1. The Problem

An intervention may be effective yet invisible to the trial designed to detect it — not because the intervention fails, but because the trial's design is structurally mismatched to the phenomenon.

Specifically: if the response function has threshold structure, if realized exposure within the treatment arm is highly heterogeneous, and if the study neither measures nor exploits this continuous exposure state, then binary assignment combined with binary analysis will systematically dilute the signal from subjects who crossed the threshold.

This problem is not new. Exposure-response analysis in pharmacology (FDA Guidance; ICH E4), MCP-Mod's continuous dose-response estimation, threshold regression and change-point analysis, adaptive enrichment designs, and the implementation fidelity literature in education and rehabilitation science all address aspects of "what to do when thresholds exist."

But these frameworks share a blind spot: they care about where the threshold is and what the effect difference across it looks like. They do not typically ask a more fundamental geometric question: are the distances on each side of the threshold symmetric?

If the distance from intervention onset to threshold (the sprouting distance) is much larger than the distance from threshold to full effect establishment (the establishment distance), i.e. the asymmetry ratio r >> 1, then signal dilution is not accidental but structural. Increasing sample size cannot recover the diluted effect size itself — it can only estimate a near-zero number more precisely.

This asymmetry ratio r is the central object of this paper. Its existence, and the prior prediction that r >> 1, derive from the mathematical structure of phase-transition windows in the dynamic programming recursion of ZFCρ (an integer complexity theory built on ZFC set theory; foundational framework in Qin, DOI: 10.5281/zenodo.18914682).

This paper's contributions were independently derived from ZFCρ's mathematical structure. They overlap with parts of the territory covered by the literatures named above, but the knowledge paths are different. This paper acknowledges these overlaps and positions its incremental contributions, but does not claim inheritance.

2. Definitions

Definition 1: Three-Phase Response Structure

Let the intervention response function g(z) depend on the subject's penetration depth z in the state space, with three segments:

g(z) = 0, for z < F (sprouting zone: microscopic perturbation present, but macroscopic net effect is zero or negative)

g(z) = δ · (z − F) / (E − F), for F ≤ z < E (flip to establishment: net effect ramps from zero to maximum)

g(z) = δ, for z ≥ E (post-establishment: full effect)

where F is the flip point, E is the establishment point, and δ is the true maximum effect size.

Definition 2: Asymmetry Ratio r

r = F / (E − F)

r measures the ratio of sprouting distance to establishment distance. r = 1 is symmetric; r >> 1 means the sprouting distance is much longer than the establishment distance.

Definition 3: Crossing Fraction π_cross

π_cross = P(Z_i ≥ F | treated)

The proportion of treatment-arm subjects who actually cross the flip point. When Z_i follows an exponential distribution Exp(μ) within the treatment arm, π_cross = exp(−F/μ). As r increases, F increases (for fixed total window width), and π_cross decreases.

3. Mathematical Source: The Ω Phase Transition in ZFCρ

3.1 Background

In the ZFCρ framework, the integer complexity function ρ_E(n) measures the minimum cost of representing integer n using two path types: addition (successor, n→n+1) and multiplication (factorization and recombination). When integers are stratified by Ω (number of prime factors with multiplicity), the dynamic programming recursion exhibits a phase transition in Ω-space from a disordered phase (successor paths dominant) to an ordered phase (multiplicative paths dominant). This transition is a gradual window, not a sharp boundary.

3.2 Data

Data come from three independent sources: mean h from presieve computation at N = 10^{10} scale, remaining step-level indicators from autocorrelation analysis (anticorr.c) at N = 10^7 scale — both released with ZFCρ Paper 44 (DOI: 10.5281/zenodo.19247859) — and the sixth indicator z/√j from ZFCρ Paper 57's spectral analysis (N = 10^{10}). All code and data available at the respective Zenodo pages.

mean h (local convexity): h > 0 means successor retains local competitiveness; h < 0 means multiplicative dominates locally. h = 0 marks full establishment of the ordered phase.

P(J>0) (multiplicative win rate): Proportion of integers where multiplicative net saving J > 0. P(J>0) = 50% is the minority-to-majority crossover.

E[A] (expected net step): Average net saving of multiplicative over successor. E[A] = 0 is the flip point where net effect turns positive.

w_shielded (Le Chatelier shielding rate): Proportion of integers whose fluctuations are absorbed by the system's frequency-intensity hedging. w_shielded = 2/3 is the natural threshold below which the exposed sector dominates and the Le Chatelier buffer fails.

Var(A) (step variance): Var(A) = 1.0 marks O(1) fluctuations, an interior characteristic of the ordered phase.

3.3 Five Crossover Points

Step-level crossovers are computed by linear interpolation. The spectral indicator z/√j is located via the approximate mapping Ω_typ ≈ ln j.

P(J>0) = 50%: Ω* = 2 + (50 − 28.7)/(57.1 − 28.7) = 2.750. Sprouting.

z/√j peak (spectral crossover): z/√j peaks at 5.03 at j = 23 (Paper 57 data), corresponding to Ω_typ = ln 23 ≈ 3.14. Multiplicative shielding begins to suppress additive-inertia accumulation. The fluctuation spectral peak precedes the net-step zero crossing — susceptibility before order parameter, consistent with standard phase-transition structure.

w_shielded = 2/3: Ω* = 3 + (70.5 − 66.7)/(70.5 − 65.1) = 3.710. Le Chatelier buffer failure.

E[A] = 0: Ω* = 3 + 0.326/(0.326 + 0.087) = 3.790. Flip (the most physically meaningful crossover).

h(Ω) = 0: Ω* = 4 + 0.004/(0.004 + 0.329) = 4.012. Establishment.

3.4 Four-Phase Structure

Phase 1: Sprouting (Ω ≈ 2.75). Multiplicative paths first win on a majority of integers, but net effect remains negative. Additive inertia begins accumulating.

Phase 2: Spectral flip (Ω ≈ 3.14). z/√j peaks. Additive inertia reaches maximum; multiplicative shielding begins but does not yet dominate. Fluctuation control shifts from correlation-dominated to screening-dominated. Occurs before step-level indicators flip.

Phase 3: Flip (Ω ≈ 3.79). E[A] = 0; net effect turns positive. Le Chatelier shielding nearly simultaneously drops below 2/3 (Ω ≈ 3.71).

Phase 4: Establishment (Ω ≈ 4.01). h = 0; successor path loses local competitiveness. Ordered phase fully established.

3.5 Extraction of the Asymmetry Ratio

Sprouting to flip: 3.79 − 2.75 = 1.04

Flip to establishment: 4.01 − 3.79 = 0.22

Asymmetry ratio: r = 1.04 / 0.22 ≈ 4.7 (rounded to ~5)

Total window width: 1.26. Flip-to-establishment is 17.5% of the window.

3.6 Status of the Ratio

r ≈ 5 is a numerical result within the ZFCρ model. The core argument depends only on the weaker condition r >> 1. Specific r values may differ across systems and must be independently estimated. ZFCρ's r ≈ 5 is a prior prediction whose cross-domain validity is an empirical question.

Structural intuition: Le Chatelier buffering operates at full strength during sprouting, requiring prolonged effort to reach the flip. Once the buffer is breached, establishment is rapid — the buffer itself has already dropped below the minimum required to maintain the old pattern.

4. Core Theorem: Signal Dilution

Let each subject i have latent state Z_i (penetration depth), jointly determined by random assignment and individual factors. Outcome Y_i = g(Z_i) + ε_i, ε_i ~ N(0, σ²).

Theorem (Signal Dilution). Under binary assignment combined with binary analysis, the ITT average treatment effect ATE = E[g(Z_i) | treated] − E[g(Z_i) | control]. When control-arm Z_i ≈ 0, ATE = E[g(Z_i) | treated]. Since only fraction π_cross of the treatment arm crosses F, the ATE is systematically diluted by sprouting-zone subjects contributing zero effect. Larger r means smaller π_cross means worse dilution.

Corollary 1 (Limitation of sample size). Increasing n improves ATE precision (lower SE) but does not change the diluted ATE itself. When π_cross is small, the diluted ATE may require impractical n for adequate power.

Corollary 2 (Best-responder signal). Conditional on Z_i ≥ F, subgroup mean effect approaches δ. This explains why post-hoc subgroup analyses show effects while overall trials are negative.

Corollary 3 ("Promising but insufficient evidence" pattern). In small studies, selection on Z inflates π_cross and yields effects near δ. In large RCTs, π_cross reverts to population base rate and the effect is diluted. This resembles publication bias but has a different mechanism.

Monte Carlo verification (Appendix B): At r = 5, depth = 0.3, n = 500/arm: binary power = 9.3% (true d = 0.8); TIZ power = 97.7%. A 1000-subject "well-powered" RCT is functionally blind to the true large effect.

5. Subject Conditions

The core theorem requires the following conditions to hold simultaneously:

Condition 1: The response function has threshold structure. Not all interventions have phase-transition responses. Purely linear causal relationships, one-time acute interventions, and surgical procedure-vs-no-procedure trials fall outside this paper's scope.

Condition 2: Realized exposure within the treatment arm is highly heterogeneous. If all subjects reach the same depth, no dilution occurs regardless of r. Sources of heterogeneity include adherence variation, physiological response differences, and implementation quality differences. This is nearly inevitable in complex interventions: metabolic therapy, psychotherapy, educational reform.

Condition 3: The study does not measure or exploit the continuous exposure state. If exposure-response analysis or adaptive titration is already being done, dilution is at least partially mitigated. The problem lies in the combination of binary assignment with binary analysis, not in binary randomization per se.

Condition 4: Outcomes are measured at the emergence layer, not the construct layer. The intervention acts on the foundational layer (construct); the outcome manifests at a higher level (emergence). If the outcome is measured directly at the construct layer (e.g., blood ketone levels rather than tumor size), the response may be continuous rather than threshold-type.

When all four conditions hold, the signal dilution theorem applies. Strongly applicable domains: metabolic interventions, rehabilitation, complex psychological interventions, educational reform, digital health and behavior change, just-in-time adaptive interventions.

6. Rays

Ray 1: Methodological Recommendations

6.1.1 From binary exposure to time-in-zone. The analysis variable should include cumulative time spent above the candidate flip point. Causal caveat: time-in-zone is a post-randomization variable; directly substituting it for treatment assignment sacrifices causal identifiability. Three legitimate uses:

(a) Mechanistic secondary analysis alongside the primary ITT.

(b) Adaptive titration design with exposure intensity randomized at the design level.

(c) Principal stratum or CACE framework, estimating causal effects for compliers. Requires monotonicity/exclusion restriction assumptions.

Additional note on immortal time bias: if time-in-zone requires being alive or highly adherent, TIZ correlates with unobserved positive confounders. Must be addressed via time-varying covariates in Cox models or landmark analysis.

6.1.2 Monitoring density must match flip-to-establishment distance. The flip-to-establishment window is only 17.5% of the total window. If monitoring intervals exceed this window, sprouting and post-flip states become indistinguishable. Monitoring frequency should be determined by the estimated flip-to-establishment distance, not by administrative convenience.

6.1.3 Exposure verification as prerequisite for negative conclusions. Before declaring an intervention ineffective, one must first establish: what fraction of subjects crossed the flip point, and for how long? Low or undetermined crossing fractions mean the negative conclusion speaks to implementation, not mechanism. Negative results after verified exposure attainment constitute genuine mechanism falsification.

6.1.4 Prior testing for phase-transition structure. Before large-scale RCTs, use small-sample intensive monitoring: track state-outcome joint trajectories, test for breakpoints via segmented regression or change-point detection, estimate r, and design the large trial accordingly.

Ray 2: Diagnostic Checklist

Suggestive signals (limited specificity; may also arise from publication bias or fidelity deficits):

(1) Small studies report strong effects; large RCTs report weak/null effects.

(2) Best-responder subgroup analyses consistently positive.

(3) Intervention intensity varies widely across studies; effect directions inconsistent.

(4) The field is chronically "promising but insufficient."

(5) Mechanistic research strongly supports efficacy; field trials fail to replicate.

Confirmatory signals (higher specificity):

(6) A measurable continuous state variable exists within the treatment arm, and the outcome shows a breakpoint/segmented relationship with it.

(7) Best responders overlap substantially with prospectively defined flip-point crossers.

(8) After controlling for fidelity and adherence, threshold/segmented models systematically outperform linear models.

Ray 3: Safety Direction

The asymmetry applies equally to toxicity. If a chronic exposure's toxicity response has r >> 1, Phase III may report "safe" because 81% of subjects never crossed the toxicity flip point. Real-world chronic exposure pushes more individuals past the threshold, producing Phase IV failure. The exposure verification recommendations apply to safety assessment as well as efficacy.

Ray 4: Worked Example — ERGO2

ERGO2 was an RCT in recurrent malignant glioma comparing ketogenic diet (KD) plus intermittent fasting plus re-irradiation versus standard diet plus re-irradiation. The primary endpoint (PFS) was not met; the trial is cited as evidence for "lack of clinical benefit."

Under this paper's framework: latent state Z = Glucose-Ketone Index (GKI); candidate flip point F = GKI ≤ 2.0. The intervention lasted only several days; blood ketone levels had high inter-individual variance; many patients did not spend sufficient time at very low GKI. But best responders (lower day-6 glucose, higher day-6 ketones) had significantly longer PFS and OS — precisely the pattern predicted by Corollary 2.

All five suggestive signals are met. Confirmatory signals are partially met or need testing (ERGO2 has GKI-continuous data but segmented regression has not been conducted; overlap between best responders and prospectively defined GKI ≤ 2.0 crossers needs verification).

Recommendations: (a) Secondary analysis using "cumulative days at GKI ≤ 2.0" as exposure variable. (b) Future RCT should randomize patients to different GKI targets rather than KD-vs-standard. (c) Only a trial with verified time-in-zone ≥ 4 weeks at GKI ≤ 2.0 that is still negative would constitute mechanism falsification.

7. Non-Trivial Predictions

Prediction 1: Best-responder effect sizes approximate the true effect in large negative RCTs

In any domain satisfying the subject conditions, within an overall-negative large RCT, the subgroup conditional on crossing the candidate flip point should show effect sizes approaching δ.

Falsification: In multiple preregistered studies, prospectively defined crossers show null effects.

Prediction 2: Time-in-zone models systematically outperform binary models

In trials satisfying the subject conditions, exposure-response analysis using the continuous state variable should systematically outperform binary ITT analysis (higher R², lower AIC/BIC, larger likelihood ratio).

Falsification: Prospectively measured TIZ models provide no additional explanatory power over binary models.

Prediction 3: The asymmetry ratio r > 1 in most construct-emergence systems

Independently estimated in systems with construct-emergence hierarchical structure, most should show r > 1 with median substantially above 1. ZFCρ's prior: r ≈ 5.

Falsification: Cross-domain data show r is not systematically skewed; r ≈ 1 or reverse-skewed is the norm. If so, this paper's scope should be sharply narrowed.

Prediction 4: Exposure-verified negative trials are the true falsifiers

Among large negative trials currently cited as evidence that "intervention X is ineffective for outcome Y," most treatment arms should have low π_cross. Only trials with verified exposure attainment that remain negative constitute strong mechanism falsification.

Falsification: A substantial fraction of subjects are prospectively verified to have crossed the flip point, yet the overall effect remains stably null.

8. Conclusion

This paper identifies and formalizes the asymmetry ratio r between distances on each side of the threshold as a key parameter affecting trial statistical power. This parameter has an independent mathematical source (the ZFCρ phase-transition window), yielding a prior prediction of r ≈ 5. Monte Carlo simulation demonstrates that at r = 5 with shallow penetration, a 1000-subject RCT achieves only 9.3% power for a large true effect (d = 0.8), while a time-in-zone design achieves 97.7%.

The four methodological recommendations (time-in-zone, monitoring density, exposure verification, prior testing) do not depend on r = 5 specifically, only on r >> 1. The framework applies to both efficacy and safety assessment.

Open questions: (1) What is the cross-domain distribution of r? (2) How should r be operationally estimated in education and psychotherapy? (3) How can this paper's diagnostic checklist be embedded in existing systematic review and meta-analysis methods?

Relationship to the SAE framework: Within SAE (Self-as-an-End; foundational paper: Qin, DOI: 10.5281/zenodo.18528813), the three-phase structure and asymmetry originate from the general geometry of construct-emergence relations. "Construct" refers to structural components at the foundational layer; "emergence" to macroscopic patterns arising from construct combinations at higher levels. Le Chatelier shielding determines the length of the sprouting phase; once the buffer is breached, emergence-layer response establishes rapidly. This asymmetry is an intrinsic feature of construct-emergence hierarchical structure — but whether it appears across domains is a prediction to be tested, not an established fact.

Full metabolic oncology application: SAE Biology Note 1 (forthcoming).

Appendix A: Worked Example Full Data (ERGO2)

See Section 6, Ray 4.

Appendix B: Monte Carlo Simulation

Design

Three-phase response function. True effect d = 0.8. Noise σ = 1.0. Window width 1.26. Penetration depths exponentially distributed. 2,000 simulations per condition. Binary design (t-test) vs TIZ (regression). Total subjects matched.

Result 1: r systematically destroys binary power

n = 200/arm, depth = 50%.

Result 2: Sample size has limited effect

r = 5, depth = 50%.

TIZ at n = 50 (56.8%) surpasses binary at n = 500 (51.2%).

Result 3: Penetration depth is decisive

r = 5, n = 200/arm.

Critical scenario

r = 5, depth = 0.3, n = 500/arm: Binary power 9.3%, TIZ 97.7%, observed d = 0.039 (true 0.8).

Code

See attached: monte_carlo_sim.py.