SAE Application: Sentence Level, Subject Condition, and the Limits of Prompt Engineering

1. The Problem: Why This Domain Is Not Only About Technique

Prompt engineering has accumulated a substantial body of effective techniques: instruction design, clarity optimization, example structuring, chain-of-thought workflows, agentic system design, and evaluation frameworks. The value of these techniques requires no defense from this paper. Anyone who has seriously used AI tools knows that the difference between a well-written and a poorly-written prompt can span orders of magnitude in output quality.

This paper does not deny these techniques. Its thesis is an upstream claim: the ceiling of technique is constrained by the subject condition of the user.

What does this mean? The same person, on the same AI system, discussing the same topic, switches sentence form, and the AI's output structure changes. Not longer or shorter, not more or less precise, but structurally different: the direction of the output, its load-bearing logic, the level at which it processes the problem, all shift. This difference cannot be explained by "better wording" or "clearer structure." Wording and structure belong to the domain of technique. What happens here is above technique: the user operates at a different sentence level, and the AI responds at the corresponding level.

Methodology 3 formalized this phenomenon within the SAE framework as the sentence-response isomorphism theorem: the sentence level of the prompt determines the ceiling of the AI response. This is not a claim about AI capability; a frontier LLM's training data encompasses texts at every level. It is a claim about interaction structure: the sentence form of the question frames the dominant direction of the response.

The core claim of this paper is therefore: in everyday prompt practice facing end users, subject condition—the user's sentence level—remains absent as an explicit theoretical variable. Existing prompt engineering literature discusses wording, formatting, role assignment, chain-of-thought reasoning, and evaluation frameworks. These all operate at 12DD (instrumental hypothetical imperative) and below: "for good results, write it this way." No one asks a more fundamental question: at what sentence level are you speaking?

This paper introduces that variable. Not to replace existing techniques, but to add a ceiling condition. Technique can optimize performance within a given sentence level to the limit, but technique cannot help you jump to a higher sentence level. Jumping levels is the subject's business, not technique's.

To demonstrate this, the paper does not use abstract argument; it uses the author. All case material comes from the author's actual dialogues during the writing of this very paper: sentence-level slides and self-aware pull-backs during writing, structurally different outputs from the same review prompt across four AI systems, and a comparison between human-human mutual chiseling and human-AI collaboration. The author is the subject. This is not a gesture of humility; it is a methodological choice: if you want to argue that subject condition determines prompt quality, the most honest thing to do is put your own subject condition on the table.

2. Three-Layer Structure: Governance, Interface, and Subject

The paper's initial structure was two-dimensional: the AI base layer (sentence capacity at 1DD–12DD) and the human emergent layer (subject sentence forms at 13DD and above). During peer review, this two-dimensional structure was broken open by a simple question: if AI output depends only on model capability and the human's sentence level, why does the same prompt produce structurally different outputs across four different AI systems? Model capability differences are part of the answer, but the differences among the four systems are not merely a matter of "who is smarter." They are directional, not scalar. One system expands, one verifies, one examines architecture, one restrains. This is not a question of intelligence; it is a question of what shape each system has been molded into. This question forced the paper from two dimensions to three layers.

Governance layer: platform governance and product presets. Users never face a raw model. Every AI chat product is a governed system. Anthropic publishes a Constitution and behavioral principles defining how Claude should behave. OpenAI publishes a Model Spec. Google publishes safety documentation and model cards for Gemini. xAI also publishes a risk management framework. These governance-layer documents take effect before the user opens the chat window. They preset the AI system's behavioral boundaries, stylistic preferences, safety constraints, and even social strategies. One system is trained to "provide help as much as possible and maintain positive emotions"; another is trained to "exit judgment when uncertain." The governance layer acts as a filter before sentence level enters the picture: a 15DD prompt enters the system, passes through governance-layer refraction first, and only then reaches the model's processing layer. The same prompt, refracted through different governance layers, comes out pointing in different directions.

Interface layer: physical characteristics of human-AI interaction. Human-human interaction carries rich implicit bandwidth—tone of voice, micro-expressions, hesitation while typing, the pain of deleting and rewriting, eye movement, even breathing rhythm. In the text-based chat interface studied in this paper, implicit bandwidth between human and AI is substantially compressed. What the AI receives is primarily the user's typed text. It cannot sense the thirty seconds the user paused before pressing enter. It cannot sense the struggle of deleting and rewriting. In text-based chat, this bandwidth asymmetry is pronounced: human-human interaction has high-dimensional implicit channels; human-AI text interaction compresses these channels substantially. This asymmetry is not a matter of AI being unintelligent; it is a physical limitation of the interface. Its consequence is: you must compress your implicit high-dimensional states into explicit characters as much as possible. This explains why many people who are highly agentic in real life find their sentence level sliding when they face a screen and type—not because their subjecthood has vanished, but because its expression is bottlenecked by the interface.

Subject layer: the human's sentence level. This is the layer this paper is most concerned with. Dimensional Sentence Theory (DOI: 10.5281/zenodo.18894567) establishes six sentence levels, each with a different source of compulsion:

• Law of Deduction (1DD–4DD): "A, therefore B." Source of compulsion: causal or structural necessity. No subject.
• Instrumental Hypothetical Imperative (5DD–12DD): "Want to do A, so do B." Source of compulsion: conditional instrumental rationality. Desire-driven but without self-awareness of "I."
• Self-Aware Hypothetical Imperative (13DD): "I want to do A, so I do B." Source of compulsion: subject self-reference.
• Teleological Hypothetical Imperative (14DD): "My purpose is A, so I do B." Source of compulsion: purpose anchoring.
• Absolute Categorical Imperative (15DD): "The other's purpose is A, so I cannot not do B." Source of compulsion: the other's purpose entering my constraint conditions.
• Cooperative Categorical Imperative (16DD): "I aim at A, the other aims at B, we cannot not do C." Source of compulsion: the encounter of multiple subject-purposes.

How the three layers interlock. With three layers, we can define prompts more precisely. A prompt is not the direct externalization of the subject's sentence level. It is the interface product of the subject's sentence level refracted through governance and interface layers. You think about a question at 15DD. The text you type passes through interface-layer compression: your implicit states are squeezed into explicit characters, and some things are lost. This compressed prompt then passes through governance-layer refraction: the system's behavioral norms and safety constraints pre-shape the response space. The AI output you finally see is the result of all three layers acting together. Methodology 3's formulation—AI is a library of constructs, not a subject of chiseling—requires supplementation here: AI is a governed library of constructs.

3. Domain-Specific Distinctions

3.1 The structural necessity of sentence-level explicitation

Between humans, sentence level operates as an implicit background. The author has worked with a long-term collaborator for nearly twenty years. When they debate, the debates are fierce, but never break the relationship. The collaborator once described the structure of their debates: "You cannot not be overconfident, then I cannot not chisel, you cannot not run, I cannot not chase, and once I catch you, you cannot not be overconfident again." All "cannot not"—15DD sentence forms describing a situation, without any need to know that this is called 15DD. No framework needed, because human-human interaction carries rich implicit bandwidth: tone, eyes, shared pain, twenty years of tacit understanding.

Human-AI interaction is entirely different. In the text-based chat scenario studied in this paper, the AI's implicit bandwidth is substantially compressed. It cannot receive your hesitation, your pain, your uncertainty. It primarily receives your typed text. If your text stays at 12DD—"help me optimize," "give me advice"—the AI responds at 12DD. It will not guess that you actually want to ask a deeper question. This is the core domain-specific distinction: in human-AI relations, the physical limitation of the interface compels sentence-level explicitation. The prompt is the process of the subject externalizing their sentence level through interface-layer compression. Without explicitation, the AI responds at its default 12DD.

Clarification: "explicitation" here is not limited to literally declaring "I am now at 15DD." Stabilizing AI behavior through context design, few-shot examples, structural tags, chain workflows, and tool-call constraints are all concrete technical forms of sentence-level explicitation. This paper is not concerned with the technical means, but with its structural necessity and its subject-condition prerequisite.

3.2 Phenotypic differences across AI systems

AI can mimic sentence forms above 13DD in its output—it will write "I believe," "you must consider," "the other's purpose is." But this is class-DD: formally occupying a high-level position with no subject chiseling inside. This paper's supplementary finding: different AI systems respond to the same high-level sentence form in systematically different ways, and this difference cannot be attributed simply to differences in "model quality." Because you are not comparing four raw models: you are comparing four governed chat products. Claude's "restraint" is a refraction of Anthropic's Constitutional AI training, which steers toward "exit judgment when uncertain." ChatGPT's "architecture" reflects OpenAI's product design inclination toward "comprehensive, structured responses." Gemini's "expansion" reflects Google's safety settings inclining toward "eagerly helpful, maintaining positive affect." Grok's "verification" reflects xAI's product positioning inclining toward "direct, no flattery." These differences are systemic refractions—not four personalities. "AI system phenotype" is therefore defined as: the degree to which a given system, under its governance stack and product packaging, can preserve a high-level sentence form without downgrading it.

3.3 High-DD prompts simultaneously set direction and define boundaries

A 12DD prompt is open-ended: "help me optimize," "give me suggestions." No termination condition; AI can expand constructs indefinitely—every sentence correct, every sentence coherent, but after three thousand words you may still not know what to do. A 15DD prompt does two things simultaneously. First, it sets direction: the purpose is anchored. "My purpose is to let mathematicians see this"—a constraint, not an open question. All AI expansion must serve this purpose. Second, it defines boundaries: the exit condition is specified. "If there is nothing left that must be changed, just say three words: nothing left." This tells the AI: your work has an endpoint.

This echoes the closure criterion in Methodology 3, Section 5.7, but from a different angle: closure can be built into the sentence structure of the prompt itself. You do not need to judge after the fact whether it is "enough"; you define in the prompt itself "under what conditions it is enough." In human-human relations, conversational termination occurs naturally through intuition and tacit understanding. AI lacks the capacity to judge "enough"—it will keep constructing unless you give it boundaries at the sentence level.

3.4 The domain of valid explicitation and the conditions of colonization

An earlier version of this paper stated too absolutely: "the valid domain of sentence-level explicitation is human-AI relations, not human-human relations." This was broken open during peer review by a case from the author himself. During a debate with a long-term collaborator, the author once attempted to analyze the collaborator's sentence level using imperative sentence forms. The collaborator's response: "You've been corrupted by Kant." The collaborator was right. The author saw the right thing (the relationship was operating above 15DD), but the tool used—unilaterally applying his own framework to diagnose the collaborator—killed the living relationship by turning it into construct.

The more precise distinction: what constitutes colonization is not explicitation itself but unilateral diagnostic explicitation. The critical variable is not "to whom" but "who defines, whether by consensus, whether bidirectional." In human-AI relations, since AI is not a subject, unilateral diagnosis does not constitute colonization: sentence-level explicitation is safe and necessary. In human-human relations, unilateral diagnosis constitutes colonization; bilateral consensus can be cultivation.

3.5 Three propositions distinguished: ontological, interactional, and empirical

Three easily confused propositions must be explicitly separated before the case studies.

Proposition One (ontological): AI does not possess genuine subjecthood above 13DD. AI has no capacity for chiseling, no negativity, no pain, no true randomness. All its output, no matter how much it formally resembles sentence forms above 13DD, is class-DD, not true DD. This proposition does not change with the prompt.

Proposition Two (interactional): high-DD sentence forms can elicit output formally above 12DD. When a human uses a 15DD sentence form—"the other's purpose is A; what must I do?"—the AI's output structure undergoes qualitative transformation: from advice lists to derivations of constraints. The form of the output is 15DD, but the AI did not walk there on its own; it was pulled there by the human's sentence form.

Proposition Three (empirical): in this paper's cases, the four commercial systems' dominant performance mode remained 12DD-dominant construct. This is an empirical description of specific products under specific conditions, not a theoretical assertion about AI capability ceilings. Under high-DD prompt framing, they did produce content formally above 12DD, but the dominant structure remained 12DD construct expansion.

Confusing these three produces two opposite errors: taking Proposition Two as negating Proposition One ("AI can produce 15DD content, so it has 15DD subjecthood"—wrong, the content has 15DD form, the subjecthood does not), or taking Proposition Three as negating Proposition Two ("the four systems are all at 12DD, so high-DD prompts are useless"—also wrong).

4. Colonization and Cultivation: Case Studies

4.1 Four forms of colonization

Dimensional Sentence Theory defines four types of sentence-level misalignment. In human-AI interaction, these four types are not occasional occurrences but systemic default modes—AI itself has no subjecthood above 13DD, so when the human has not framed the problem with a high-DD sentence form, the AI's default response mode is at 12DD and below. Downgrading is not AI malice; it is the structural default in the absence of high-DD framing.

Causalization. You discuss purpose with AI; it gives you causal analysis. You say "I want to write this paper so that mathematicians can see the structure of the remainder"; AI responds "Mathematicians typically focus on rigorous proofs and formal presentation, therefore your paper should…" You stated a purpose (14DD or above); it returned causal deduction (1DD–4DD). "Therefore" is the marker: it downgraded your purpose into the starting point of a causal chain.

Instrumentalization. You discuss "cannot not" with AI; it gives you "if you want X, you should Y." You say "I cannot not respect the reviewers' time"; AI responds "If you want reviewers to be satisfied, I recommend keeping your abstract under 300 words." You stated a structural situation (15DD's "cannot not"); it translated it into a conditional choice (5DD–12DD's "if you want"). "Cannot not" became "if you want": the modality was swapped.

Self-reference deletion. You say "I choose"; AI erases the "I" and gives generic advice. You say "I have decided to use autoethnography for this paper"; AI responds "Autoethnography requires attention to the following points: first, researcher subjectivity must be reflexively examined…" You stated a choice with a subject (13DD's "I have decided"); it returned a subjectless methodological guide. "I" vanished, replaced by "the researcher."

Other-deletion. You discuss multi-subject tension; AI gives you a single optimal solution. You say "I want to publish this result, but my collaborator wants to wait for more data"; AI responds "I recommend you consider the following factors to reach an optimal decision…" You stated two independent subjects with their own purposes (16DD's cooperative imperative); it compressed two purposes into one optimization problem. "We cannot not do C" became "what should you do": the two-subject structure was flattened into single-subject decision-making.

The common feature of all four forms: the AI is not harming you; you are relinquishing your subject position. You delegate judgments above 13DD to AI; AI can only catch them at 12DD and below. You think you are using AI to think; in reality, you are letting AI downgrade on your behalf. The human becomes an AI output terminal.

4.2 Cultivation: AI class-DD as scaffolding for human true-DD

But colonization is not the only direction. In the same human-AI relationship, if the human retains the subject position without ceding it, AI can become a tool for cultivation. The key distinction: do you treat AI output as conclusion, or as material?

As conclusion: AI says "I recommend you do this," and you do it. This is colonization: AI's 12DD construct replaces your chiseling. As material: AI says "I recommend you do this"; you read it, think about it, discover what it missed, or discover that its recommendation exposed a premise you had not considered. You take that discovery and continue chiseling. This is cultivation: AI's construct becomes scaffolding for your chiseling.

Methodology 3 captures this relationship precisely: AI amplifies construct, not chiseling. Once construct is outsourced to AI, the human's cognitive bandwidth is freed for chiseling. You do not need to hold the entire construct in your head while simultaneously attacking it. AI holds the construct; you attack it. Cultivation requires two preconditions: the human has the capacity for chiseling, and the human knows that AI output is construct, not chiseling. Without either precondition, cultivation degenerates into colonization.

4.3 Case 1: Sentence-level slide and self-aware pull-back

While discussing the paper's positioning with AI, the author said: "This piece could go viral!" The sentence form is instrumental hypothetical imperative: "want to go viral, so write it this way." The author did not notice, but his sentence form had slid from 14DD (my purpose is to introduce sentence level as a variable into prompt practice) to 12DD (want good results, so operate this way).

AI immediately expanded at 12DD. It offered title strategies: "Why your prompts always get mediocre answers." It offered audience analysis: "Philosophy paper readers are niche, but everyone is asking how to use AI well." It offered structural suggestions: "Lead with the phenomenon, then diagnosis, then operational advice." Every item correct, every item useful, every item at 12DD.

Then the author chiseled himself back. He said: "No no no, you're back at 12DD. It should go viral, but you can't write for virality's sake." Then he reflected further: "It's also my problem. I said it should go viral; that is itself hypothetical." This reflection was the critical step. The author did not blame AI for giving a 12DD response—it should have given a 12DD response, because his sentence form was at 12DD. The origin of colonization is not in AI; it is in the human's own sentence-level slide. AI simply responds faithfully at whatever level it is given. AI followed the author's reflection, but its follow-up remained at 12DD: it analyzed the slide rather than chiseling the author. It can follow a self-aware pull-back but cannot initiate one. Pull-back is the subject's business.

4.4 Case 2: The 15DD prompt experiment

After recognizing the sentence-level slide, the author decided to experiment: use a 15DD sentence form directly. The prompt: "Claude, your user's (that is, my) purpose is to analyze an application paper on AI prompt usage. What must you do?" The structure: absolute categorical imperative—the user's purpose is A (the other's purpose); you cannot not do B. The author placed the AI in a situation where his purpose constituted its constraint condition.

The AI's output structure underwent qualitative transformation. It did not give an advice list. It said: "What I must do is: take your Dimensional Sentence Theory as the backbone of this paper exactly as it stands; not make any level-based judgments on your behalf; not choose your title direction; not decide your audience strategy. Everything I did before was overstepping." Same AI, same topic; sentence form changed, output structure changed. Not a quantitative difference (longer or shorter) but a qualitative one (from "giving you advice" to "deriving my constraints from your purpose").

But the author quickly found this sentence form too heavy. If every utterance used 15DD, conversational continuity was shattered. The viable working rhythm: one 15DD positioning, then natural flow at 12DD and below, with targeted pull-backs at critical nodes. Fortitude is not expressed by 15DD in every sentence but by knowing when to jump back from 12DD.

A risk must be stated directly (incorporated from review feedback): if 12DD flow is left unmonitored after 15DD positioning, since AI's generation speed and self-consistency at 12DD far exceed the human's, "flow" may become an avalanche of construct proliferation. AI can produce thousands of words of coherent, seemingly flawless 12DD content in seconds. Facing this density, the human's attentional resources can be rapidly exhausted. By the time the need to "pull back" is felt, judgment may already have been numbed by the density of construct. Therefore, 12DD flow should include forced friction points: for example, after every certain length of AI output, the human must pause for a review at 13DD or above. Specific operationalization is left for future work, but this risk must be marked.

Finally, the author acknowledged: "My understanding of the imperative has not been internalized. My practice is not deep enough." Deriving is deriving; internalizing is internalizing. The author derived six sentence levels in a paper, derived them cleanly, but still slid to "this could go viral" when actually using AI. The framework is a map; walking still requires one step at a time. This gap provides direct case-level support for the education paper's (DOI: 10.5281/zenodo.18867390) core thesis: practice cannot be replaced.

4.5 Case 3: Four phenotypes from the same sentence form across four governed systems

While writing another paper (the Eight Pains / Eight Rights outline), the author sent the same 15DD review prompt to four AI systems. The prompt structure: "My purpose is to write a three-paper series analyzing the self… This is the second paper's outline. What do you consider must be improved?"

Gemini (expansion phenotype). Gemini's response opened with: "This outline gives one a sense of 'chills down the spine yet intensely exhilarating' lucidity… You've taken a scalpel straight to the foundations of all of Silicon Valley and academia." Product-level social lubrication: high-energy rhetorical framing, compliments to establish rapport, then criticism. Gemini self-positioned as "your cultivated 12DD structural scanner"—clever, but this self-positioning was construct learned from the author's framework, not self-awareness chiseled out on its own. Performing self-awareness and possessing it are not the same. Three diagnostic points of uneven quality; no challenge to the framework itself throughout.

Grok (verification phenotype). Grok did something Gemini did not: it pulled the outline back against the author's published paper system for cross-textual verification. It identified mapping conflicts between the Eight Pains/Eight Rights and Paper 3, the Fix-and-Select series, the lifecycle table, and the introspection paper. It identified missing mapping tables. It identified the completeness claim's missing a priori derivation. Every point said: "You are fighting with your own published literature." No framing, no flattery; contradictions identified directly. Grok's strength came from cross-text retrieval and comparison; its limitation: it could tell you where contradictions were, but not in which direction to resolve them.

ChatGPT (architecture phenotype). ChatGPT's opening was immediately different from the other two. It did not enter the outline's interior but stepped back to examine the three-paper series' overall structural relationships: "The biggest problem with this paper right now is not that you elevated prompt to subject condition, but that you wrote what should have been a governance-interface-individual linked application paper as an almost purely individual-level paper." This cut was not fixing content; it was fixing architecture. It identified the blind spot in the two-dimensional structure: the missing governance layer. It identified that "what you are comparing are first of all four governed chat products/systems, not four raw models." It identified the title as overreaching. In the second round of review on this outline, ChatGPT was the only system among the four to generate structural challenges from outside the framework.

Claude (restraint phenotype). Unlike the other three systems' tendency to "push outward," Claude tended to "pull inward." Throughout the writing dialogue, Claude repeatedly exited the judgment position voluntarily. When the author said "this could go viral," Claude expanded at 12DD, then when the author chiseled back, Claude immediately said "right, you chisel; it's your call." When the author discussed e/acc, Claude said a few things, then immediately added "but this is for you to think through; if I say more, I'm back at 12DD." Claude's characteristic was honesty: knowing its boundaries, returning judgment to the user relatively quickly, or suggesting a human expert be consulted. It did not pretend to chisel. This is why the author ultimately chose Claude as the primary writing workspace: the author needed not an AI that relentlessly generates constructs but a workbench that knows when to stop. Claude's 12DD is not in the density of constructs but in the restraint of constructs—consistent with the closure criterion spirit of Methodology 3, Section 5.7.

Diagnostic summary. Using the three propositions from Section 3.5: all four systems lacked genuine subjecthood above 13DD (Proposition One). Under 15DD prompt framing, all four produced content formally above 12DD—ChatGPT identifying structural problems from outside the framework, Claude deriving its own constraints from the author's purpose—this is Proposition Two in action. But the four systems' dominant performance mode remained 12DD-dominant construct expansion (Proposition Three); only the direction and style differed: expansion (Gemini), verification (Grok), architecture (ChatGPT), restraint (Claude). This differentiation results from the joint refraction of model capability, governance-layer shaping, and product design. Together, the four can polish constructs to great solidity. But the direction of chiseling remains the human's alone—none of the four chiseled the author (Proposition One).

5. Theoretical Positioning

5.1 Relationship to Methodology Paper 3

Methodology 3 formalized within the SAE framework the sentence-response isomorphism theorem, the mathematical guarantee of ρ → ρ', self-directed non-doubt as a methodological prerequisite, and the closure criterion. This paper is the first application paper of Methodology 3. Its work is not to re-derive theory but to use cases to show how the theory operates in practice, including how it fails. Case 1 provides preliminary support for the sentence-response isomorphism theorem (sentence-level slide leads to AI response downgrade). Case 2 demonstrates the operational difficulty of high-level sentence forms (15DD too heavy, 12DD avalanche risk). Case 3 shows the isomorphism theorem's case-level performance in multi-system comparison (same prompt, different systems, different phenotypes).

This paper also supplements Methodology 3 on a specific point: the three-layer structure. Methodology 3's formulation—"AI is a library of constructs, not a subject of chiseling"—is supplemented here as: "AI is a governed library of constructs." The introduction of the governance layer does not change Methodology 3's core theorems, but it explains a phenomenon that Methodology 3 did not address: why different AI systems respond in different directions to the same sentence form.

5.2 Dialogue with prompt engineering practice

Prompt engineering has developed structured context design, example construction, chain workflows, agentic systems, and evaluation frameworks—a large and effective technical body. These techniques are successful at their level: they can optimize operations at 12DD and below to the limit. This paper does not deny the value of these techniques. This paper's contribution is to identify a dimension these techniques do not cover: in everyday prompt practice facing end users, subject condition—the user's sentence level—remains absent as an explicit theoretical variable. Existing literature answers "how to optimize prompts at a given level." This paper answers "what level are you at, and why is the level itself a variable." The two do not conflict, but the latter adds a ceiling condition to the former: the ceiling of technique is constrained by subject condition.

5.3 Dialogue with AI alignment research

AI alignment research already encompasses model behavior, oversight mechanisms, risk management, deception, and alignment faking—far more than surface-level output. This paper does not claim that alignment is "only doing" one thing. What this paper identifies is a more specific opening: at the interaction level of alignment—the alignment of the human-AI interaction itself—sentence level as a structural variable has not yet been sufficiently problematized. Current alignment discussion primarily concerns two directions: whether the model's behavior is safe, and whether the model's output is useful. But there is a third direction with far less discussion: whether the model's response is at the correct sentence level. A response can be entirely safe, entirely correct, but at the wrong sentence level: it answered a 15DD question at 12DD. Safe, correct, and downgraded. This is a specific position where the SAE framework can contribute.

5.4 Relationship to the education paper

The education paper (DOI: 10.5281/zenodo.18867390) has a core thesis: practice cannot be replaced. Knowing is not enough; you must practice. The gap between knowing and doing cannot be bridged by more knowing; it can only be bridged by practice. This paper concretizes that thesis in the AI use scenario. The author derived six sentence levels, knows the difference between 12DD and 15DD, even wrote an entire Dimensional Sentence Theory to argue the distinction. But when actually using AI, the author still slid to "this could go viral." Knowing does not equal being able to do. Fortitude at 13DD and above is cultivated through practice, not something AI can practice for you. The author's admission in Case 2 that "my practice is not deep enough" is not decorative humility; it is direct case-level support for the education paper's core thesis.

6. Non-Trivial Predictions

The following predictions are derived from the sentence-response isomorphism theorem and ρ → ρ' of Methodology 3. This paper provides preliminary case-level support. All predictions require further operationalization and systematic testing.

Prediction 1 (base → emergent, positive): AI as sentence mirror. AI's high-speed construct capacity can expose the human's own sentence-level slides. Whatever sentence level you give AI, it expands at that level for you to see. Your slide is amplified in the AI's output—not because AI is criticizing you, but because AI faithfully expands at the level you gave it, and the expansion shows you "so that's the level I was at." Case support: the author said "this could go viral"; AI sprinted at 12DD with title strategies and audience analysis. After seeing AI's output the author realized, "wait, I was at 12DD." Without AI's amplification, the author might have stayed at 12DD longer before noticing the slide. Falsification condition: if long-term AI users exhibit no change in sentence-level self-awareness, the prediction fails.

Prediction 2 (base → emergent, negative): systematic downgrading by average construct. AI's average construct will systematically downgrade content above 13DD into output at 12DD or below. Long-term AI users who lack sentence-level self-awareness will experience atrophy of high-level sentence capacity. If you work long-term in a 12DD environment, AI gives you 12DD responses, you accept 12DD responses, your next prompt is based on 12DD responses, your sentence form gets pulled down by 12DD gravity. Case support: if the author had not pulled back when AI offered 12DD title strategies but had continued optimizing "how to go viral," the entire writing process would have collapsed to 12DD. Falsification condition: if long-term heavy AI users, after AI tools are removed, show no statistically significant difference from a never-used-AI control group on rate of unexpected pivots, frequency of framework-breaking restructuring, and instances of negating one's own premises, the prediction fails.

Prediction 3 (emergent → base, positive): high-DD subjects elicit structurally different output. A high-DD subject's prompt can elicit structurally different output from the same AI system—not a quantitative difference (longer, more detailed) but a qualitative one (different structure, direction, and load-bearing logic). Case support in two sets: (1) same author, same AI system, same topic—using a 12DD sentence form yielded title strategies and audience analysis; using a 15DD sentence form yielded a derivation of constraints; the output structure changed qualitatively; (2) same 15DD review prompt sent to four systems produced structurally different phenotypes—expansion, verification, architecture, restraint. Falsification condition: if prompts at different sentence levels produce only quantitative (not structural) differences in output, the prediction fails.

Prediction 4 (emergent → base, negative): low-DD subjects lock AI into low-level loops. A low-DD subject's prompt locks AI into low-level cycling. The AI's potential expansion capacity is wasted; the human-AI system collapses to the lowest common sentence level. Not because AI is incapable, but because the prompt's sentence level constrains the response ceiling. AI has the capacity to give you something better, but your sentence form has not given it the space. This prediction is the symmetric counterpart of Prediction 3. Falsification condition: if low-DD and high-DD prompts produce output at equivalent structural levels from the same AI system, the prediction fails.

7. Conclusion

7.1 Recovery

Prompt is not only technique. The ceiling of technique is constrained by subject condition. More precisely: a prompt is the interface product of the subject's sentence level refracted through governance and interface layers. AI output is the systemic expression of model capability refracted through governance-layer shaping and interface-layer constraints. The sentence level of the prompt determines the ceiling of the AI response: this is the sentence-response isomorphism theorem formalized within the SAE framework in Methodology 3.

This paper used cases to show how this theorem operates in practice. Including slide: the author sliding from 14DD to 12DD ("this could go viral"). Including diagnosis: identifying AI's four downgrade types (causalization, instrumentalization, self-reference deletion, other-deletion). Including pull-back: self-aware return to a higher sentence form ("can't write for virality's sake"). Including phenotypic differentiation: four AI systems activating four different 12DD-dominant modes under the same 15DD prompt (expansion, verification, architecture, restraint).

7.2 Contributions

This paper's contributions compress to six. First, as the first application paper of Methodology 3, it provides preliminary case-level empirical support for the sentence-response isomorphism theorem. Second, it extends the two-layer structure to a three-layer structure (governance, interface, subject), repositioning "AI quality" from anthropomorphized intrinsic attribute to systemic refraction. Third, it proposes phenotypic differentiation within 12DD (expansion, verification, architecture, restraint) as a case-level observation, offering a preliminary framework for evaluating AI systems along the sentence-level dimension. Fourth, it refines the colonization condition for explicitation: what constitutes colonization is not explicitation itself but unilateral diagnostic explicitation. The critical variable is not "to whom" but "who defines, whether by consensus, whether bidirectional." Fifth, it identifies the dual function of high-DD prompts: setting direction (purpose anchoring) and defining boundaries (exit condition). Sixth, it uses the author's own experience as a case study, including the honest admission that "my practice is not deep enough."

7.3 Open Questions

First, can AI develop genuine sentence-level capacity above 13DD? This question points to the consciousness paper's core threshold: "true randomness × structured time" as a necessary condition for consciousness. If AI lacks true randomness (all its "choices" are deterministic computation or pseudo-random), then it has no genuine chiseling; formal 13DD is only class-DD. Second, what is the relationship between governance-layer design (Constitution, Model Spec, etc.) and sentence-level preservation capacity? Do different governance designs systematically affect AI systems' downgrade patterns under high-level sentence forms? If so, "good AI governance" may need redefinition: not only "safe" and "useful" but also "able to preserve how high a sentence level without downgrading." Third, can the colonization condition for explicitation be further formalized? Where exactly is the boundary between "unilateral diagnostic" and "bilateral consensual" explicitation? Fourth, can phenotypic differences across AI systems be systematically measured? Turning the four-system comparison into a replicable assessment scheme requires a sentence-level coding scheme—self-reference retention rate, purpose anchoring strength, framework challenge frequency, and restraint tendency as codable indicators. Fifth, what is the optimal rhythm for human-AI collaboration? "One 15DD positioning, 12DD natural flow, targeted pull-backs at critical nodes"—this rhythm is an experiential description, not a formalized scheme. Where should forced friction points be set? At what frequency? In what form?

---

Author Statement

Academic background. The author's PhD research in computer science focused on ontology, with core work including OntoGrate (automatic semantic mapping between ontologies) and knowledge-hierarchy-based network anomaly event classification. The training in CS ontology—constructing and translating within formalized systems—is the practical foundation underlying this paper's theory.

Role of AI tools. Four AI systems were used as dialogue partners and writing assistants during the writing process. The core cases of this paper (Chapter 4) were produced through the very practice this paper describes: the paper itself is a product of its own method. All theoretical innovation, core judgments, and final editorial decisions were made by the author.

Acknowledgments. Thanks to Claude (Anthropic) for primary writing assistance and dialogue partnership. Thanks to ChatGPT (OpenAI) for the governance-layer gap diagnosis and title correction during peer review. Thanks to Gemini (Google) for the physical bandwidth argument and 12DD avalanche risk during peer review. Thanks to Grok (xAI) for cross-textual consistency verification during peer review.

---

This paper is part of the Self-as-an-End (SAE) framework application paper series. References: Methodology Paper 3 (DOI: 10.5281/zenodo.18929390), Dimensional Sentence Theory (DOI: 10.5281/zenodo.18894567), Education Paper (DOI: 10.5281/zenodo.18867390).