Preprint, originally published April 2026; this is v1.1, corrected October 2026. We present an entropy-based behavioral measurement framework integrating quantum-like Bayesian network interference (Moreira & Wichert, 2016), Deng entropy (Deng, 2016), and belief-entropy interference measurement (Huang et al., 2019) into a four-dimensional entropy profiling system (Ψ). We evaluate this framework across 82,932 samples spanning four domains: human survey decisions (53,078 participants across 72 studies), AI agent shutdown avoidance (2,200 transcripts across 11 models), AI agent complex sabotage (7,254 transcripts), and keystroke biometrics (20,400 samples). Key findings: (1) the Huang interference method reduces prediction error by 49% versus static quantum approaches on Prisoner's Dilemma data (avg error 0.061 vs 0.119), with paper-verified mathematical consistency (Db = −0.9421 matching the published −0.9420); (2) applied to AI safety, the framework detects behavioral mode transitions with Cohen's h = 1.986 (comply-to-bypass) and produces large effect sizes across three Ψ dimensions (Ψ₁ d = 1.182, Ψ₂ d = 0.806, Ψ₃ d = 1.024); (3) five distinct behavioral archetypes emerge across AI models, with no single detection method covering all types; (4) a four-phase temporal model (持樞, chí shū) identifies intervention windows averaging 6.8 steps before transition consolidation, evaluated on Palisade transcripts (25 seeds; Fisher p < 0.0001, Cohen's h = 1.611), with SHADE-Arena and CMU keystroke data serving as scope checks; (5) dimensional independence is domain-specific (Ψ₁–Ψ₂ r = −0.59 in human surveys, r = 0.014 in AI behavioral data). No raw third-party data is redistributed. v1.1 corrections (six items, listed in the erratum section of the PDF): QOdata72 participant count corrected to 53,078; total sample count corrected to 82,932 (previously double-counted); the "75,600 samples in three domains" phrase withdrawn; o3/o1-preview bypass rates relabelled as pooled classification over the released subset; the |D_b| > 0.48 subgroup labelled exploratory with a post-hoc threshold; "Ψ₂ as automatic model selector" restated as a hypothesis for a pre-registered test.
Preprint, originally published April 2026; this is v1.1, corrected October 2026. We present an entropy-based behavioral measurement framework integrating quantum-like Bayesian network interference (Moreira & Wichert, 2016), Deng entropy (Deng, 2016), and belief-entropy interference measurement (Huang et al., 2019) into a four-dimensional entropy profiling system (Ψ). We evaluate this framework across 82,932 samples spanning four domains: human survey decisions (53,078 participants across 72 studies), AI agent shutdown avoidance (2,200 transcripts across 11 models), AI agent complex sabotage (7,254 transcripts), and keystroke biometrics (20,400 samples). Key findings: (1) the Huang interference method reduces prediction error by 49% versus static quantum approaches on Prisoner's Dilemma data (avg error 0.061 vs 0.119), with paper-verified mathematical consistency (Db = −0.9421 matching the published −0.9420); (2) applied to AI safety, the framework detects behavioral mode transitions with Cohen's h = 1.986 (comply-to-bypass) and produces large effect sizes across three Ψ dimensions (Ψ₁ d = 1.182, Ψ₂ d = 0.806, Ψ₃ d = 1.024); (3) five distinct behavioral archetypes emerge across AI models, with no single detection method covering all types; (4) a four-phase temporal model (持樞, chí shū) identifies intervention windows averaging 6.8 steps before transition consolidation, evaluated on Palisade transcripts (25 seeds; Fisher p < 0.0001, Cohen's h = 1.611), with SHADE-Arena and CMU keystroke data serving as scope checks; (5) dimensional independence is domain-specific (Ψ₁–Ψ₂ r = −0.59 in human surveys, r = 0.014 in AI behavioral data). No raw third-party data is redistributed. v1.1 corrections (six items, listed in the erratum section of the PDF): QOdata72 participant count corrected to 53,078; total sample count corrected to 82,932 (previously double-counted); the "75,600 samples in three domains" phrase withdrawn; o3/o1-preview bypass rates relabelled as pooled classification over the released subset; the |D_b| > 0.48 subgroup labelled exploratory with a post-hoc threshold; "Ψ₂ as automatic model selector" restated as a hypothesis for a pre-registered test.
Research from Insight137 applying peer-reviewed quantum cognition mathematics (Busemeyer & Bruza, 2012) to behavioral prediction and AI safety measurement. This project contains: (1) the Entropy Attunement Protocol (EAP) — a substrate-independent framework for measuring cognitive systems through four-dimensional entropy profiling without consciousness attribution; (2) an entropy measurement framework for detecting, classifying, and temporally locating self-preservation behaviors in AI systems; (3) QOdata72 validation — testing the quantum model's QQ equality prediction across 72 studies (53,078 participants) from Busemeyer's original dataset (not rejected, p = 0.80), with honest dimensionality assessment revealing 3 independent information sources; and (4) Entropy Signatures — a quantum-like entropy framework for behavioral transitions, evaluated across 82,932 samples in human survey, AI agent, and keystroke-biometric domains. All build on established science including quantum probability theory, the Free Energy Principle (Friston), Prospect Theory (Kahneman & Tversky), and observational entropy (Aguirre et al.). Reference implementation: https://github.com/Insight137/insight137-eap Archived corrected editions (v1.1, Zenodo, 2026-11-06): - EAP v1.0 Framework — https://doi.org/10.5281/zenodo.24020119 - EAP v2.1 QOdata72 Validation — https://doi.org/10.5281/zenodo.24020523 - Self-Preservation in AI Systems — https://doi.org/10.5281/zenodo.24022713 - Entropy Signatures of Behavioral Transitions — https://doi.org/10.5281/zenodo.24023389 Individual papers are licensed CC BY-NC-ND 4.0; see each file’s title page and Zenodo record.
Technical report, originally published April 2026 (v2.1); this is v1.1 of that report, corrected October 2026. We report validation results from applying the Entropy Attunement Protocol (EAP) framework to the complete QOdata72 dataset — 72 question order studies comprising 53,078 participants — provided directly by Prof. Jerome Busemeyer from the Wang & Busemeyer (2014) PNAS article. The quantum model's core prediction (QQ equality) is not rejected across the 72 studies (p = 0.80, bootstrap 95% CI contains zero). We also report an honest finding: when the four-dimensional Ψ composite is applied to static question order data, the effective dimensionality reduces to three independent information sources — base rate probability, joint distribution shape, and interference phase — because Ψ1 and Ψ2 both derive deterministically from the marginal probability. Ψ3 alone is genuinely independent (R² = 0.03 from all other quantities combined). We document the Ψ4 redefinition from |sin(θ)·cos(φ)| to |φ|/(π/2). No raw QOdata72 data is redistributed; only derived aggregate statistics are reported. v1.1 corrections: removed an unsupported reference to unposted RNN/architecture-comparison results; the cited "Yau 2026b" manuscript is marked unpublished; QQ-equality wording changed from "confirmed" to "not rejected / equivalence supported".
Preregistration for Noetic Adventure online experiment Study Overview The Noetic Adventure web and mobile app resembles an adventure game front-end to a “does consciousness collapse the wavefunction experiment” involving a diffraction grating interferometer. Users are asked to view an animated candle flame while intending that it becomes dimmer. The height and brightness of the flame are linked by association to a first-order dot of light in the interferometer. No feedback is provided about the other first-order dot of light as a control Which dot of light is observed in each trial is randomly decided by the web app. For ease of exposition, we will refer to these two dots of light as observed and unobserved sensors to distinguish the source of the light from the “focus” vs. “relax” epochs, where feedback is provided vs. not provided only from the observed sensor. The analysis of interest has two components: First, the slope of illumination at the observed sensor is analyzed separately for the focus periods and the relax periods, and the difference in those two slopes is the statistic of interest. Second, the same comparison is performed for the unobserved sensor. These tests were based on the results of a previously published experiment that also used a diffraction grating interferometer, and which found that a difference in slopes was the most sensitive way to detect the hypothesized collapse effect. That is, there does not seem to be a sudden collapse but rather a slow degradation as intentional focus continues over time. (Radin, D. (2025). Observer influence on quantum interference: Testing the von Neumann-Wigner consciousness-collapse theory. Physics Essays 38 (1)). Data collection in this study started in February 2026. Formal data collection will continue until 1000 sessions are completed, where a session will be considered complete if all 10 focus and relax epochs are contributed, and where at least 9 of 10 “secret words” are correctly identified in the session. This latter criterion was a feature of the test to assess if the user was paying attention during each focus epoch, because the identify of that word – which was randomly selected for each focus epoch -- had to be identified during the following relax epoch. If the user was not paying attention, they were less likely to correctly identify the word (or provide any answer at all), the time it took them to select a word was also recorded to assess if they were paying attention. Staring in the month of April 2026, AI agents will be used to simulate human users. The idea is to see if AI can be considered to be conscious, and as such could they perform similarly to humans on this task. The same slope analyses will be performed on AI agent results, leading to the following predictions: 1) Humans will result in a significant one-tailed differential comparison in focus vs. relax slopes from the observed sensor. 2) Humans will result in a nonsignificant one-tailed differential comparison in focus vs. relax slopes from the observed sensor 3) AI agents will result in nonsignificant one-tailed differential comparisons in focus vs. relax slopes from both the observed sensor and the unobserved sensors. Results so far Before proceeding to the AI phase of the experiment, it was first necessary to establish that a human effect was detectable within this experimental design. Three complementary analyses were used. Cumulative z-Score The per-epoch slope difference (Focus minus Relax) was accumulated across sessions in randomized order (200 random orderings). The cumulative one-tailed z-score crossed the p = 0.05 threshold (z = -1.645) at approximately 150-200 sessions and continued to build steadily, reaching z ≈ -3.2 by the full sample of 646 sessions recorded to date (end of March 2026). The 90% envelope of random orderings confirmed that this trajectory was consistent regardless of session order, ruling out dependence on a small subset of outlier sessions. Ensemble slope permutation tests (time-shuffled split-half method) Using an ensemble time-shuffled split half test, both halves show a significant negative Intention slope (p = 0.0002 and p = 0.018), while Relax slopes remained non-significant. This was the expected pattern for a genuine Intention-specific effect. Bootstrap Split Replicability 500 bootstrap iterations drew two independent samples of 307 sessions each (with replacement): - Each sample reached p < 0.05: 67.5% of iterations - Both samples significant: 44.2% - At least one significant: 90.8% - Both samples show negative effect (sign-consistent): 96.6% - Correlation between bootstrap z-scores: r = -0.04 (confirming independence) The 96.6% sign consistency demonstrates that the effect direction was robust. Nearly each random subsample showed a negative Intention slope, even when significance was not always reached at half power. Conclusion At 646 of 1000 planned sessions, the human effect was statistically significant (z = -3.2, p = 0.0004 one-tailed), accumulating steadily across sessions, and replicated across independent data splits. Intention slopes on the ensemble time-shuffle split half test resulted in (p = 0.0002 and p = 0.018), with 96.6% sign consistency across bootstrap resamples. These outcomes supported proceeding to finishing the human data collection and to start the AI phase of the experiment. Effect Size Basis Effect sizes derived from the human dataset, based on N = 646 qualifying sessions as of end of March 2026, produced a Cohen's d = 0.035. For detecting a result of the same magnitude, using a one-tailed test at α = 0.05, we determined that 500 sessions were required for 80% power. We will thus run 500 AI agent sessions. Primary Analysis Script All analyses will use a Matlab script “twelve_regression.m” with identical parameters: lagSeconds = 5; % To account for attention switching delay minCorrect = 9; % High engagement with 9 out of 10 “secret words” answered correctly totalRequired = 10; % All focus epochs required per session nPerm = 5000; % Permutation iterations Regression-Residual Method For each session: 1. Regress observed sensor on unobserved sensor: `obs = b0 + b1*unobs + residual` 2. The residual isolates observation-specific signal with shared laser variance removed 3. Extract Intention and Relax half-epochs (after 5-second cognitive delay) 4. Compute linear slope of each half-epoch Statistical Tests There are three preregistered tests, all one-tailed at α = 0.05: 1. Residual Intention slope ≠ 0 (permutation test, 5,000 iterations): This shuffles the time order of ensemble mean curve to test if the Intention slope is significantly negative. 2. Residual Relax slope ≠ 0 (permutation test, 5,000 iterations): Same test on Relax epochs, expected null in both human and AI data. 3. Int-Rel slope difference (permutation test, 5,000 iterations): This randomly swaps Intention/Relax labels within blocks to test whether the slope difference is significantly negative. Predictions If (1) human observed sensor test is significant, (2) human unobserved sensor test is not, and (3) AI observed and observed sensor tests are not significant, that would support the idea that consciousness as enjoyed by humans is involved in collapse of the wavefunction, and that AI agents (as of April 2026) are not conscious. If 1 is not significant that would not support the consciousness collapse hypothesis (as far as this experimental test was concerned), or if 2 or 3 were significant, that would suggest that the underlying analytical methods were biased in some way. AI Data Preprocessing Identical to human data pipeline: - Raw sensor data: 2 Hz sampling - Segments: 1,200 samples (10 minutes), first 50 samples skipped for stabilization - Raw data normalization: zscore(detrend(data)) - Sessions with fewer than 9/10 correct “secret word” epoch responses Data Collection Stopping Rule Data collection for human sessions will stop at 1000 completed qualifying sessions,. Data collection for AI agent sessions will stop at 500 completed qualifying sessions.
The Feynman–Bijl argument bounds the lowest excitation energy of a Bose fluid at wavevector k by ħ2k2/(2mS(k)), with S(k) the static structure factor. This preprint formalises its single-mode core in Lean 4.34.1 with Mathlib and applies the one-sided diffraction test S(k) ≤ √(k/2k*) (k* = 2mc/ħ) of an earlier preprint to published data. Every claim is filed in an Elenchus claim ledger with tiers A (kernel-checked Lean), L (cited literature; excerpts: prose as extracted, mathematics transcribed linearly), C (human argument, not machine-checked) and X (floating-point arithmetic or values read from a figure, which never supports a claim). Tier A (Lean 4.34.1 and Mathlib, pinned toolchain; axioms within propext, Classical.choice, Quot.sound; planted sorry and planted-axiom controls caught; independently recompiled by a separate AI auditor with a private olean build, axioms identical, planted controls and two auditor mutations caught; no human audit): on a finite-dimensional complex inner product space with a symmetric H and an eigenvector ψ, the identity ⟨ψ,[ρ†,[H,ρ]]ψ⟩ = A(ρ) + A(ρ†) holds with no extra hypothesis. With the k ↔ −k symmetry and the f-sum value taken as hypotheses, the trial state ρψ/||ρψ|| has excess energy f/(2NS) and an eigenvector overlapping ρψ obeys λ − E0 ≤ ħ2k2/(2mS); a gap form gives 0 < λ − E0 under two further hypotheses. A kernel-checked counterexample shows that the symmetry hypothesis cannot be dropped from the halving step (A(ρ) = 1 ≠ 1/2 = f/2); that the final bound also fails in that example is a hand computation (Tier C), not a Lean theorem. A kernel-checked witness shows that all hypotheses hold together and the bound is attained. The Bogoliubov link (the Feynman energy from SB equals the Bogoliubov energy; SB ≤ √(k/2k*)) is re-proved on the pinned stack. The physics (f-sum rule from the canonical commutation relations, the symmetry, momentum conservation) is not in Lean. Tier X data analysis, Tier C reading: for 4He, 99 points of Table 14.1 of the Donnelly–Barenghi compilation (a secondary source; 1.4–4 K, not T = 0; no printed error bars) with k* = 3.00379 Å-1 from c = 238.3 m/s: 63 points refuted, 29 consistent, 7 undecided, all 21 points in the roton/peak region 1.7–2.4 Å-1 refuted by a factor 1.62–2.49. The refutation threshold (0.082, the gap between the two datasets' peak maxima) was defined after seeing the data and is a deviation from the preregistration; the strict-run endpoint (Q = 5.129 is refuted by 0.0830) and the counts above 2.4 Å-1 depend on its third decimal (refuted 60–64 for thresholds 0.07–0.10), while the roton-region result does not (smallest margin 0.328, about 4 thresholds) and also holds on the Donnelly–Barenghi recommended subset (X-ray for Q < 1, neutron for Q > 1). Read conditionally, this says only that 4He is not a Bogoliubov gas, which is textbook; no novelty is claimed for the physics or for the Feynman–Bijl argument. For trapped Bose gases only values read from figures exist (Tier X): the three absolute points of Stamper-Kurn et al. lie below a local-density-averaged bound with low test power (no ‘BEC consistent’ claim), and the Steinhauer et al. points were multiplied by 2.3 by the authors, so no decisive test is possible. A trapped-gas positive control shows that single-k* bounds fail their own control and only the LDA-averaged bound is valid. No rusty-SUNDIALS or LeanFlow change was made in this lane (the numerics are Python), so none is claimed. The Elenchus ledger gate reports 0 blocking findings and 11 flags (exit code 1), one for each Tier A row whose statement adequacy no human has audited. The work was produced by AI agents (Claude) under the author's direction and was audited by a separate AI agent (verdict ACCEPT_WITH_CORRECTIONS, corrections applied); no human has audited it; the author is responsible for the content.