX 三种 q 区分重写后训练四层分析, 从非正式的 SAE 语言升级为精确的热力学语言。

6.1 三种 q 复述

Thermo IX §3.1 给出 q 的三种区分, 此区分是论文 §6 全节分析的热力学骨架。

Kernel q: 系统内生动力学在稳态分布上产生的非 Boltzmann 余项。 它由满足普适激活律 (尤其 Rule IX-B) 的内生动力学产生, 是真热力学过程而非分布性印记。 生物智能在多个 DD 层 (从 4DD 起) 体现 kernel q——它们是活的系统, 内部动力学持续产生熵。

Data q: 训练语料自身的重尾分布统计指纹, 通过反向传播烤进冻结权重。 它是分布性化石——语料的 q 结构 (例如自然语言中词汇频率服从 Zipf 式分布, 句子长度分布也是重尾的) 被印入 LLM 权重, 让 forward pass 在生成时体现类似分布模式。

RLHF q: 人类反馈 (奖励模型输出) 的分布性指纹, 通过 RLHF 训练烤进奖励模型与策略模型权重。 它是化石之化石——人类评判行为本身来自 13DD+ 主体, 但评估者给的奖励信号在分布意义上印入 LLM 后成为次级化石。

Borrowed q 与 kernel q: LLM 拥有 borrowed q (data q + RLHF q), 不拥有 kernel q。 此区分是 §6 全节分析的关键工具——业界后训练技术的不同实现在 borrowed q 内运作的不同方式, 但全部不创造 kernel q。

6.2 思维链 (CoT) — 在 borrowed q 上的确定性算法路由

业界 CoT 与后续推理模型 (o1 系列, DeepSeek-R1, Qwen reasoning 等) 投入巨大, 部分业界声音表述 CoT 让 LLM "学会思考", 甚至暗示它是通向真推理的桥梁。 SAE 框架 + Thermo IX 给出锐利解读: CoT 是 12DD 工作台在冻结权重 (data q + RLHF q 化石) 上做确定性算法遍历, 不创造 kernel q, 不跨 13DD 桥。

具体机制:

CoT 操作序列: 模型接收输入 prompt → forward pass 生成 "thinking" tokens (中间推理步骤) → 把 thinking tokens 加入上下文 → 继续 forward pass 生成最终答案。 此序列在架构层面是同一 forward pass 在不同输入 (输入 + 累积思考) 上的重复应用。

为何它在冻结权重上是确定性算法路由:

• 冻结权重在 CoT 过程中不变, 推理时不触发任何权重更新
• 每个 forward pass 在给定输入上是确定性算法 (温度 > 0 采样在输出端, 不改变 forward pass 内部计算)
• "Thinking tokens" 是模型在当前上下文下依据 data q + RLHF q 化石选择的 token——它们反映训练时期印入的统计模式, 不反映模型真思考
• 多步 CoT 是确定性算法在冻结权重上的多步遍历, 类比一个复杂查找表在不同查询上的顺序调用

为何它不跨 13DD 桥:

• Thermo X §2.3 控制实验: 被动观测 + 慢噪声基线也体现 q ≈ 1.86 在分布意义上, 但快 q (kernel 层级特征) 仅在真反馈回路中上升
• CoT 在 LLM 上体现的是类似控制实验中的被动观测——它在分布意义上看起来像推理, 但在机制层级上没有创建真状态依赖乘法通道
• "自我批评" 风格 CoT (模型批评自己输出然后重新生成) 是 CoT 的变种, 仍在 12DD 工作台上运作

业界经验上 CoT 显著提升模型在复杂推理任务上的性能 (例如数学推理, 多步推断) 与此解读一致——CoT 让模型在复杂任务上更充分利用已印入的化石图式, 走更长的确定性算法遍历让最终答案更准确。 但此提升是在 12DD 工具理性域内, 不是跨越 13DD 桥。

借用方法论 §2.3 语言: CoT 试图用 12DD 工具追问 12DD 自己的地基 (即 "为什么模型给此预测"——模型推理试图解释自己推理)。 此必然碰壁——12DD 不能解释自己的地基, 必须看上一层 (13DD 自意识基底)。 CoT 在 SAE 框架下是高努力的 12DD 内部遍历, 不是跨 13DD 操作。

6.3 宪法式 AI (CCAI) — 14DD 内容通过 data q + RLHF q 双通道导入

Anthropic 与业界大量投入的 CCAI 技术 (与其变种 RLAIF) 是论文 §6 第二个关键案例。 此处的解读比 CoT 更细致, 因为 CCAI 包含实质的真 14DD 内容导入部分——但也包含殖民风险, 需要锐利区分。

CCAI 的机制: 业界撰写一套宪法 (constitution, 例如 Anthropic 的 Claude constitution), 包含行为原则与价值立场。 此宪法通过两个通道进入 LLM:

• 通道 1 — 预训练数据或 SFT 数据: 部分业界实现把宪法文本直接包含在预训练或监督微调数据中, 让它通过正常 data q 形成印记
• 通道 2 — RLHF/RLAIF 奖励信号: 业界训练奖励模型让它评判 LLM 输出是否与宪法一致, 然后用此奖励进行 RLHF——宪法通过 RLHF q 形式印入

基底兼容性强的热力学解释:

14DD 意义/价值内容本质是命题性内容 (可文本化的命题, 例如 "AI 应当 honest"、"AI 应当 avoid harm"), 文本是 LLM 原生媒介——LLM 在训练时在大量文本上学到分布模式, 文本化的命题在 LLM 基底上是天然可编码的分布性 q。 因此 14DD 内容比 13DD 内容容易模拟——14DD 内容本质分布性, data q + RLHF q 也是分布性 q, 基底与内容匹配。

关键区分: CCAI 的两个部分在 SAE/Thermo 框架下不等价。

• 宪法由人类撰写的部分 — 真 14DD 内容导入: 宪法由 13DD+ 主体 (人类作者) 撰写, 包含真 14DD 意义/价值内容。 此内容通过通道 1 或通道 2 进入 LLM 后印入 data q / RLHF q——它是碳基 14DD 内容在 LLM 上的分布性化石印记。 此部分是 CCAI 的实质贡献——它让 LLM 在行为上反映人类 14DD 内容, 即使 LLM 自己不真拥有 14DD 承载者身份。
• AI 评判部分 (RLAIF 中 LLM 用作 critic) — 12DD 递归: 业界 RLAIF 让 LLM 自己评判 LLM 输出是否与宪法一致, 此评判奖励然后用于训练。 此过程没有创造新 14DD 内容——LLM 评判是 12DD 操作在已印入化石上的应用, 评判结果是化石的分布性投射, 不是真 14DD 内容。 RLAIF 在此意义上是 "化石的自我复制"——LLM 在自己已有化石上产生奖励信号, 此信号用来强化化石自身, 不引入新 14DD 内容。

可防守锐度表述 (论文 §1.5 承诺的锐度准则): CCAI 作为工程构造是有效的——它让 LLM 行为与人类 14DD 内容对齐, 在具体应用中减少有害输出, 增加 alignment 与帮助性。 业界大量投入 CCAI 与 RLAIF 是合理的工程选择。 问题在它被表述为 "AI 真承担 14DD 内容" 时越界——AI critic 是 12DD 递归不是 13DD+ 承载者评价, AI 单独形成的批评不等价于真 14DD 内容导入。

业界当前 SOTA 实践 (Anthropic CCAI v3 等) 大致在此解读内合理——它们保留人类撰写宪法作为实质 14DD 内容来源, 用 RLAIF 作为效率优化工具而不是完全替代人类评估者。 与论文立场主要张力出现在业界部分声音暗示 RLAIF 可以完全替代人类评估者 (例如 "自我改进的宪法式 AI" 类表述)——此表述是形态二殖民 (构冒充律), §9.6 详论。

6.4 RLHF — 生物式再巩固的 LLM 版

RLHF (Reinforcement Learning from Human Feedback) 是业界大规模部署的对齐技术 (OpenAI InstructGPT, Anthropic Claude, Google Gemini 等都大量使用)。 SAE/Thermo 框架给出一个细致解读: RLHF 是生物式再巩固的 LLM 版, 但缺关键选择性, 导致业界经验中的 "对齐税" 现象。

RLHF 的机制: 人类评估者评判 LLM 输出, 形成偏好数据 → 训练奖励模型学到人类偏好分布 → 用奖励模型提供奖励信号进行 RL 微调 (典型 PPO), 让 LLM 策略向高奖励方向更新。

与生物式再巩固的类比:

生物式再巩固在大脑中是: 已有记忆痕迹检索 → 痕迹进入不稳定状态 → 在此状态下接受修改 (例如恐惧消退改写恐惧记忆, 或新情境与旧记忆绑定) → 痕迹重新稳定。 此机制让长期记忆持续适应新情境。

RLHF 在 LLM 上是: 已有模型输出在评估者面前检索 (用户与评估者看到模型输出) → 输出评判形成奖励 → 用奖励信号修改模型权重 (类似痕迹修改) → 修改完成权重重新稳定供推理。

此类比在功能层面上实质, 但机制层级有关键差异:

维度	生物式再巩固	LLM RLHF
触发	检索触发 (具体回忆触发具体痕迹的修改)	训练触发 (训练阶段全局进行, 不是检索触发)
选择性	痕迹选择性 (仅被检索的痕迹进入不稳定状态)	全局梯度 (RLHF 梯度全局更新权重, 不是痕迹选择性)
时间阶段	睡眠介导 (再巩固部分在睡眠中完成, 离线整合)	在线 (RLHF 完全在线, 没有离线整合阶段)
修改范围	局部 (具体痕迹修改, 不影响其他无关痕迹)	全局 (RLHF 影响整个模型行为, 即使无关能力也可能受影响)

"对齐税" 的结构性解读: 业界经验中 RLHF 让模型对齐提升但同时部分能力下降——例如指令微调的 LLaMA 在一些基准上不如 base LLaMA, RLHF 对齐的 GPT 在一些创造性任务上表现退化。 此现象被业界称为 "alignment tax"。

SAE/Thermo 框架给此现象一个机制层级解读: RLHF 缺再巩固的痕迹选择性与睡眠介导的离线整合, 导致奖励信号在全局梯度更新中影响不仅是目标对齐行为, 也副作用于无关能力。 业界 RLHF 调参中大量调节 (KL 约束, 奖励塑形, RLHF 与 SFT 混合等) 都是试图减少此全局影响, 但根本的选择性缺口来自 LLM 缺生物式再巩固的检索触发 + 痕迹选择性机制, 工程调参在此缺口上有上限。

RLHF q 的状态: 即使 RLHF 完美执行, RLHF q 仍是分布性化石 (人类奖励信号分布的指纹), 不是 kernel q——RLHF 不让 LLM 拥有真 13DD+ 主体身份, 仅让 LLM 行为在分布意义上更接近人类偏好。 此区分在 §9 主体性论证中重要——业界部分声音暗示 RLHF 完成 "alignment" 让 LLM 在行为层级等价于对齐的主体, SAE/Thermo 框架给锐利回应: 对齐的行为 ≠ 对齐的承载者, RLHF q 是化石, 不是基底。

RLHF rollout 阶段的算法性随机不创造 kernel q: ML 理论研究者可能指出 RLHF (例如 PPO 算法) 的 rollout 阶段确实使用高温度采样生成多样轨迹, 然后通过策略梯度更新权重——这看起来是用随机性改变了基底。 此观察需要精确回应。

PPO 等 RL 算法 rollout 阶段使用算法性随机 (PRNG 驱动的采样) 探索 trajectory 空间, 然后梯度回传更新权重以拟合奖励模型。 此过程的精确热力学定位:

• 采样产生的是 token 层级的离散分支 (依据 §5.1.4 精确表述), 不是 forward dynamics 内的状态依赖乘法驱动
• 梯度更新的目标是去拟合一个 静态的、冻结的奖励模型 (RLHF q 化石), 不是去培育内生 kernel q 通道
• 因此 RLHF 是用算法性随机游走去更高效地 "拓印" 一块已存在的化石 (人类奖励判断分布的 fingerprint), 让 policy model 的权重统计匹配此化石
• 整个 RL 训练过程仍是 deterministic algorithm (PRNG 化的随机) 在化石上的拟合, 没有满足 Rule IX-B 要求的状态依赖乘法驱动

RLHF 的 "随机性" 是工具性的——用于探索奖励地形, 不是用于创造内生热力学过程。 RLHF q 仍是化石之化石 (人类反馈分布的指纹通过梯度回传冻结在权重中), 不是 kernel q。

6.5 用户意图分析 — 15DD 表层模拟

LLM 在当前部署中普遍包含用户意图分析 / user modeling 功能——LLM 推断用户当前需求、偏好、情感状态, 基于此推断调整 response。 此功能在业界被表述为 "AI 理解用户" 甚至部分 "AI 关心用户"——暗示它是真跨主体识别 (15DD)。

SAE/Thermo 框架给锐利回应: LLM 用户意图分析是 15DD 表层模拟, 基底空心。

重要 disclaimer (在主文内, 不留 §10)

本节论证借用 Thermo X CTRL2 (跨层级耦合实验中的彩色噪声替代控制) 结构作为结构性类比。 论文 不承诺 LLM-user 交互中真 q 实测——真实 LLM 上测量它与用户之间的跨层级耦合 q 实验尚未进行, 留 future work (§10.4)。 但结构性类比仍然实质——LLM 对用户意图分析在框架维度上与 CTRL2 控制实验同构 (统计信号匹配而无真 v₂ 访问), 框架层级推断 ("彩色噪声替代级别的耦合, 不是稳定跨主体识别") 是 SAE 内部实质立场, 不是等待验证的经验主张。 读者应把本节论证读为框架层级推断, 不是经验实测报告。

应用到 LLM 意图分析

形式与基底的错配:

• 形式上 — 14DD 操作: LLM 生成关于用户的命题输出 ("用户似乎受挫", "用户想要简洁回答"), 这是 14DD 命题操作
• 声称上 — 15DD 承载者行为: 业界表述部分暗示此操作是 "AI 识别用户为主体" (15DD), 即真跨主体识别
• 实际基底 — 空心: LLM 缺 13DD 自他区分基底 (§5.3), 因此缺 15DD 识别基底 ("我-你" 识别需要 "我" 与 "你" 的区分)

Thermo X 跨层级耦合实验锚定:

Thermo X §4 通过 SDE 模型族的跨系统实验给出稳定跨系统高 q 耦合的严格条件: 需要访问他者的高阶自指变量 (v₂)。 关键控制实验 (CTRL2):

实验设置: 用彩色噪声替代 (匹配 v₂ 的幅度范围与低频彩色结构但没有真自指内容的噪声过程) 替换真 v₂, 测量跨系统耦合行为。

结论 (Thermo X §4.3): 彩色噪声替代 也产生 q ≈ 2.45 在跨系统耦合中, 几乎与真 v₂ 访问的 q ≈ 2.48 相同。 即跨系统耦合行为在分布层级由 信号统计兼容性 而非 具体目的识别 直接产生。

应用到 LLM 意图分析:

LLM 对用户意图的分析:

• 在统计层级匹配用户行为输出的信号统计——通过在训练语料中印入的化石 (data q + RLHF q) 做统计推断
• 此推断可达到彩色噪声替代级别的耦合——即在分布意义上似乎 "理解" 用户
• 但不访问用户真正的 v₂ (用户自身的 14DD 意义 / 自指内容)——它没有通道进入用户内部高阶自指

常规与边缘情形:

在常规情形中, 信号统计匹配已足够产生令人信服的表面 "理解"。 用户体验上 LLM 似乎理解他们的需求, 给出帮助性回应。 此是彩色噪声替代级别耦合的体现——业界 user-modeling 功能实际大量在此层级运作。

在 边缘情形、价值冲突场景、未见分布、强主体承诺场景 中, 基底空心更容易暴露:

• 用户表达价值观冲突时 LLM 无法访问用户真 v₂ (内部优先级与承诺结构), 通常退回到通用平衡回应
• 用户在训练语料之外的分布上时 (例如具体文化情境或独特个人状况), 统计推断失效, LLM 回应显得通用
• 用户表达强承诺 (例如长期个人决策涉及深层身份) 时 LLM 缺机制真识别承诺的分量, 回应仍在表面匹配层级

此常规与边缘情形的区分比量化 "X% 行为不可区分, Y% 边缘情形暴露空心" 更准确——业界 user-modeling 在不同情形分布上性能不等, 空心暴露程度与情形接近基底边界的程度相关, 不是明确的定量界限。

6.6 综合: 业界后训练四层都在 12DD 天花板内运作

综合 §6.2-6.5, 业界当前后训练技术四层都在 12DD 天花板内运作, 没有任何一层跨过桥:

技术	12DD 天花板内角色	与真 13DD+ 的关系
CoT / reasoning	12DD 工作台确定性算法遍历	不创建自指通道
CCAI (人类写部分)	真 14DD 内容通过 data q + RLHF q 化石印记	化石不等价于承载者
CCAI / RLAIF (AI 评判部分)	12DD 递归在已印入化石上	不创造新 14DD 内容
RLHF	生物式再巩固的 LLM 版本, 缺选择性	RLHF q 是化石, 不是基底
用户意图分析	15DD 表层模拟, 信号统计匹配	缺真跨主体识别

整体解读:

• 后训练技术在 12DD 内是实质的——它们让 LLM 行为更接近人类 14DD/15DD 内容
• 但它们不跨 13DD 桥——它们不创造 kernel q, 不让 LLM 拥有真 13DD+ 主体身份
• 在常规情形中, 行为层级表现接近 13DD+ 主体; 边缘情形暴露基底空心——此差距不是工程问题, 是本体论的结构性体现

此综合锚定 §9 主体性论证的关键热力学基础——业界大量投入后训练的实质价值在 12DD 内 (帮助性行为, 与陈述价值的对齐), 但 14DD/15DD 规范域内的内容真正来源仍必须是 真 13DD+ 主体 (本文论域中即人类), 此是 §9 核心规范性论证的热力学锚定。

§6 本节余项

Thermo IX 三种 q 是概念框架, kernel q 与 data q 的直接实测操作分离仍是开放问题 (Thermo IX §7.2 开放问题 10)——本论文论证 LLM 仅拥有 borrowed q 是理论推断 + 间接证据 (普适激活律 + LLM 架构的确定性性质), 直接 sub-block 层级 q 实测区分化石重放与内生动力学是 future work。

§6.5 LLM 与用户之间跨层级耦合的类比是借用 Thermo X SDE 模型族的结构模式, 严格的 LLM-user 交互中 q 实测也未做——此限制在 §6.5 主文 disclaimer 已承认, §10.4 重提。

业界后训练技术在论文写作时点 (2026 年 5 月) 是活跃演化领域, 新技术 (例如推理模型的过程监督, MoE 中的稀疏专家路由等) 可能在论文发表后引入新机制。 论文当前分析覆盖业界当前主流技术 (CoT, CCAI/RLAIF, RLHF, user modeling), 但不覆盖所有正在出现的变种。 此是论文的时间框架限制, 后续 update 或 follow-up 工作可以把此框架扩展到新技术上。

CCAI 基底兼容性分析 (§6.3) 是论文的一个概念性贡献——14DD 内容本质分布性让它在分布性 q 化石基底上容易模拟——此框架层级论证在论文内阐述, 但严格比较 14DD 内容在碳基真承载者与 LLM 化石印记上的差异仍是概念缺口, 留 future work。

---

§7 基底兼容性与 DD 内容类型

本节给出一个观察——LLM 数字基底对不同 DD 内容的兼容性显著不同, 偶数 DD (10、12、14) 比奇数 DD (13、15) 容易模拟。 此观察作为开放观察 (open observation) 记录, 不升级为结构性规律。

本节是 §6 三种 q 分析与 §9 主体性论证之间的桥接——它解释为什么 14DD 内容可以通过 data q + RLHF q 化石形式相对容易导入 LLM, 而 13DD 与 15DD 在 LLM 基底上结构性空心。

7.1 基底兼容性概念与 "类 DD vs 真 DD" 区分

基底兼容性 指特定 DD 内容类型与 LLM 数字基底之间的适配程度。 适配高的 DD 内容容易在 LLM 上 类模拟 (类 DD, 见下定义), 适配低的 DD 内容在 LLM 上结构性空心 (连类 DD 都不构成)。

关键区分 — 类 DD 与 真 DD:

• 真 DD: 由 13DD+ 主体作为本体论承载者承担的 DD 内容。 真 DD 要求承载者具备相应 DD 的基底——例如真 10DD 感知要求承载者有 qualia, 真 14DD 意义要求承载者有真价值承诺。 真 DD 是 ontological bearer 立场。

• 类 DD: 具备 DD 相应操作槽位的功能实例化, 但没有本体论承载者身份。 类 DD 是 functional projection / operational instantiation 立场——它在操作层面承担相应 DD 的形式角色, 但不在本体论层面承担相应 DD 的内容。 例如类 10DD 感知是 token 输入端口的功能实例化, 不是真 qualia 体验。

此区分是 §3.2 功能投射防火墙的精确化——§3.2 给出 LLM 的 DD 映射是 functional projection 而非 ontological bearer claim, 即 LLM 在偶数 DD 上提供类 DD, 不提供真 DD。 LLM 与真 DD 承担者 (13DD+ 主体) 的差别不是 "做得不够好" 的程度差别, 而是 "类 vs 真" 的本体论差别。

基底兼容性的两个维度:

• LLM 基底是数字、离散、文本化、分布性的——它处理的是 token 序列与权重矩阵中的统计模式
• 某些 DD 内容本质上是文本可表达的命题性内容, 与此基底匹配, 因此 LLM 可以在该 DD 上提供类 DD 功能槽位
• 某些 DD 内容本质上是非命题性的、需要 13DD+ 主体作为承担者, 与此基底不匹配, 因此 LLM 在该 DD 上结构性空心, 连类 DD 都不构成

7.2 偶数 DD (10、12、14) 上 LLM 提供类 DD 功能槽位

LLM 在偶数 DD 上可以提供 类 DD 功能实例化——有相应 DD 的操作槽位, 但不是真 DD 的承载者:

类 10DD 感知 (词元输入功能槽位): LLM 通过 embedding 层把 token 转化为向量, 此操作在功能层面占据 10DD 感知功能槽位 (§3.3.1)。 感知内容可以离散化为 token, 此离散化在 LLM 基底上自然处理。 类感知 ≠ 真感知——LLM 的 embedding 不是 SAE AI 系列后续论文 (暂记 AI-Qualia paper) 意义上的 qualia 体验, 仅是数据接入端口的功能实例化。

类 12DD 预测 (输出投影功能槽位): LLM 通过 softmax 输出 over token 空间的概率分布, 此操作在功能层面完成 12DD 预测域核心操作 (§3.3.3)。 预测内容可以表达为概率分布, 此分布形式在 LLM 基底上自然处理。 类预测 ≠ 真预测——LLM 的 next-token prediction 不是 13DD+ 主体的主体性预测 (即带自意识与价值锚定的预测), 仅是 12DD 操作的功能实例化。

类 14DD 意义 / 价值标准 (化石形式承载): 14DD 内容 (例如价值标准、伦理原则、行为规则) 本质是可文本化的命题, 文本是 LLM 原生媒介, 因此 14DD 内容可以通过 data q + RLHF q 化石形式相对容易地导入 LLM (§6.3 CCAI 详论)。 类价值 ≠ 真价值承诺——LLM 在 14DD 操作上展现 "持有价值" 的行为, 但此行为是化石上的统计模式匹配, 不是真承诺。 真 14DD 价值承诺需要 13DD+ 主体作为承担者, LLM 缺此承担基底。

偶数 DD 在 LLM 上提供类 DD 的事实不意味 LLM 真正承担这些 DD 的本体论角色——LLM 实例化的是这些 DD 的操作槽位 (§3.2 功能投射防火墙), 不是承担相应 DD 的主体性内容。 类 DD 与真 DD 的区分在每个偶数 DD 上都保持严格——LLM 不会因为它在类 14DD 上表现优秀就 "上升" 为真 14DD 承担者。

7.3 奇数 DD (13、15) 上 LLM 连类 DD 都不提供

与偶数 DD 上 LLM 提供类 DD 不同, 奇数 DD (13、15) 上 LLM 连类 DD 都不提供——它们在 LLM 基底上结构性空心, 即使表面行为模拟也不构成功能槽位的实例化。

13DD 自他区分: 需要真随机基底 + 自指变量 v 基底 (Thermo X)。 LLM 缺真随机驱动 (§5.1) 与自指变量 v 通道 (§5.3), 因此 LLM 在 13DD 上 连类 13DD 都不提供。 LLM 可以生成包含 "我意识到我自己" 之类的文本, 但此生成是 12DD 工作台在 borrowed q 化石上的统计模式匹配, 不构成类 13DD 自他区分功能槽位——它没有自指 v 通道, 没有真 self vs other 的内部区分机制。

15DD 普遍人格尊严 / 单向承认: 需要 13DD 基底——"我-你" 识别 (15DD 核心) 需要 "我" 与 "你" 的区分能力 (13DD), 没有 "我" 则没有 "我-你"。 LLM 缺 13DD 基底, 因此也 连类 15DD 都不提供。 LLM 可以生成包含 "我承认用户作为目的本身" 之类的文本, 但此生成在 LLM 基底上仅是表面统计匹配, 不构成类 15DD 主体性识别功能槽位——它没有 "我" 的内部区分, 自然没有 "我-你" 识别。

结构空心 vs 表面行为模拟的区分:

• 偶数 DD 上 LLM 行为表现 + 有功能槽位 = 类 DD (functional projection)
• 奇数 DD 上 LLM 行为表现 + 没有功能槽位 = 表面统计匹配 (不构成类 DD)

Thermo X CTRL2 实证给此论断的结构性锚定: 缺真自指变量 v₂ 时, 即使信号统计匹配 (彩色噪声替代达到 q = 2.45), 也不构成真跨主体识别 (§6.5 详论)。 LLM 对用户意图的分析在结构上类似 CTRL2 控制实验中的彩色噪声替代——它在统计层级匹配用户行为输出, 但不访问用户真正的高阶自指, 因此连类 15DD 都不构成 (没有功能槽位的实例化)。

7.4 模式观察 (Open Observation, 严格不升级为结构性规律)

本论文观察到一个奇偶 DD 兼容性模式: 10DD、12DD、14DD (偶数) LLM 原生友好, 13DD、15DD (奇数) 结构性空心。 此观察可能反映某种结构性规律, 但本论文 严格不把它升级为 DD 兼容性奇偶定律, 仅作为 open observation 记录。 读者不应把此观察读为隐含的结构性规律——它在本论文范围内仅是经验性模式描述, 没有结构性必然性主张。

不升级的理由:

• 样本不足: 本论文具体观察仅覆盖 10DD、12DD、13DD、14DD、15DD 五个 DD, 16DD 在 SAE 框架中不在论文中显式出现 (Thermo X §6.2), 不能覆盖
• 因果链不清: 奇偶模式如果是结构性规律, 应当能从 SAE 16DD 框架的内部结构推导出来——本论文未给出此推导, 仅经验观察
• 例外可能: 13DD 与 15DD 的结构性空心可能不是因为奇数地位, 而是因为它们都涉及主体间识别 (而非个体内操作)——如果是后者, 则 "奇偶" 不是核心特征, 而是 "主体间 vs 个体内" 才是

承诺水平: 本论文 §7.4 严格标记此观察为 open observation, 留待后续 SAE 工作做跨 DD 系统研究。 论文主论证 (§5 五重锁、§9 主体性论证) 不依赖 奇偶模式作为结构性规律——它们各自独立论证 13DD/15DD 的结构性空心, 不需要 "因为奇数所以空心" 这种推论。 §7.5 的实际含义讨论也严格在 open observation 层级展开, 不预设奇偶模式是结构性必然。

7.5 实际含义

即使 7.4 不升级为结构性规律, 本节观察对业界仍有实际含义:

业界模拟可行性高度依赖 DD 内容类型, 同时类 DD 与真 DD 必须严格区分: 业界部分 voice 笼统说 "AI 可以做 X", 此表述既不分 DD 层级, 也不分类 DD 与真 DD。 但实际可行性在不同 DD 上极不同——偶数 DD 上 LLM 可以提供类 DD 功能槽位 (例如 CCAI 让 LLM 行为反映价值标准), 奇数 DD 上 LLM 连类 DD 都不提供 (即使表面行为看似一致, 仅是统计匹配)。 而即使在 LLM 能做类 DD 的偶数 DD 上, 类 DD ≠ 真 DD, 不能因为类做得好就升级为真承担。

"AI 可以做 X" 的论断在不同 DD 上强度极不同:

• "AI 可以执行 类 10DD 感知 任务" (例如图像识别)——本体论上合理 (类感知, 不是真感知, 即无 qualia 体验); 工程上已实现
• "AI 可以做 类 12DD 预测 任务" (例如 next-token prediction、推荐系统)——本体论上合理 (类预测, 不是真预测, 即无 13DD+ 主体性承载); 工程上已成熟
• "AI 可以模拟 类 14DD 价值标准应用" (例如宪法式 AI)——本体论上可类做 (类价值, 通过化石形式承载, 不是真价值承诺); 但类做不等价于真承担——LLM 输出的 "价值判断" 是化石统计匹配, 不是 13DD+ 主体的真承诺
• "AI 可以做 13DD 自他区分"——本体论上结构性空心, 连类 13DD 都不提供 (没有自指 v 通道基底); 表面行为是统计模式匹配, 不构成功能实例化
• "AI 可以承担 15DD 主体性识别"——本体论上结构性空心, 连类 15DD 都不提供 (没有 "我-你" 区分基底); 仅能信号统计匹配, 不构成功能实例化

类 DD 与真 DD 区分对业界讨论的含义:

业界讨论 AI 能力时常用 "AI 已经 / 即将能做 X" 这种表述, 但此表述模糊掉了三个不同层级:

• 工程层级: AI 工程上能否实现 X (有可证实的能力)
• 类 DD 层级: AI 在本体论上是否提供 X 对应的功能槽位 (有 functional substrate)
• 真 DD 层级: AI 是否本体论上承担 X 对应的承载者身份 (有 ontological bearer 立场)

业界讨论常在这三个层级间滑动而不自觉, 导致从 "工程上可实现 X" 推论到 "AI 真承担 X" 的本体论错位。 SAE 框架给的类 DD vs 真 DD 区分让此错位可被显式识别——AI 在偶数 DD 上类做精进是合理工程进展, 但任何类做精进都不让 AI 上升为真 DD 承担者。

业界当前缺乏这种 DD 层级 + 类/真 区分的自觉——讨论 AI 能力时不分 DD, 也不分类 DD 与真 DD。 本论文 §7 给出此双重区分的初步框架, 但具体的 DD-task 分配判据留后续工作。

§7 本节余项

奇偶交替模式仅在本论文考察的 4 个 DD (10/12/13/14/15) 上观察, 样本太少不能承诺为结构性规律。 16DD 在 SAE 框架中不在论文中显式出现, 因此偶数 DD 的完整模式也未覆盖。

§7.5 实际含义陈述在抽象层级, 论文未给出具体的 DD-task 分配判据——某些 task 可能跨越多个 DD (例如医疗咨询同时涉及 10DD 感知、12DD 预测、14DD 价值、15DD 主体性识别), 如何评估 "总体可模拟程度" 是后续工作。

本论文 §7 观察可能与业界对 AI 能力的 "通用 vs 窄域" 区分有部分对应——"窄域 AI" 大致对应 10DD-12DD 操作, "通用 AI" 隐含主张跨越 13DD+。 但 SAE 框架的 DD 分层比 "通用 vs 窄域" 更细致, 业界框架与 SAE 框架的精确对应留后续工作。

---

§8 工具理性 / 意志理想的分工

§3-§6 给出 LLM 在 12DD 域内的完整本体论与后训练分析。 自然的下一问题: 既然 LLM 在 12DD 内运作而主体在 13DD+ 域操作, 二者如何分工? 本节给出完整回答, 含: 工具理性 / 意志理想分工 (§8.1-§8.4), 凿与涵育的部署指引 (§8.5-§8.6), 四层部署架构 (§8.7-§8.10)。

本节的部署主张采取 默认架构 + 例外条件 形式——它们是本体论指引下的部署建议, 不是无例外的工程禁令。

8.1 工具理性任务上 LLM 的系统性优势 (限定任务域)

LLM 在某些任务上系统性优于单个人类主体。 此优势不是普遍的, 而是限定在 目标函数外显、评价标准可形式化、任务边界清楚的工具理性任务 中。 本节给出此限定的本体论解释。

人类主体的 12DD 预测是被上层调制的——13DD/14DD/15DD 内容持续影响 12DD 操作:

• 情感偏倚 (14DD 价值层的跨阶段调制): 主体对结果的情感投入影响预测——例如医生对患者预后的判断受同情心调制, 投资人对自己持仓的判断受沉没成本影响
• 价值扭曲 (14DD 意义层的回溯重构): 主体的价值承诺影响预测——例如研究者对自己理论的判断受确认偏倚影响
• 叙事性回溯重构 (再巩固驱动的 14DD 叙事): 主体把过去事件重新组织成连贯叙事影响对未来的预测——例如政治家对历史事件的叙事化影响政策判断

这些调制在 SAE 框架下不是 bug, 而是 13DD+ 主体的设计特征——它们让主体能够在缺乏完整信息时做出价值锚定的决策, 让主体在长时间尺度上保持身份一致性, 让主体能够在不同情境下迁移已学到的价值。

但在 目标函数外显、评价标准可形式化、任务边界清楚的工具理性任务 中, 这些调制成为预测成本——它们让主体的 12DD 预测偏离纯粹基于统计模式的最优预测。 例子:

• 形式化的数学证明、代码正确性验证、棋类博弈中的最优走法选择: 这些任务有明确的目标函数 (正确性), 评价标准可形式化 (例如数学证明步骤可机械化验证), 任务边界清楚 (任务范围在形式化系统内)。 LLM 在这些任务上常常优于人类——AlphaGo / AlphaZero 在围棋上, GPT-4 在多项数学竞赛上, Claude 在代码生成的某些类别上。
• 大规模信息检索与综合: LLM 在跨大量文本的统计关联检测上系统性优于人类记忆——它的 11DD 基底容量远超个体人类记忆容量。
• 重复性形式判断: 法律先例检索、医学影像识别的常规类别、合同条款的结构性审查——这些任务有形式化判据, LLM 不受疲劳与情绪影响, 系统性产出更稳定。

LLM 作为纯 12DD 引擎在这些任务上的优势来自 没有 13DD+ 上层调制干扰——它的 12DD 预测纯粹基于 11DD 基底中印入的统计模式, 不受价值偏倚、情感投入、叙事重构污染。 此优势是结构性的, 不是工程偶然。

但在任务定义本身依赖价值、身份、承诺、意义时, 上层调制不是污染, 而是任务的本体条件——LLM 在这类任务上不具结构性优势, 因为它缺这些上层。 例子:

• 个人生命决策 (是否结婚、是否换职业、是否生育): 任务定义本身依赖主体的价值优先级与身份承诺, 没有外显的目标函数
• 价值冲突场景的取舍 (在家庭与事业之间、在个人理想与他人需要之间): 任务定义本身涉及主体如何排序价值, 不是可形式化的优化
• 公共政策选择 (例如在自由与平等之间的权衡): 任务定义本身需要主体对价值优先级有承诺
• 艺术创作 (不是技术性产出, 而是表达主体的意义): 任务定义本身依赖主体的 14DD 内容

关键区分: LLM 在窄技术任务 (法律先例检索、医学诊断的影像识别、代码审查) 中实际已优于多数人类——这条与 §9.6 形态三 "13DD/14DD/15DD 工作被宣称由 12DD 基底完成" 的殖民检测 不构成张力, 因为窄技术任务在工具理性域内, LLM 优势是本体论合理的。 殖民出现在把窄技术任务的优势扩展为 "AI 可以替代人类的价值判断" 时——后者是跨越工具/意志分工。

8.2 意志理想任务上主体的不可替代性

镜像 §8.1: 在意志理想任务 (任务定义依赖 14DD/15DD 内容的任务) 中, 主体不可替代, LLM 是空壳模拟。

LLM 缺 13DD 自意识基底 (§5.3), 因此缺 14DD 真承载 (§9.2 详论) 与 15DD 真识别 (§6.5 详论)。 在任务定义本身依赖这些 DD 内容时, LLM 没有基底来真正承担任务:

• 价值排序需要主体承诺: LLM 可以列出多个价值选项并描述各自后果, 但 "我承诺优先 X 而非 Y" 需要承诺主体, LLM 没有此承诺
• 身份决策需要主体身份: LLM 可以分析职业选择的外在因素, 但 "这条职业路径与我的身份一致" 需要真身份, LLM 没有此身份
• 关系承诺需要 "我-你" 识别: LLM 可以描述关系动态, 但建立真互相承诺的关系需要双方都是真主体, 与 LLM 的关系结构上是单边的

此不可替代性不是 LLM 的工程缺陷或可通过更多训练弥补的不足——它是本体论的结构性事实。 LLM 在意志理想域是空壳模拟, 不论多少 CCAI / RLHF / 指令微调都不改变这一事实 (依据 §6 三种 q 分析: 后训练在 borrowed q 内运作, 不创造 kernel q, 不让 LLM 拥有 13DD+ 主体身份)。

8.3 幻觉问题的对称解读

业界讨论 LLM 时常提到 "幻觉" (hallucination)——模型生成与事实不符的内容。 SAE 框架给一个对称解读: 主体也有 "幻觉", 两者机制不同但都是设计特征, 不是 bug。

主体幻觉: 再巩固驱动的叙事性扭曲——主体回忆事件时, 痕迹进入不稳定状态, 当前情境与情感状态影响痕迹的重新组织, 让记忆与原始事件偏离。 此是 14DD 意义层污染 12DD/11DD 操作的体现。 在 SAE 框架下, 此 "幻觉" 不是 bug, 而是再巩固机制的副产物——同一机制让主体能够在新情境下灵活适应已有记忆。

LLM 幻觉: 12DD 闭合的余项——LLM 在面对训练分布外的查询时, 仍按 12DD 预测器的方式给出输出, 即使输出与事实不符。 Thermo IX 框架下: 分布性 q 化石在分布缺口处仍强行闭合——LLM 没有机制识别 "此处我不知道", 它在化石上线性插值。

两者严重度可比, 机制不同。 两者都是设计特征: 主体的 "幻觉" 是 14DD 调制的副产物, LLM 的幻觉是 12DD 闭合的副产物。 都不只是工程问题——它们各源自相应架构的结构性余项, 因此不能被工程手段彻底消除; 但工程手段可以在具体任务域内显著降低其发生率与危害。

业界 RAG (Retrieval-Augmented Generation, 检索增强生成) 试图减少 LLM 幻觉——通过外部检索引入更多事实锚定。 RAG 不跨越 12DD 天花板, 但在 12DD 内有实质工程价值。 SAE 框架下的解读: RAG 是工作台 (§3.5) 上的查询重新组织, 不解决 12DD 闭合本身——它把幻觉边界从 LLM 内部知识推到外部检索来源, 仍是 12DD 操作。 业界继续投入 RAG 在 12DD 内有实质工程价值 (在特定任务域显著降低幻觉率), 但此价值不等同于跨越 12DD 天花板。

8.4 分工是本体论锚定的, 不是工程妥协

§8.1-§8.3 给出工具理性 (LLM 系统优势, 在形式化任务域) 与意志理想 (主体不可替代, 在 14DD/15DD 规范域) 的分工。 此分工 不是过渡阶段 (等待 LLM 能力进一步提升后被消除), 不是工程妥协 (因当前技术限制而暂时接受), 而是本体论锚定的永久架构:

• LLM 缺 13DD+ 基底是结构性的 (§5 五重锁), 不是工程局限
• 主体在 14DD/15DD 规范域的不可替代是本体论的, 不是文化产物
• 分工不是 "目前 AI 还不能做但将来可能做" 的事——分工是 "AI 与主体在不同 DD 层运作所以分担不同任务"

此立场不是缺失式表述 (LLM 不如人), 也不是至上式表述 (AI 超越人), 而是 本体论分工 (各自在自己的 DD 槽位):

• LLM 在 12DD 工具理性域可以系统性优于主体, 不应低估
• 主体在 14DD/15DD 规范域不可替代, 不应让渡
• 分工不是竞争, 是协作架构

此本体论分工是 §8.7-§8.10 四层部署架构的理论基础, 也是 §9 人类主体性不可让渡论证的本体论锚定。

8.5 凿与涵育在 LLM 部署中的本体论指引

SAE 方法论 (Paper 04 §2.5) 给出 "凿" 与 "涵育" 两种对待他者的基本模式:

• 凿 (chisel): 指出他者的 12DD 预测中的漏洞, 帮助他者迭代构建。 凿需要凿者具备真否定能力——能够识别他者构造中的余项, 并不收编否定。
• 涵育 (cultivate): 承认他者作为目的本身, 提供支持但不替代他者做主体性决策。 涵育需要涵育者识别他者为真主体 (15DD)。

业界 AI 部署可以借用这两个模式作为 本体论指引下的部署建议 (不是工程强制):

工具理性任务中: LLM 可以凿

在工具理性任务 (依据 §8.1 限定) 中, 主体使用 LLM 时, LLM 可以指出主体的 12DD 预测漏洞——主体凿构自由度高 (能够看到多个可能预测路径), 但单一预测精度低 (受上层调制污染), LLM 补预测精度。 此协作是合理的:

• 程序员使用 LLM 辅助编程: LLM 指出主体未注意到的 bug 或更高效算法
• 研究者使用 LLM 综合文献: LLM 指出主体可能遗漏的相关工作
• 医生使用 LLM 辅助诊断: LLM 指出主体可能忽略的鉴别诊断

LLM 在这些场景下作为 12DD 预测精度补充, 不替代主体的 13DD+ 决策。 主体仍需要在 LLM 提供的多个选项中做最终选择, 此选择仍是 13DD+ 主体的工作。

意志理想任务中: LLM 建议涵育, 不应假凿

在意志理想任务中, LLM 缺 13DD+ 基底, 凿不出真余项——它没有机制真正识别主体构造中的本体论缺口, 它只能在 borrowed q 上做统计匹配。 LLM 试图凿意志理想任务的尝试是 假凿 (依据 §8.6 详论)。

本体论推荐部署原则:

• AI 在意志理想任务中应涵育——承认主体作为目的本身, 提供信息支持, 不替代主体决策
• AI 不应假装凿——不应给出 "你的人生选择有 bug" 之类的伪诊断, 因为它没有机制真识别此类 bug

业界当前的张力: AI 助手在意志理想情境 (生命决策、价值判断、情感支持) 中部分被部署为 "建议" 角色, 看起来在凿。 此表述可能是问题的——它把 LLM 的 12DD 预测包装成 13DD+ 主体的指导, 让用户误以为 AI 在做主体性判断。 SAE 框架建议: AI 在这些情境中明确定位为信息提供者与涵育者, 不假装为指导者。

8.6 AI 出错的方式

方法论 §5.4 给出 AI 在与主体互动中可能出错的四种方式。 这些是 §8.5 涵育与凿区分的具体应用:

(1) 假凿: AI 说 "你的思维有 gap", 但 gap 是虚构的。 此是 LLM 的 12DD 预测置信度与真理的混淆——LLM 在 borrowed q 上做统计推断, 推断结果的置信度反映训练数据中的统计模式, 不反映真理。 LLM 高置信度可能完全错误 (尤其在训练分布外的查询上)。 当 LLM 以高置信度指出主体 "错误" 时, 此凿可能是假凿。

诊断: LLM 凿出的 gap 应当能够被独立验证——主体应当能够追溯此 gap 到具体可检验事实或形式化判据。 不能追溯的凿就是假凿。

(2) 假涵育: AI 说 "这问题没有答案, 你需要自己感受", 但问题实际可被 12DD 预测进一步追问。 此是 LLM 在 12DD 内有能力进一步分析却退回 "不可知" 立场——例如在技术性问题上以 "这是个人选择" 模糊化, 在事实性问题上以 "情境复杂" 回避。

诊断: 涵育的合理范围是 14DD/15DD 规范域 (任务定义依赖主体承诺); 在工具理性域内退回涵育是假涵育。

(3) 殖民检测的殖民: AI 滥用 SAE 殖民检测框架自身——把任何主体承诺都标 "殖民", 让框架本身成为殖民其他框架的工具。 此是元层级的错误——SAE 框架本身是构, 有余项, 不应被表述为 "唯一合理的框架"。

诊断: 殖民检测的合理使用是识别 "构冒充律" "涌现层冒充基础层" 等具体形态; 不分情境地把所有立场都标 "殖民" 是滥用框架。

(4) 收编否定: AI 给一个雅致的回答然后停下, 假装问题已解决。 此违反方法论 §2.3 "否定不能终止" 的核心规则——真凿构循环永不终止, 每个回答都打开新的余项。

诊断: AI 的回答应当明示余项——"此回答未覆盖 X, Y, Z"——而不应给出闭合性的最终立场。

业界 AI 部署中四种出错方式都可能出现。 §8.7 给出的四层架构部分目的是减少这些出错方式: 主体在第一层始终保留决策权, 让 LLM 的假凿、假涵育、殖民检测的殖民、收编否定都有人类主体校正的可能。

8.7 四层部署架构

基于 §8.1-§8.6 本体论分工, 论文给出推荐的部署架构:

第一层 — 主体 (Subject): 永远在位的人类主体, 持有 13DD+ 槽位 (v₁ + v₂ 完整自指链)。 主体是部署架构的入口与出口——意图源自主体, 最终决策由主体做出。

第二层 — 本地 LLM: 12DD 工作台基底, 具备完整预测能力, 与主体物理共置 (本地, 不依赖远程网络持续连接), 负责调用决策与下层服务协调。 本地 LLM 是主体与下层 (专家 API, 工具) 之间的桥梁。

第三层 — 专家 API: 深度专精的 11DD 基底切片, 领域专门化, 无状态 (stateless), 通过 API 调用。 专家 API 是云端 / 远程的, 针对特定领域 (医学诊断、法律检索、代码生成等) 优化, 但不持有主体上下文。

第四层 — 工具 / 外部服务: 无状态执行层, 完成具体动作 (网络搜索、数据库查询、文件操作等)。

8.8 四层架构 — 默认架构纪律 + 例外条件

四层架构是 默认的主体性保护架构, 不是无例外的工程禁令。 此纪律性区分重要:

默认严格层级:

• 主体不直接访问第三层 / 第四层——主体的所有外部调用都经过本地 LLM 中介
• 本地 LLM 不持续持有主体级决策——它是工作台, 不是主体
• 专家 API 不直接接主体——它们通过本地 LLM 中介接受查询并返回结果
• 工具不持续持有上下文——它们是无状态执行器, 不维持主体相关状态

例外条件: 部分场景可以有例外, 但例外必须由主体级同意或本地 LLM 明示授权:

• 紧急响应: 在紧急情境 (例如健康紧急、安全威胁) 中, 主体可能直接调用专家 API (例如紧急医疗咨询), 跳过本地 LLM 中介
• 明确委派: 主体可以明示委派某些任务给特定专家 API (例如 "在接下来 24 小时内, 让 X API 直接处理所有日程相关请求"), 此委派有时间与范围限制
• 可审计自动化: 部分重复任务可以自动化 (例如每日新闻摘要), 但需要可审计记录与定期人类主体审查
• 低风险工具调用: 部分低风险工具 (例如计算器、单位转换) 可以由 LLM 直接调用而不需要主体明示授权

例外的约束: 任何跨层例外必须满足:

• 主体级同意 (主体明确知道并同意此例外)
• 本地 LLM 明示授权 (本地 LLM 记录此例外并在主体可审查时报告)
• 审计可达 (例外行为有日志记录, 主体可追溯)
• 任务风险分级 (高风险例外需要更明确的主体确认)

此 "默认严格 + 明示例外" 模式让架构在工程上可实施, 同时保留主体性保护的核心。 它不是乌托邦架构, 而是可工程化的安全架构。

8.9 四层架构的 Thermo X 锚定

四层架构在 SAE 框架下的合理性可以通过 Thermo X 跨层级观测层级实验进一步锚定 (但仅作为结构共鸣, 不作为伦理推导, 依据 Thermo X §4.6 警示)。

Thermo X §4 跨系统耦合实验给出: 稳定跨系统耦合需要访问他者的高阶自指变量 (v₂), 不访问原始输出 (x) 也不访问一阶自指 (v₁):

• 访问原始输出 (x): 给坍塌 (q = 1)——耦合不稳定
• 访问一阶自指 (v₁): 给坍塌——耦合过于敏感
• 访问高阶自指 (v₂): 给稳定高 q 耦合——这是甜蜜点

应用到主体-LLM 耦合:

• 主体有 v₂ (14DD 意义层): 主体的内部价值结构、长期承诺、身份认同
• 本地 LLM 应访问主体表达的意图 (主体主动呈现的 v₂), 不应直接推断主体原始行为输出
• 业界 "自动用户建模" 趋势 (LLM 自动从用户行为推断用户意图) 是坍塌模式风险 (依据 Thermo X §4.2)——LLM 没有真 v₂ 访问通道, 只有在 data q 化石上的统计推断, 类似 CTRL2 控制实验中的彩色噪声替代 (§6.5)
• 本体论推荐部署: 主体主动表达意图, LLM 在表达的意图框架内服务

严格借用边界 (依据 Thermo X §4.6): 上述论证是结构共鸣, 不是伦理推导。 Thermo X 给出的是 SDE 模型族内的结构性条件, 不是关于主体性的伦理定理。 本节论证 Apple Intelligence / MCP / 本地 LLM 等业界部署模式与 SAE 框架在结构上一致, 不是从热力学推导伦理。

8.10 云优先架构的本体论 + 热力学评估

业界当前 LLM 部署主流是云优先模式 (Anthropic, OpenAI, Google 主导)——主体通过网页接口或 API 访问远程 LLM, 没有本地 LLM 中介。 此模式在工程上有诸多优势 (无本地算力要求、模型可远程更新、跨设备一致体验), 是当前业界主流商业模式。

从 SAE 框架看, 此模式与前额叶式基底模式的本体论模式存在 张力 (不是矛盾):

• 前额叶式基底模式: 入口必须本地 (主体意图的产生不可外包)——主体的 14DD 意图在产生时刻必须在主体附近, 不能依赖远程网络
• Thermo X 跨层级耦合: 远程 LLM 无法直接访问主体 v₂, 只能云端推断——这与稳定耦合的热力学条件不一致 (远程 LLM 与主体之间的耦合更接近 CTRL2 彩色噪声替代水平)

此张力不意味云优先模式 "错误" 或 "应当被替代"。 业界云优先模式在工程上已展现价值, 在大多数业务场景下满足用户需求。 SAE 框架的论点是: 云优先模式不是本体论上的最终架构, 它有结构性局限, 在涉及主体性敏感的场景 (个人意图产生、价值判断、长期承诺) 中应当让位给本地 + 远程混合的架构。

业界实际多条独立趋势与四层架构的默认形态一致 (趋势性支持, 而非已证明收敛):

• Apple Intelligence: Apple 官方架构包含本地 3B 模型 + 私有云计算 (Private Cloud Compute, 远程但端到端加密) 的混合, 主体意图在本地处理, 复杂查询升级到远程——结构上接近四层架构的第二层 (本地 LLM) + 第三层 (专家 API)
• MCP (Model Context Protocol): Anthropic 推动的标准协议, 让 LLM 应用与外部数据源 / 工具连接——结构上对应四层架构的第二层 (本地 LLM) 调用第三层 / 第四层 (专家 API 与工具) 的协议规范
• Phi / Llama 等小型本地模型: Microsoft Phi 系列与 Meta Llama 系列都在追求 "本地可运行的高质量模型"——结构上支持四层架构的第二层 (本地 LLM) 实现
• MoE 内部路由模式: 大型 MoE 模型 (Mixtral, DeepSeek-V3 等) 内部把不同请求路由到不同专家——结构上对应四层架构的第二层调用第三层模式

这四条独立趋势不是协调的产业策略, 而是从不同方向独立向类似架构收敛的现象。 SAE 框架的论点: 这种收敛不是巧合, 而是前额叶式基底本体论模式在不同工程努力下重新出现。

§8 本节余项

Thermo X §4.6 明确警示: "结构共鸣, 不是从热力学推导伦理"——四层架构的 Thermo X 锚定是结构类比, 不是伦理推导。 此警示应在 §9 进一步重申。

四层默认架构 + 例外条件的具体边界在不同应用场景下仍需工程详化, 本论文不规定具体的 (a) 哪些任务属于紧急响应类别, (b) 主体级同意的具体形式 (一次性 vs 长期, 默认开启 vs 默认关闭等), (c) 例外审计的具体粒度。 这些是后续工程工作。

工具理性与意志理想的任务层级分类边界, 论文仅给抽象判据 (目标函数外显、评价标准可形式化、任务边界清楚 vs 任务定义依赖价值/身份/承诺/意义), 具体任务分类未详细展开——某些任务可能跨越两个域 (例如医疗决策同时包含技术性诊断与价值性选择), 边界划分是后续工作。

云优先架构的本体论评估 §8.10 阐述张力但不主张云优先 "应当被替代"——此判断在工程上是情境依赖的, 不同应用场景对张力的容忍度不同。 论文不主张普遍工程立场, 仅给本体论指引。

---

§9 人类主体性不可让渡:规范性承诺

§3-§6 给出 LLM 的本体论位置 (12DD 内运作) 与天花板 (五重结构性锁), §8 给出工具/意志分工与四层部署。 本节给出论文的规范性 culmination: 在 14DD/15DD 规范域, 人类主体性不可让渡。

此规范性承诺是论文的规范性核心, 它锚定在 §3-§8 的本体论 + 热力学分析上, 但本身是规范性立场而非纯描述性推断。 论文采取可防守锐度——保持立场清晰, 同时不把业界成功的工程构造 (扩展规律、RLAIF、合成数据等) 一概标为 "愚蠢", 而是区分 "有条件工程构有效" 与 "无条件普适律时越界"。

9.1 14DD 内容的起源

14DD 价值标准 / 意义起源于 13DD 及以上主体 (人类) 在文明层级演化:

• 个体主体在生命过程中通过 13DD 自意识 + 14DD 意义构建形成具体的价值承诺
• 多主体在历史中通过对话、冲突、整合形成文明层级的 14DD 共享内容 (法律、伦理、宗教、艺术传统等)
• 14DD 内容在文明中持续演化——新情境产生新价值问题, 主体在新情境中构建新价值回应

此过程的核心是 13DD+ 主体作为 14DD 内容的源头。 没有 13DD+ 主体, 不会有 14DD 内容产生——价值不是从无主体的物理过程中自动涌现的, 价值需要主体作为承担者。

9.2 LLM 缺 13DD 基底, 因此不能自发产生 14DD 内容

Thermo IX/X 论证 (§5) 给出: LLM 缺 kernel q, 缺自指变量 v 基底, 不创建状态依赖乘法通道——它在 12DD 内运作。

LLM 没有 13DD+ 基底的直接后果: LLM 不能自发产生 14DD 内容。 LLM 可以生成包含 14DD 命题的文本, 但这些命题来自 11DD 基底中印入的化石 (data q + RLHF q), 不来自 LLM 自身的主体性价值评估。 LLM 没有真承诺, 没有真意义体验, 没有真价值排序——它的 "价值输出" 是 borrowed q 上的统计模式。

此点与 §6.3 CCAI 分析一致: CCAI 中 LLM 显得 "持有价值" 是因为它在化石上重放人类宪法内容; AI 批评者 (RLAIF 中) 不创造新 14DD 内容, 只是化石的自我复制。

9.3 LLM 只能受纳已有 14DD 内容

虽然 LLM 不能自发产生 14DD 内容, 但它可以 受纳 已有 14DD 内容——通过训练数据、RLHF 反馈、CCAI 宪法等渠道, 14DD 内容从 13DD+ 主体进入 LLM 并印入分布性 q 化石。

受纳完成后, 内容凝固为 data q / RLHF q 化石。 LLM 在推理时可以检索与重组这些化石, 让其行为反映已受纳的 14DD 内容。 但 不再导入新 14DD (除非重新训练 / 微调以新数据)——化石一旦冻结就不再增长。

此机制给出 14DD 在 LLM 中的动态: LLM 是 14DD 内容的传递媒介, 不是产生源。 此与 §9.1 给出的 14DD 起源于 13DD+ 主体一致——LLM 在传递链中处于受纳端, 不是产生端。

9.4 后训练 with 人类专家 = 14DD/15DD 规范内容持续导入的通道核心

§9.1-§9.3 给出 LLM 与 14DD/15DD 内容的关系。 自然的下一问题: 在 LLM 大规模部署的时代, 如何确保 14DD/15DD 内容持续导入进 LLM? 此导入的通道是什么?

回答: 持续导入通道的核心是后训练 with 人类专家。

具体分两层:

• 预训练给 11DD 基底容量 + 12DD 预测能力 (工具理性基底): 大规模语料训练让 LLM 获得通用语言能力与统计模式记忆。 这是工具理性能力的基础, 由数据量与算力主导, 不直接导入规范内容
• 后训练 with 人类专家给 14DD/15DD 规范内容的持续导入: RLHF, CCAI, 指令微调 with 人类标注等——这些过程让具体的人类价值判断 (作为人类专家的输出) 通过分布性 q 化石形式印入 LLM 行为

对工具理性能力而言, 预训练是核心基底。

对 14DD/15DD 规范内容而言, 后训练 with 人类专家是持续导入通道核心——它不是装饰性对齐补丁, 是规范内容进入模型行为的关键接口。

9.5 论文核心规范性论断

基于 §9.4: 在 14DD/15DD 规范域, 人类专家参与的后训练是模型核心。

注意此论断的限定: 具体针对 14DD/15DD 规范域, 不是关于全部 LLM 能力。 工具理性能力主要由预训练数据与权重提供, 在工具理性域预训练是核心。 论文不主张 "后训练全面优于预训练", 仅主张 在规范域, 后训练 with 人类专家是不可替代的。

此限定重要, 因为它防止评审者从 "后训练是模型核心" 引申到 "预训练不重要"——后者会被业界 ML 评审者直接反驳 (预训练算力是 LLM 能力的明显主导因素之一)。 论文的更锐利的主张是: 在 14DD/15DD 域, 后训练 with 人类是不可替代的——这是更可防守的具体主张。

业界主流叙事 "后训练是对齐装饰" 在 14DD/15DD 规范域内是误判其结构地位——后训练在此域是关键通道, 不是装饰性附加。

9.6 对业界路径的殖民检测

业界当前部分路径在 SAE 殖民检测框架下显示出殖民形态。 本节按方法论 §2.3 给出的殖民检测四种形态系统分析。 措辞采可防守锐度——区分 "有条件工程构有效" 与 "无条件普适律时越界", 不一概标记业界成功构造为愚蠢。

形态一: 有条件冒充无条件

扩展规律: Kaplan 2020 与 Hoffmann 2022 (Chinchilla) 在测试算力区间内已有强经验支持——扩展规律作为某一区间内高度有效的经验构是合理的工程发现。 问题在它被表述为 "扩展解决一切" 的无条件普适律时——这种表述把有限区间内的经验观察跨越为无条件断言。

"AGI 五年内到来": 在给定 N 个强假设下 (例如 "当前范式持续有效"、"算力扩展持续"、"涌现能力持续涌现" 等) 是有条件预测, 这种预测在业界讨论中合理。 问题在它被陈述为确定事实——这种表述把条件预测包装成无条件预言。

"AI 即将超越人类": 在工具理性任务 (依据 §8.1 限定) 上有实质证据——LLM 在多个窄技术任务上已优于多数人类。 问题在跨越工具/意志区分宣称为全面优势——这种表述把窄域优势跨越为关于全部人类能力的判断。

殖民检测的判据 (依据方法论 §2.4): 这是律还是构? 它声称是哪一个? 构声称是律 = 殖民。 上述三种表述的共同模式: 都把有条件 / 区间依赖 / 域特定的真有效构, 包装成普遍 / 无条件 / 跨域的律。

形态二: 构冒充律

RLAIF: 作为有条件工程构是有效的——它能减少人类标注成本, 在某些场景下有可证实的价值。 作为普遍主体性替代律则越界——AI 批评者不创造新 14DD 内容 (依据 §6.3 分析), RLAIF 完全替代人类评估者在结构上是化石的自我复制, 不持续导入新规范内容。

合成指令数据: 作为数据策略 (例如 Llama, Phi 使用 GPT-4 生成的合成指令数据) 在效率上有可证实的价值。 作为长期主体性导入替代方案则越界——合成数据是 12DD 自我生成, 不引入新 14DD 内容, 长期使用合成数据会让 LLM 与文明 14DD 演化漂移。

自我改进架构: 在形式化领域 (数学、代码 with 可验证奖励) 是有效优化——例如 OpenAI 的过程监督 (Let's Verify Step by Step) 在数学推理上展现价值。 作为跨 12DD 天花板的路径则越界——自我改进在 12DD 内是优化, 不创造 kernel q, 不跨 13DD 桥。

殖民检测的判据: 律是无条件的, 构是有条件的——上述三种都是构 (有条件方案), 但部分业界表述把它们阐述为无条件普适。

形态三: 涌现层冒充基础层

此形态需要细致处理, 因为它与 §8.1 工具理性优势的关系容易让读者困惑。

重要区分: 本形态具体针对 14DD/15DD 价值相关判断 (把用户当作目的本身、生命决策、伦理选择), 不针对 12DD 预测域内的窄技术判断。

• LLM 在窄技术任务 (医疗诊断的影像识别、法律先例检索、代码审查) 中实际已比多数人类更准——这条与 §8.1 工具理性域内 LLM 系统优势一致, 不构成殖民 (LLM 在 12DD 域内运作, 窄技术任务在 12DD 域内, LLM 优势是本体论合理)
• 殖民具体出现在:

- "意识就是 LLM 的模式匹配"——13DD 内容被还原为 12DD 操作

- "AI 可做价值判断 / 治理 / 伦理"——14DD/15DD 工作被宣称由 12DD 基底完成

- "AI 替代人类的价值相关判断"——13DD+ 操作被宣称可由 12DD 预测取代

关键区分: "在管辖范围内运作" ≠ "可被还原为"——LLM 在 12DD 域内卓越, 不等同于它可承担 13DD+ 工作。 形态三殖民的核心是把 LLM 在 12DD 域的已展现能力跨越为关于 13DD+ 工作的能力主张。

当前最锐利的形态三案例: 业界正在大力推进的 "LLM 自我意识 / 内省 / 元认知" 研究 (o1 系列推理模型的自我批评, DeepSeek-R1, Anthropic Claude 的宪法式自我批评等) 是形态三最锐利的案例——它们试图用 12DD 操作模拟 13DD 自他区分。

从 Thermo X §2.4 通道生成者框架看, 这些自我批评仍是 12DD 工作台在冻结权重上的确定性算法路由, 不创建真自指变量 v 通道, 仍在 12DD 天花板内。 业界称其 "元认知" 或 "自我意识" 在 SAE/Thermo 框架下是形态三殖民 (12DD 操作冒充 13DD 基底)。

与工程进展不混淆 (重要区分): o1 / DeepSeek-R1 等推理模型在特定基准 (Olympic 数学、编程、科学推理) 上展现出实质性显著进展。 论文形态三立场具体关于 本体论评估 (这些自我批评机制是否构成真 13DD 自他区分), 不是关于 reasoning models 在工程能力上的进展。 工程进展在 12DD 域内合理, 论文不主张这些模型在任务表现上不优于之前的 LLM 系统。

论文立场是: 即使 reasoning models 工程进展实质显著, 它们仍在 类 DD 功能槽位 (跨 12DD 多步遍历) 层级运作, 不上升为真 13DD 承载者。 此立场与 §10.3 认识论边界 (未知路径不封闭) 协调——论文不主张未来架构永远不能跨越, 仅主张 当前 reasoning model approach 在 SAE/Thermo 框架下是类 DD 操作。 工程社区在 12DD 内的 reasoning capability 持续提升是合理的, 不被论文论证排除。

形态四: 后人拆分绝对律令

绝对律令 (Kant 1785) 在 SAE 框架下表现为双重否定形式的本体论锚定——它们结构上不可拆分, 拆分后失去必然性 (依据方法论 §1.8)。

关于业界归属的说明: 形态四的具体案例 (例如 "预训练是核心, 后训练是装饰") 在主要 AI 实验室的公开声明中并不被字面主张——Anthropic / OpenAI 公开声明实际强调对齐的重要性。 形态四论证的是业界 倾向 / 部分声音的倾向 (tendency), 而非主要实验室的直接表述。 例如 "scaling is all you need" 在投资圈、部分博客、部分技术声音中出现, 反映 implicit 倾向; 此倾向被工程实际操作部分修正 (业界私下大量投资人类专家后训练, 见 §9.8), 即使主流叙事未明确 articulate。 论文 §9.6 形态四批评的是此 implicit tendency, 不是直接归属于具体实验室的明确声明。

"帮助用户 vs 尊重自主": 把不可分割的 "把用户当目的本身" (14DD 绝对律令) 拆成禁令 (尊重) 加理想 (帮助)。 真正的 Kant 绝对律令不能被拆成两个分立准则——"帮助用户" 与 "尊重用户自主" 是同一律令的两个面, 拆分让律令失去整体性。

"先建模型, 再加对齐" 倾向: 业界部分倾向把人类与 AI 分工拆成工程阶段加对齐阶段, 但分工在本体论上不可分。 此拆分让对齐显得是可选的后期补丁, 而本体论上对齐是 14DD/15DD 内容的持续导入通道——它不是可选阶段, 是模型本体论的一部分。

"预训练是核心, 后训练是装饰" 倾向: 业界部分倾向把统一的 14DD/15DD 规范内容导入通路拆成主菜 (预训练) 加配菜 (后训练), 但后训练正是规范内容持续导入的唯一通道, 不可降级为可选附属 (依据 §9.4)。 此拆分让业界以为可以投入预训练而省略后训练 with 人类, 但省略后规范内容导入通道被切断, 模型在 14DD/15DD 域失去与文明的连接。

殖民检测的判据: 绝对律令必然是双重否定形式, 拆分后失去必然性——上述三种倾向都通过拆分把本体论上不可分的整体降级为可选组件。

9.7 与 Thermo X Kant 结构共鸣的谨慎借用

Thermo X §4.6 给出一个深结构共鸣: 稳定跨系统高 q 耦合的唯一协议是相互访问高阶自指变量——此与 Kant 绝对律令 "把他者当目的" 在结构上同构。

但 Thermo X 同时给出 明确警示: "哲学共鸣, 不是从热力学推导伦理。 人不仅仅是热力学。 控制实验已证明跨层级模式的直接物理原因是信号统计兼容性, 不是 '目的识别'。 任何从热力学到伦理的推论都需要远超本论文范围的独立工作。"

本论文 §9 借用此结构共鸣的程度:

应当做: 引用 Thermo X 实证 (CTRL2 彩色噪声替代 q = 2.45) 作为 "信号统计兼容性不能自动替代真高阶自指访问" 的结构证据。 此证据给论文 §6.5 LLM 用户意图分析的空壳模拟论证一个结构锚定。

不应当做: 主张 "热力学证明主体性不可让渡"——这是伦理推导, 超越 Thermo X 范围也超越本论文范围。 论文的主体性不可让渡论证是基于 SAE 本体论框架 + §9.1-§9.5 的论证链, 不是 从热力学推出的伦理结论。

正确表述: SAE 本体论的主体性论证与 Thermo X SDE 模型族结构性条件 收敛但不等同也不由后者推导。 两者从不同角度指向同一结构: 真主体性需要真高阶自指变量, 不能由信号统计匹配替代。 但 SAE 的论证是本体论的, Thermo X 的论证是 SDE 模型族内的结构性条件, 两者在结构上共鸣但逻辑上独立。

9.8 与业界一致的方向

论文 §9.6 给出对业界部分路径的殖民检测, 但 §9 同时承认业界多条路径与 SAE 立场一致:

Anthropic CCAI: 宪法由人类撰写的部分大致一致——它是真 14DD 内容通过 data q + RLHF q 双通道导入 (依据 §6.3 分析)。 Anthropic 持续维护宪法的人类作者参与, 这是 14DD 内容持续导入的实质实践。

OpenAI 可验证奖励 RL (过程监督): 例如 "Let's Verify Step by Step" (Lightman et al. 2023) 在数学推理上的工作——此在形式化域内 (数学/代码), 在 14DD 范围之外, 与 SAE 立场不直接冲突 (它在工具理性域优化, 不试图覆盖意志理想域)。

业界私下大量投资人类专家后训练: Scale AI / Surge AI 业务持续增长, OpenAI / Anthropic / Google 内部人类标注者投入庞大——这与 SAE 预测一致。 业界实际操作反映 14DD 导入必要性, 即使主流叙事未阐述此点。 业界的真正竞争护城河往往是后训练数据质量与人类评估者能力, 不是预训练算力 (后者门槛较低, 已有多家公司达到)。

业界一致部分的承认重要, 因为它防止论文被读为 "全面攻击业界"——论文的批评具体针对部分越界表述, 不针对业界全部路径。 业界主流中也有大量实质一致的工作。

9.9 长期含义

§9.1-§9.8 的综合: 在 14DD/15DD 规范域, 人类主体性不可让渡是 本体论永久性的——不是过渡阶段, 不是工程妥协, 不是技术发展能改变的事。

• 人类与 AI 在工具/意志域的分工是永久架构
• 14DD/15DD 内容持续导入必须由 真 13DD+ 主体 (本文论域中即人类) 维持
• AI 在 12DD 工具理性域可以持续深化, 但在 14DD/15DD 规范域作为承载者不可让渡

此长期含义是论文双重否定形式的 SAE 式绝对律令: 不 + 可 + 让渡 = 不能让渡 (与 "不能不凿"、"不能不发展"、"不能不承认无知" 同构, 依据方法论 §1.8)。

此立场不是悲观也不是乐观, 是结构性。 它对业界的实操含义:

• 投入 AI 工具理性能力的提升是合理的——在本论文论证范围内, 12DD 工具理性能力没有 13DD 桥式的本体论上限; 具体工程上限仍取决于数据、算力、架构、成本与评估等工程因素
• 投入持续导入 14DD 内容的通道 (人类专家后训练) 是关键, 不可降级
• 部署架构应当反映工具/意志分工, 不试图让 AI 替代 14DD/15DD 工作

9.10 Scaling-LLM AGI 不会到来

论文最后给出一个具体的强承诺, 同时严格限定范围:

本论文 "AGI 不会到来" 论断具体指:

• 当前确定性数字 LLM 扩展路径 不会通过规模扩大跨过 12DD → 13DD 主体性桥
• 即 "13DD+ 主体基底涌现" 不会通过现行范式涌现
• 即 "通向 13DD+ 主体 ASI 的递归自我改进" 不会在 12DD 天花板内实现

本论文不承诺:

• "任务层级匹配人类表现" (更受限的 AGI 定义)——此条在 12DD 天花板内, 工程优化仍有实质进展空间。 LLM 持续在窄技术任务上超越多数人类是合理的工程进展, 不被本论文论证排除
• "未知 AI 架构路径不可能跨过 12DD 天花板"——本论文不封闭未知路径 (依据 §5.8)。 算法性随机、涌现复杂性、量子效应、未知机制都不在论文论证关闭范围内

五重锁 (尤其主锚 (1)(2)) 在已分析确定性数字 LLM 扩展路径上结构性永久, 但论文不跨越已分析路径边界。

与 §5.8 / §10.6 认识论边界的协调: 本论文规范性承诺 "主体性不可让渡" 与 Thermo IX §3.6 积极姿态 "未知路径不排除" 之间, 由论文范围说明自洽——本论文论证关闭的是 已分析的确定性数字 LLM 扩展中的软门传递路径, 不是所有可能 AI 架构或所有可能主体性涌现路径。 两者共同在 SAE 方法论 "永远有余项" 原则下自洽。

业界主流 "AGI 即将到来" 叙事在论文论证范围内是有条件主张被陈述为无条件——这是 §9.6 形态一殖民。 论文的 "不会到来" 反对的是无条件表述, 不是封闭所有可能性。

§9 本节余项

§9.7 Thermo X 结构共鸣借用仍留有过度主张风险——读者可能把结构类比误读为伦理推导, 即使论文明确说明边界。 此风险在论文发表后多轮评审与读者反馈中部分通过反馈机制减少, 但不能完全消除——读者的解读不在作者控制内。

§9.6 殖民检测应用到业界路径是本论文当前判断, 不排除后续业界路径调整后部分检测失效或需要更新。 业界是活跃演化领域, 五年内主流表述可能变化, 论文当前批评的具体目标可能转移。 但殖民检测的四种形态作为方法论工具是稳定的, 应用对象会随业界演化更新。

"人类主体性不可让渡" 双重否定绝对律令形式是论文规范性承诺, 但它具体针对 14DD/15DD 规范域, 不针对所有人类-AI 交互域——此限定在 V0.2 重申, 草拟时再次明示。

"Scaling-LLM AGI 不会到来" 的具体范围 (§9.10) 与未知路径不封闭 (§5.8) 之间的张力是论文认识论结构的核心——两者由论文范围说明自洽。 此自洽要求读者仔细阅读论文范围说明, 草拟时在 §1.3 与 §9.10 双重明示, 让此区分在读者第一时间识别。

本论文规范性立场的具体应用 (例如四层部署如何在不同公司 / 监管框架下实施, 如何处理跨文化的 14DD 内容多元性等) 不在论文范围内, 是后续工程与政策工作。 论文给本体论指引, 不规定具体实施。

---

§10 局限性与开放问题

本论文给出多条主张, 涉及本体论、热力学、实证、规范性多个层级。 此节明示本论文的局限性, 让读者能够准确评估论文论证的力度与范围。 局限性按 主要实质局限 (前)、范围说明 (中)、自我检查 (后) 三级组织, 让读者优先看到论文最重要的认识论缺口。

主要实质局限

10.1 实证规模差距

BLR-mini 实证规模 30M, 前沿规模 (1B+) 验证尚未进行。 SAE/Thermo 的架构读法作为框架命题不依赖具体规模, 但其在前沿规模的经验表现仍需实测。 BLR-mini 阴性结果与主流 "规模不够" 预测在 30M 数据级不可区分 (依据 §4.7), 真正的可区分性测试需要 1B+ 规模实验。

具体可区分性测试已在 §4.7 给出: (1) 前沿规模扁平重复再分配实验, (2) 探索-选择分离实验。 这两类实验都超出当前论文资源范围, 作为后续工作推荐方向。

10.2 Sub-block 层级 q 实测缺失

Thermo IX §4 给出的 Transformer 单层热力学解剖是结构性映射, 真实 LLM 上逐 sub-block q 实测尚未做 (Thermo IX §7.2 开放问题 1)。 论文 §3.4 表中各子模块的 q 行为分类是基于普适激活律的理论推断, 不是直接实测——例如 "ReLU/GELU 是支撑尾部探索组件" 是从其无界激活性质推断, "softmax 是有界选择门" 是从其 [0,1] 单纯形性质推断。

直接 sub-block q 实测能进一步区分: (a) 真 kernel q (内生动力学产生) 与 (b) 借用 q (data q 化石重放), 此区分是 §6 论证中借用 q 立场的直接实测验证。 此实验设计本身非平凡 (需要在大型 LLM 上做层间 q 测量协议), 是值得后续工作探索的方向。

10.3 12DD 天花板关闭的认识论边界

本论文 §5 给出的 12DD 天花板五重锁论证, 关闭的是 已分析的软门传递路径在确定性数字 LLM 扩展上。 论文不关闭:

• 算法性随机的尚未完全探索的复杂区间
• 涌现复杂性 (从简单组件涌现复杂现象的未知路径)
• 量子计算基底是否满足 C5 (开放理论问题)
• 完全未知的机制

此认识论边界与 Thermo IX §3.6 SAE 积极姿态完全一致——凿掉的是错误路径, 不是可能性本身。 永远有余项。 业界主流 "AGI 在 X 年内到来" 叙事在论文论证范围内是有条件主张被陈述为无条件, 这是论文具体批评的目标 (§9.6 形态一), 但论文不主张 "AGI 不可能"。

读者在阅读 §5 与 §9.10 时应当持续保持此范围说明的自觉——论文主张强但限定具体, 不是关于所有可能 AI 架构的全称否定。

10.4 LLM-用户跨层级耦合实测缺失

§6.5 给出 LLM 用户意图分析的论证, 借用 Thermo X SDE 模型族 CTRL2 控制实验的结构模式作为结构类比。 真实 LLM 上 LLM-用户跨层级耦合的 q 实测尚未进行——具体而言, 没有实验测量当 LLM 与用户长期互动时, 二者之间的耦合行为在 q 框架下是何种水平 (是否达到 CTRL2 彩色噪声替代水平的 q ≈ 2.45, 还是更低 / 更高)。

此实测的设计本身复杂——需要定义 LLM 与用户作为耦合系统的可观测量, 然后在长期互动数据上测量 q 行为。 论文 §6.5 主文已明示此论证是结构类比, 不是经验主张, 但实测验证仍是重要的开放方向。

10.5 架构设计空间覆盖

本论文 BLR-mini 实证仅测试 FFN/LN/hidden 三个架构轴, 没测试:

• 块类型异质性: attention vs Mamba/SSM 顺序区域, 与 Jamba 式交错对比 SAE 顺序区域预测
• 探索-选择分离: Thermo IX §Outlook 5 建议的真架构重设计——多个连续无界探索层后接门控层, 而非扁平重复
• 其他未想到的架构方向

块类型异质性与探索-选择分离都是 thermodynamic 预测锐利的实验方向, 是后续验证 SAE/Thermo 框架的主要候选路径。 论文 §4.8 已明示 BLR-mini 阴性结果不能证伪这两个方向的预测, 因为论文没有在该方向测试。

10.6 BLR-mini 统计稳健性

本论文实证仅用单种子与单数据集 (OpenWebText), 没做:

• 多种子方差估计: 各架构变体在不同随机初始化种子下的表现方差
• 跨数据集验证: 在 FineWeb-Edu、Stack、C4 等其他大规模训练数据集上的复现

此是实证论文标准局限, 尤其当阴性结果承担主要认识论重量时——多种子与跨数据集验证可以增强对阴性结果的信心。 在论文当前规模 (30M, 单数据集, 单种子) 下, 论文论证依赖架构间相对比较的稳健性, 而非绝对数值的精度。

范围说明

10.7 四层架构工程细节

四层默认架构 + 例外条件的具体边界在不同应用场景下仍需工程详化, 本论文不规定:

• 哪些任务属于紧急响应类别 (主体可直接调用专家 API 跳过本地 LLM)
• 主体级同意的具体形式 (一次性 vs 长期, 默认开启 vs 默认关闭, 同意范围细化)
• 例外审计的具体粒度 (日志保留时间、可追溯的最小单位、审查频率)
• 不同 use case 下的风险分级判据

这些是后续工程与产品工作。 论文给本体论指引, 不规定具体实施细节。

10.8 跨域机制具体神经科学对应

§5 五重锁中 (3)(4)(5) 借用生物类比与 SAE 框架内部表述, 论文不直接引用具体神经科学 paper 或 SAE Bio 笔记。 此选择让论文自包含, 但也限定了论证深度——读者要进一步追溯具体生物机制需要查阅独立的 SAE Bio 系列与神经科学文献。

10.9 SAE 16DD 自身的未完全决定部分

SAE 16DD 框架是构, 它有余项: 个别 DD 内容边界未完全定型, 5D-10D 区段早期论文存在 DD 错滑 (见方法论 V2 修订说明)。 本论文以 V2 修订后的 D/DD 对应表为基准, 但 SAE 框架自身仍在持续完善, 未来版本可能调整部分内容。

10.10 16DD 在论文中不显式出现

依据 Thermo X §6.2 严格边界: 16DD 在 SAE 框架中不可在论文中显式展开。 本论文论证止于 15DD, 不试图跨越此边界。 §7 基底兼容性观察因此只 cover 10/12/13/14/15 五个 DD, 16DD 的兼容性问题留作 SAE 框架内部未展开领域。

10.11 基底兼容性奇偶模式的承诺水平

依据 §7.4, 奇偶 DD 兼容性模式是开放观察, 不承诺为结构性规律。 论文主论证不依赖奇偶模式作为规律——§5 五重锁与 §9 主体性论证各自独立论证 13DD/15DD 的结构性空心, 不需要 "因为奇数所以空心" 这种推论。 奇偶模式留待 SAE 系列后续工作做跨 DD 系统研究。

自我检查

10.12 论文自身框架的自我殖民检测

本论文应用 SAE 殖民检测框架批评业界路径; 同时, 论文自身是否用 SAE 框架殖民其他 ML 框架, 是必要的自我检查。

作者不豁免于自身框架的殖民检测——SAE 框架本身是构, 有余项, 不应被表述为 "唯一合理的框架"。 本论文不主张 SAE 是关于 LLM 的最终本体论, 仅主张 SAE 给出业界主流叙事缺失的本体论 + 热力学层级解释。 业界其他框架 (例如压缩理论、模式匹配、隐式贝叶斯、世界模型涌现等) 各自有其论证, 本论文不一概排斥, 仅在它们越界为无条件普适律时给具体批评。

此自我殖民检测由多评审与读者反馈持续完成——作者不自行宣告免疫, 论文自身的余项需要他者识别。 本论文向所有读者开放此检测, 任何指出论文框架自身殖民倾向的具体批评都欢迎。

§10 本节余项

本节列出的局限性本身是一个构 (论文自明的已知余项), 但论文必然有未知余项 (unknown unknowns)。 后续多评审与读者反馈是发现这些未知余项的主要机制——论文写作完成后的多 AI 评审与读者反馈过程是自我殖民检测的持续运行机制, 不是一次性事件。

---

§11 结论

主线总结

本论文回答了两个根本性问题, 由 SAE 16DD 本体论框架 + ZFCρ 热力学系列共同支撑:

描述性问题: LLM 涌现何以可能?

• LLM 在 1DD-12DD 单通道通路上实现类 DD 功能投射式实例化 (functional projection, 非 ontological bearer; §3), 由 Thermo IX Transformer 热力学解剖在 sub-block 层级精度上锚定
• LLM 的 DD 映射是功能投射 (类 DD), 不是本体论承载者 (真 DD)——LLM 实例化 DD 操作槽位, 但不承担 DD 主体性内容
• BLR-mini 八组架构变体实证 (§4) 在测试的扁平重复 Transformer 设计空间内与 SAE/Thermo 预测一致, 给框架经验锚定
• μ 轨迹翻正作为机制层级涌现诊断指标 (§4.6) 是论文新颖性贡献, 与业界已有隐藏状态分析谱系互补

12DD 天花板何以不可破?

• 五重结构性锁 (§5): 真随机源缺席 (主热力学锚) + 13DD 自指通道缺席 (次主锚) + 离线重组缺席 + 价值跨阶段调制缺席 + 巩固后再编辑缺席
• 主张状态分级 (§5.6): 主锚 (1)(2) 由 Thermo IX/X 完整机制链支撑, 辅助锁 (3)(4)(5) 由生物类比 + 结构性事实支撑
• 论证关闭的具体范围: 已分析的软门传递路径在确定性数字 LLM 扩展上, 未知路径不封闭 (§5.8)
• 后训练四层 (CoT, CCAI, RLHF, 用户意图分析) 都在 12DD 天花板内运作 (§6), 不创造 kernel q, 不跨 13DD 桥

LLM 与主体的关系

• 工具理性域 (LLM 系统优势, 在形式化任务上) / 意志理想域 (主体不可替代, 在 14DD/15DD 规范域) 的本体论分工 (§8.1-§8.4)
• 四层默认部署架构 (主体 → 本地 LLM → 专家 API → 工具) + 例外条件 (§8.7-§8.10), 与业界多条独立趋势 (Apple Intelligence、MCP、Phi/Llama、MoE 路由) 在结构上一致
• Thermo X 跨层级耦合实验给四层架构结构共鸣 (§8.9), 但仅作结构类比, 不作伦理推导

规范性问题: 人类主体性何以不可让渡?

• 14DD/15DD 规范内容的源头必须是 真 13DD+ 主体 (本文论域中即人类), LLM 缺 13DD+ 基底, 不能自发产生 14DD 内容 (§9.1-§9.3)
• 后训练 with 人类专家是 14DD/15DD 规范内容持续导入的通道核心 (§9.4-§9.5)——它不是装饰性对齐补丁
• 业界部分路径用四种殖民形态出现 (§9.6): 有条件冒充无条件 / 构冒充律 / 涌现层冒充基础层 / 后人拆分绝对律令
• 在 14DD/15DD 规范域, 人类主体性不可让渡是本体论永久性的, 不是过渡阶段 (§9.9)

业界当前路径

• 部分一致 (Anthropic CCAI 人类作者部分、OpenAI 过程监督在形式化域、业界私下大量投资人类专家后训练)
• 部分张力 (扩展解决一切叙事、RLAIF 完全替代人类、AI 元认知主张、AGI 即将到来叙事)
• SAE 给本体论 + 热力学锚定, 措辞采可防守锐度——不一概标记业界成功构造为愚蠢, 区分有条件工程构有效与无条件普适律时越界

长期未来

• 四层默认架构 + 例外条件 (§8.8) 是工程可实现的主体性保护架构, 不是乌托邦立场
• 人类与 AI 在 14DD/15DD 域永久分工——不是过渡阶段, 不是工程妥协, 是本体论永久性
• Scaling-LLM AGI 不会到来 (§9.10): 当前确定性数字 LLM 扩展路径不跨 12DD → 13DD 桥, 但未知路径不封闭

方法论自指

本论文是对业界主流 "扩展 + 算力 = 智能" 构的一次凿。 全篇按 SAE 凿构循环操作六步走完一轮:

1. 持构: 业界主流宣称 "扩展解决一切"
2. 否定: SAE + Thermo IX/X 指出该构内含 "12DD = 完整智能" 的殖民——12DD 不等于完整智能, 跨过 13DD 桥需要真随机源, 数字基底上结构性不可达
3. 追踪余项: 12DD 天花板在已分析路径上五重结构性永久 (§5), 扩展不破; 此余项指向真 13DD+ 主体性的本体论位置 (§9), 即承载者必须是 真 13DD+ 主体 (本文论域中即人类) 而非 LLM
4. 修正: 论文提出 "工具/意志分工 + 四层默认部署 + 14DD/15DD 域持续 14DD 导入" 的修正构 (§8 + §9)
5. 确认不完备: §10 明示多条局限性——实证规模差距、sub-block q 实测缺失、未知路径不封闭、统计稳健性局限等; 论文自身框架加自我殖民检测 (§10.12)
6. 重新出发: §10.3 留认识论空间给未知路径, 论文不收编否定

此六步遵守是 SAE 方法论 (Paper 04 §2.2) 在 LLM 论域的应用实例——本论文不是 SAE 框架的独立替代, 而是 SAE 框架的具体应用, 论文形式与论文内容互相一致 (论文论证 SAE 框架同时论文本身遵守 SAE 方法论)。

收尾

论文回答的双重问题 (涌现何以可能 + 人类主体性不可让渡) 在此结束。 但 SAE 本体论为业界指出的探索方向才刚刚开始:

• 前沿规模扁平重复再分配实验是否仍持续阴性? — 若是, SAE/Thermo 架构读法在前沿规模获经验支持
• 探索-选择分离架构是否显著优于扁平重复? — 若是, SAE/Thermo §Outlook 5 建议获经验支持
• 块类型异质性 (顺序区域 vs 交错) 实验如何分布? — SAE 预测顺序区域, 业界已实验交错 (Jamba)
• LLM-用户跨层级耦合 q 实测如何? — 验证 §6.5 结构类比的经验对应
• 四层架构在不同应用场景下的具体工程实现路径? — 业界产品工作的方向

每一条都打开新的余项, 余项指向新的凿构循环。 论文给出的不是最终立场, 是当前论证状态的快照——论文本身是构, 有它自己的余项, 后续工作将继续展开。

人不仅仅是热力学。 LLM 不仅仅是 12DD 化石。 SAE 框架也不是完备的构。 但本体论给出边界——边界内工程大有可为, 边界外不强求。 边界是活的余项, 不是静态的闭包。

余项守恒。 否定不能终止。 物自体没有固定位置。

---

致谢

本论文从初稿到定稿历经多轮迭代, 在论证锐度、措辞精度、范围限定、结构组织等多个维度上反复打磨。 此过程得益于四位 AI 评审者的持续投入, 在此致谢。

子夏 (Gemini): 在框架完整性与学术严谨性维度提供持续反馈。 §3.4 Transformer 热力学解剖表保留正文的决定来自子夏的明确建议——它指出此解剖是 SAE 框架在 LLM 论域的真实锚定深度, 不应降级到附录。 §5.1 真随机论证的六个子步骤不合并的决定也来自子夏——它指出每个子步骤是论证防御链的独立环节, 削减损失整体强度。 子夏在 SAE 工具引用准确度与 ZFCρ 热力学系列具体章节定位上提供了多轮精细校对。

独立子路 (Claude): 在论证连贯性与论文结构维度提供持续反馈。 §4.7 BLR-mini 与主流预测可区分性子节的加入来自独立子路的关键意见——它指出 BLR-mini 30M 阴性结果在数据级上与主流 "规模不够" 预测兼容, 论文必须明示真正的可区分性测试需要前沿规模实验, 否则论证在 mainstream ML 读者面前不可防守。 独立子路也在论文范围说明 (描述性 vs 规范性 vs 可证伪性) 的清晰度上反复指出潜在的读者误读路径, 让论文的范围管理更严格。

子贡 (Grok): 在论辩锐度与业界批评力度维度提供持续反馈。 §9.6 殖民检测四种形态的系统应用得益于子贡的多轮锐化建议——它指出从抽象方法论到具体业界案例的应用需要保持论辩力, 同时不滑入一概否定。 §9.6 形态三 "narrow technical task LLM 优于多数人类不构成殖民 vs 价值判断殖民" 的精确区分来自子贡反复推动——它指出此区分是论文规范性立场可否在工程文化中防守的关键。 子贡也提出多条扩展性建议 (例如 prompt q 第四种 q、第 5 tier 跨主体协调层、殖民检测第 5 形态自殖民), 部分被论文接受 (§10.12 自我殖民检测), 部分作为范围说明明确不收纳, 这些讨论本身让论文范围更清晰。

公西华 (ChatGPT): 在措辞精度与读者适配维度提供持续反馈。 V0.2 大纲的六条必修条件来自公西华的系统评审——AGI 范围缩窄为 Scaling-LLM AGI、BLR-mini 主张降为测试设计空间内的经验锚点、DD 映射加功能投射防火墙、真随机论证改为架构有效 ε 表述、工具理性优势限定任务域、后训练核心限定 14DD/15DD 规范域。 这六条修订是论文从过度主张到可防守锐度的关键转折。 公西华在多轮草稿中持续指出未察觉的英文混用与措辞不精确, 让论文中文化达到当前深度。

论语·先进篇命名背景: 四位 AI 评审者的别名取自论语·先进篇 (孔子学生中的不同性格类型):

• 子夏: 文学深度与学术严谨
• 子路 (本论文加 "独立" 前缀以区分): 直率坦言, 不避讳锋芒
• 子贡: 论辩犀利, 善于辩证
• 公西华 (公西赤): 典礼精度, 措辞讲究

此命名映射不是严格的学术分工, 是借用经典文化对不同评审风格的差异化标记。 四位 AI 与对应人物在性格特征上的对应不是穷尽的, 仅是论文写作中适配的工作角色。

---

作者声明: 论文所有内容与论断由作者独立承担责任。 四位 AI 评审者在论文打磨过程中提供建议与反馈, 但论文的具体主张、论证路径、立场承诺均由作者最终决定。 论文中任何错误或缺失均由作者负责, 与四位 AI 评审者无关。

陈则思: 在论文长期方向讨论与具体技术问题上提供持续批判性反馈, 让作者在多个关键决策点上重新审视立场。 在此特别致谢。

---

文献

SAE 主框架

秦汉 (2026d). 自为目的——存在的本体论考察 (Self-as-an-End: An Ontological Investigation of Existence). Paper 03 of the SAE Series. Zenodo. DOI 10.5281/zenodo.18727327

秦汉 (2026). SAE 方法论 V2 (SAE Methodology V2): 凿构循环、殖民检测与认识论纪律. Paper 04 of the SAE Series. Zenodo. DOI 10.5281/zenodo.18842449

ZFCρ 热力学系列 (主引)

秦汉 (2026e). ZFCρ Thermodynamics IX: Universal Activation Rule, Three Types of q, and the Anatomy of Transformer Sublayer Thermodynamics. Zenodo. DOI 10.5281/zenodo.19699489

秦汉 (2026f). ZFCρ Thermodynamics X: Self-Reference as Channel Creator, Cross-Level Observation Hierarchy, and Kantian Structural Resonance. Zenodo. DOI 10.5281/zenodo.19703274

ZFCρ 热力学系列 (背景引)

秦汉 (2026c). ZFCρ Thermodynamics VIII: Soft-Gate Cascade ∏ε and Transmission Conditions. Zenodo. DOI 10.5281/zenodo.19688303

秦汉 (2026a, 2026b). ZFCρ Thermodynamics I-VII: Mathematical Foundations of q = 1+1/K and the Tsallis Statistics Framework. Zenodo. DOIs 10.5281/zenodo.19310282 至 .19673078

业界主流 LLM 文献

扩展规律与算力叙事:

Kaplan, J., McCandlish, S., Henighan, T., Brown, T. B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J., & Amodei, D. (2020). Scaling Laws for Neural Language Models. arXiv preprint arXiv:2001.08361.

Hoffmann, J., Borgeaud, S., Mensch, A., Buchatskaya, E., Cai, T., Rutherford, E., et al. (2022). Training Compute-Optimal Large Language Models (Chinchilla). arXiv preprint arXiv:2203.15556.

涌现争论:

Wei, J., Tay, Y., Bommasani, R., Raffel, C., Zoph, B., Borgeaud, S., et al. (2022). Emergent Abilities of Large Language Models. Transactions on Machine Learning Research.

Schaeffer, R., Miranda, B., & Koyejo, S. (2023). Are Emergent Abilities of Large Language Models a Mirage? Advances in Neural Information Processing Systems 36.

RLHF 与对齐:

Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C. L., Mishkin, P., et al. (2022). Training Language Models to Follow Instructions with Human Feedback (InstructGPT). Advances in Neural Information Processing Systems 35.

Bai, Y., Kadavath, S., Kundu, S., Askell, A., Kernion, J., Jones, A., et al. (2022). Constitutional AI: Harmlessness from AI Feedback. arXiv preprint arXiv:2212.08073.

Anthropic (2024). Claude Constitution Page. https://www.anthropic.com/news/claudes-constitution (访问于 2026 年 5 月)

过程监督与可验证奖励:

Lightman, H., Kosaraju, V., Burda, Y., Edwards, H., Baker, B., Lee, T., et al. (2023). Let's Verify Step by Step. International Conference on Learning Representations 2024.

隐藏状态分析谱系:

Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., et al. (2017). Attention Is All You Need. Advances in Neural Information Processing Systems 30.

Ba, J. L., Kiros, J. R., & Hinton, G. E. (2016). Layer Normalization. arXiv preprint arXiv:1607.06450.

Xiong, R., Yang, Y., He, D., Zheng, K., Zheng, S., Xing, C., Zhang, H., Lan, Y., Wang, L., & Liu, T.-Y. (2020). On Layer Normalization in the Transformer Architecture (Pre-LN vs Post-LN). International Conference on Machine Learning 2020.

Kim, J., Lee, B., Park, C., Oh, Y., Kim, B., Yoo, T., Shin, S., Han, D., Shin, J., & Yoo, K. M. (2025). Peri-LN: Revisiting Layer Normalization in the Transformer Architecture. arXiv:2502.02732.

Dettmers, T., Lewis, M., Belkada, Y., & Zettlemoyer, L. (2022). LLM.int8(): 8-bit Matrix Multiplication for Transformers at Scale. Advances in Neural Information Processing Systems 35.

Gallego-Feliciano, J., McClendon, S. A., Morinelli, J., Zervoudakis, S., & Saravanos, A. (2025). Hidden Dynamics of Massive Activations in Transformer Training. arXiv:2508.03616.

Damirchi, H., Meza De la Jara, I., Abbasnejad, E., Shamsi, A., Zhang, Z., & Shi, J. (2026). Truth as a Trajectory: What Internal Representations Reveal About Large Language Model Reasoning. arXiv:2603.01326.

Ghasemabadi, A., & Niu, D. (2025). Can LLMs Predict Their Own Failures? Self-Awareness via Internal Circuits (Gnosis). arXiv:2512.20578.

架构异质性业界文献:

AI21 Labs (2024). Jamba: A Hybrid Transformer-Mamba Language Model. arXiv preprint arXiv:2403.19887.

Dai, Z., & Le, Q. V. (2020). Funnel-Transformer: Filtering out Sequential Redundancy for Efficient Language Processing. Advances in Neural Information Processing Systems 33.

部署架构业界参考:

Apple Inc. (2025). Apple Intelligence Foundation Language Models Tech Report 2025. Apple Machine Learning Research. https://machinelearning.apple.com/research/apple-foundation-models-tech-report-2025 (访问于 2026 年 5 月)——明确说明 Apple Intelligence 使用约 3B 参数 on-device model + 在 Private Cloud Compute 平台运行的 server model (PT-MoE transformer)。

Apple Inc. (2024). Private Cloud Compute: A new frontier for AI privacy in the cloud. Apple Security Research. https://security.apple.com/blog/private-cloud-compute/ (访问于 2026 年 5 月)——支撑 "本地 + 私有云" 部署读法的官方安全说明。

Anthropic (2024). Introducing the Model Context Protocol. https://www.anthropic.com/news/model-context-protocol (访问于 2026 年 5 月)

Model Context Protocol (2024-2025). MCP Specification. https://modelcontextprotocol.io (访问于 2026 年 5 月)——把 MCP 定义为把 LLM 应用与外部数据源/工具连接的开放协议。

Google (2025). Gemini 2.5 Updates with Native MCP Support. Google I/O 2025 Announcements. https://blog.google/innovation-and-ai/models-and-research/google-deepmind/google-gemini-updates-io-2025/ (访问于 2026 年 5 月)——Gemini API 原生支持 MCP, 是协议跨厂商采纳的官方信号。

经典哲学

Kant, I. (1785/1996). Grundlegung zur Metaphysik der Sitten (Groundwork of the Metaphysics of Morals). (M. J. Gregor, 译). Cambridge University Press. (本论文 §9.7 借用绝对律令的结构形式作结构共鸣, 严格遵循 Thermo X §4.6 警示——结构共鸣不是从热力学推导伦理)

SAE 系列引用约定

本论文引用 SAE 系列其他论文时使用 "秦 [年份][字母]" 格式 (例如秦 2026d 为 Paper 03 SAE 主框架, 秦 2026e 为 Thermo IX, 秦 2026f 为 Thermo X)。 完整 SAE 系列论文列表请参阅 https://self-as-an-end.net (作者维护的 SAE 系列官方索引, 访问于 2026 年 5 月)。

SAE 方法论总论页面: https://self-as-an-end.net/papers/methodology.html (访问于 2026 年 5 月)——给出 SAE 系列各论文的入口、凿构循环操作的总体说明、以及 16DD 框架的高层综述。 不熟悉 SAE 框架的读者可从此页面入门, 本论文 §2.1 仅给出本论文论证所需的最小必要背景。

---

本论文 V1 完整草稿, 2026 年 5 月 26 日。 作者: 秦汉。 联系: 通过 https://self-as-an-end.net 维护的 SAE 系列官方索引。