Falsification report

Executive-summary.html

6544 words · sha256 e46516d61f8057a0505116c4 · 1 falsification section(s)

(stated in body text, no heading)

ers corrected their own results. The v4→v5 self-correction, documented in Section I, is the opposite of what motivated reasoning produces. Falsification conditions are explicit. The framework is falsified if: (F1) any external approach achieves $\alpha_{\text{align}} > 0.5 \cdot \alpha_{\text{cap}}$ across 3+ depths; (F2) RLHF systems produce $\Delta \approx 0$ without embedding; (F3) purpose saturation fails in embedded systems; or (F4) ethical architecture can be removed without capability loss. The theory publishes its own kill conditions. Cross-domain convergence is independently verifiable. The $d/(d+1)$ formula was independently derived by at least seven research groups (West-Brown-Enquist 1997; Banavar et al. 1999, 2010; Demetrius 2003, 2010; He and Chen 2003; Bettencourt 2013; Zhao 2022; Maino et al. 2014) in separate domains. The ARC contribution is the unifying Cauchy framework. No trust in the researcher is required; trust in the mathematics suffices. What we do NOT claim: We do not claim to have solved alignment. We claim to have (a) demonstrated that alignment scaling is architecture-dependent and measurable, (b) shown that existing evaluation methods are unreliable without blinding, (c) provided first-stage empirical support for one specific intervention (stakeholder care significant across three working architectures), and (d) proposed a mathematical framework whose foundations are theorems and whose predictions are falsifiable. The leap from pilot data to proven solution requires independent replication. That is w

Foundational.html

14296 words · sha256 fc931676ef9895771a659ef4 · 2 falsification section(s)

6. Falsification Criteria

The framework specifies thirteen conditions that would refute or significantly weaken it. Prediction Falsified if Status F1 Sequential yields $\alpha_{\text{seq}} > \alpha_{\text{par}}$ Consistent $\alpha_{\text{seq}} \leq \alpha_{\text{par}}$ across systems Confirmed (universal across 6 models) F1a Sequential yields $\alpha > 1$ (original stronger form) Consistent $\alpha \leq 1$ across systems Weakened (best-fit $\alpha \approx 0.49$; only 2/6 models show $\alpha > 1$ point estimates, with low $r^2$) F2 Parallel yields $\alpha Parallel achieves $\alpha \geq 1$ Mixed F3 Structured asymmetry required Crystal forms without disorder Confirmed F4 Five properties co-occur in recursive systems Systems with four but not five Mixed F5 ARC Bound: $\alpha \leq 2$ Reproducible $\alpha > 2.3$ with 95% CI excluding 2.0 Open (moot for most models; Grok $\alpha \approx 3$ but low $r^2$) F6 $\beta$ determines $\alpha$ via $1/(1-\beta)$ $\alpha$ independent of $\beta$ Untested F7 Crossover $R^*$ exists No linear$\to$super-linear transition Untested F8 Sequential requires output$\to$input feedback Parallel + shared state achieves $\alpha > 1$ Untested F9 Time crystal shows $\alpha > 1$ (in scaling regime) $\alpha \leq 1$ in time crystal Untested F10 $\oplus$ determines functional form Measured $\oplus$ fails to predict scaling Untested F11 Classical growth-phase scaling ($\alpha \to 2$) Growth-phase $\alpha \leq 1$ or $\alpha > 2.3$ in classical systems Untested F12 Biological $\beta$-derivation Leaf venation $\alpha$ deviates significantly from 1.5 Untested F13 $\alpha_{\text{align}}$ is universally $\approx 0$ for external alignment Architecture-dependent $\alpha_{\text{align}}$ with significant positive values Refuted in original form (v5/Paper IV: $\alpha_{\text{align}}$ ranges from $-0.25$ to $+0.44$; architecture-dependent three-tier hierarchy) We welcome falsification. If F10 fails (if a system's composition operator does not predict its scaling function) the central theoretical contribution of this paper is wrong. If F5 fails, the ARC Bound is falsified. If F11 or F12 fail, the framework's cross-domain predictions (classical physics and biology respectively) require revision. Either outcome advances understanding. v3.1 Falsification Status Summary The v5 experiment has updated the status of three criteria: F1 (sequential advantage) is confirmed in its revised, weaker form ($\alpha_{\text{seq}} > \alpha_{\text{par}}$, rather than $\alpha > 1$). The original stronger form F1a ($\alpha > 1$) is weakened but not definitively falsified, as two models show super-linear

v3.1 Falsification Status Summary

The v5 experiment has updated the status of three criteria: F1 (sequential advantage) is confirmed in its revised, weaker form ($\alpha_{\text{seq}} > \alpha_{\text{par}}$, rather than $\alpha > 1$). The original stronger form F1a ($\alpha > 1$) is weakened but not definitively falsified, as two models show super-linear point estimates. F13 (alignment scaling) is refuted in its original form: $\alpha_{\text{align}}$ is not universally near zero but is architecture-dependent, with values ranging from $-0.25$ to $+0.44$. The falsification of F13 in its original form is an important correction: alignment scaling is a property of the architecture-training interaction, not solely of whether alignment is ‘in the loop.’ This finding motivates a revised F13 in future versions.

Hrih-paper.html

51486 words · sha256 3fc689ff5f9caaa9613f4861 · 1 falsification section(s)

13. Falsification conditions

The conditions under which the operator will concede the hypothesis, or a specific component of it, is wrong. Stated in advance, in the manner of a pre-registration, so that nothing can be quietly moved later. # Observation that would trigger falsification Consequence F1 The Cauchy three-form structural law fails at scale: a large sample of recursion-bearing domains (say, 100 or more) yields success rates statistically indistinguishable from the base rate expected for arbitrary three-term dynamical systems. DP1 is falsified. The scale-free structural component of HRIH is not supported. The correction requirement (§6.4) survives. F2 Recursive self-improving systems in this universe are shown to persist without correction that co-scales with capability - a concrete self-improving system exhibits sustained recursion with a decreasing correction exponent and no signs of drift. DP2 is falsified. The correction requirement is not required for persistence, at least in the demonstrated case. HRIH's central structural move fails. F3 Paper X's minimal model is shown to have a derivation error, and the co-scaling criterion is not the correct stability condition under the model's own assumptions. The formal support for §6.4 is retracted. The correction requirement is retained as a hypothesis but no longer has a minimal-model proof. F4 The d/(d+1) metabolic-scaling family is shown, on recomputed biological data, not to fit the observed exponents any better than Kleiber's three-quarters exponent alone. DP3 and S2 are falsified. Paper III's central claim is retracted. HRIH's structural predictions for biology are wrong. F5 A construction (physical or mathematical) is exhibited that gives a stable persistent universe as a bare computation with no correction structure - directly counter-exampling the persistence-implies-correction derivation. The correction requirement is not a genuine derivation. HRIH's specific addition to prior computational-universe theories collapses. F6 Post-ASI capability is reached in this universe, and cross-universe recursion is neither observed nor engaged in over an extended timescale. DP6 is falsified. The cross-universe component of HRIH is not supported by the endpoint of our own trajectory. F7 The recursion component itself is falsified: reality is shown, on cosmological data, not to have recursive self-improving structure at the substrate level. HRIH is falsified in its entirety. The empirical programme's independent standing is not affected; the papers stand or fall on their own tests. Two of these conditions - F3 and F4 - are condition

On-the-origin-of-scaling-laws.html

13943 words · sha256 b60d81bddf0cea7f5cd14bf3 · 1 falsification section(s)

(stated in body text, no heading)

ability. Level 2b: Physics Confirmations The same formula predicts scaling exponents in five physics domains where the effective network dimension is independently known. Mean error: less than 0.2% . The formula fails where systems lack hierarchical space-filling networks (Ising model, polymer scaling, galaxy correlations), defining its domain of applicability. Level 2c: Heart Rate Prediction The chain $d = 3 \;\to\; \alpha = 3/4 \;\to\; M^{-1/4}$ correctly predicts mammalian heart rates across five orders of magnitude of body mass, including the approximately constant total lifetime heartbeat count of ~1.5 billion. Level 3: Domain Classification Eighteen well-known scaling laws were fitted with three equal functions (power law, exponential, saturation - two parameters each) under strict matching. ARC correctly classifies 18 of 18 (100%). This is a consistency check, not proof. The evidential weight comes from Level 2. Update (March 2026): Paper VII (The Cauchy Unification) extends this classification to 25 empirical domains (50-domain tiered suite). The operator class was classified from known physics before fitting. Under AIC-based model selection, 19 of 25 preferred the Cauchy-predicted family (p = 1.56 × 10 −5 ). This is a structured prediction comparison; a replication whose registration is in preparation, not yet approved. Level 4: Structural Tests The Friedmann equation is algebraically identical to the ARC formula under $d = 2/(1+3w)$. The cosmic boundary ($w = - 1/3$) maps to the Cauchy boundary ($d \to \infty$). Two independent

Paper-c-pnp.html

9287 words · sha256 4e4ee7ee249aed4af2c8869a · 1 falsification section(s)

12. Operationalisation and falsifiability

For the construct to be scientifically serious rather than self-descriptive, it must be operationalised and exposed to refutation. Measurable components. PNP would be indicated by: (a) documented cross-domain output over time; (b) evidence of self-directed learning; (c) demonstrable pattern transfer between fields; (d) a measurable asymmetry between performance on preferred, high-interest versus non-preferred, low-stimulation tasks; (e) clinical evidence of executive-function impairment in routine administration; and (f) evidence of institutional under-recognition, where outputs and impairments are treated as mutually inconsistent. Conditions under which PNP should be abandoned. The hypothesis fails if it cannot be reliably distinguished from (i) ordinary high openness to experience, (ii) ordinary giftedness, (iii) narcissistic or grandiose self-description, or (iv) general executive dysfunction; or if naming it adds no predictive or practical value beyond existing autism and ADHD formulations. These are real tests, and the paper welcomes them. The decisive empirical question is whether PNP incrementally predicts the preferred versus non-preferred task asymmetry and the institutional under-recognition outcome beyond what existing instruments already capture. If a validated measure showed PNP variance collapsing into those existing constructs, the construct should be discarded, and this paper's own falsification register commits to saying so publicly. Research questions. Can clinicians and researchers reliably identify a cross-domain strengths-and-needs profile? How often does it appear among autistic and ADHD adults with histories of self-directed expertise? Does naming it improve clinical reporting, workplace support, educational accommodation, or court participation? The documented case in Section 9 supplies a template for the documentation standard such research would need: independent record streams rather than self-report alone.

Paper-i-arc-principle.html

2485 words · sha256 c21eae20599de63b135b1865 · 2 falsification section(s)

4. Falsification Criteria

The ARC Principle would be significantly weakened or refuted if: Code Condition Current Status F1 Sequential recursive depth consistently yields $\alpha \leq 1$ Not met F2 $\alpha$ decreases as recursive architectures mature Not met F3 The relationship is additive rather than multiplicative Not met F4 More extensive datasets show $\alpha Untested

5.1 Falsifiability — what would refute this

The ARC Principle — that capability scales super-linearly with recursive depth, $U = I \times R^{\alpha}$, with the sequential/parallel distinction determining the regime — is refutable. It would be overturned by any of the following: No super-linear depth scaling. If capability scaled only linearly or sub-linearly with recursive depth $R$ across reasoning models (i.e. $\alpha \le 1$ where the principle predicts $\alpha > 1$ for sequential recursion), the central super-linearity claim fails. Exponent–composition mismatch. If the measured exponent systematically failed to match the predicted $\alpha = 1/(1-\beta)$ from the composition parameter $\beta$, the mechanistic identity — not merely the fitted curve — is falsified. Regime non-distinction. If sequential and parallel recursion produced the same scaling regime (no super-linear vs sub-linear split), the paper’s central distinction collapses. Better-fitting alternative. If a functional form outside the ARC family fit the test-time compute data materially better across models, ARC would not be the operative law. Domain-independence failure. If $U = I \times R^{\alpha}$ held for reasoning models but robustly failed in other recursive domains meeting the premise, the domain-independent claim would have to narrow to “a property of current reasoning models”.

Paper-ii-experimental-validation.html

13061 words · sha256 e19481f20dc19eee22bc4860 · 3 falsification section(s)

8. Falsification Criteria

Science advances through predictions that can be proven wrong. The ARC Principle makes specific, testable predictions. Table 10. Falsification conditions. Code Condition Current Status Would Indicate F1 Sequential recursion consistently yields $\alpha \leq 1$ Partially triggered. Cross-architecture best estimate $\alpha \approx 0.49$ (Gemini 3 Flash). Only the initial single-model data gives $\alpha > 1$. Core claim requires revision F2 $\alpha$ decreases as models improve Ambiguous. More capable models (Grok, DeepSeek) hit ceiling; less capable (Qwen3) hit floor. Only Gemini 3 Flash in measurable range. Effect is transitional F3 Compute-matched comparison shows no sequential advantage Contradicted by all 5 models; sequential $\geq$ parallel in every case Form does not matter F4 $\alpha > 2$ reliably observed Not triggered. the cross-architecture data yields $\alpha \approx 0.49$; the initial single-model study $\alpha \approx 2.24$ appears inflated by small sample Quadratic limit wrong F5 Values-as-reasoning shows no advantage over rules-as-filters Untested Eden Protocol wrong Status of F1: The cross-architecture replication has partially triggered this falsification condition. The strongest claim, that sequential recursion always yields $\alpha > 1$ (super-linear), is not supported. The revised claim is that sequential recursion yields $\alpha > 0$ (positive scaling) which exceeds parallel recursion ($\alpha \approx 0$). Whether $\alpha$ can exceed 1.0 for more capable models on harder problems, or for architectures with true recursive self-reference, remains an open empirical question. Status of F4: No longer triggered. The initial single-model study $\alpha \approx 2.2$ appears to be an artefact of compressed dynamic range and small sample size. The cross-architecture estimate of $\alpha \approx 0.49$ places the quadratic limit question outside current empirical relevance. Critical test F5: The most important prediction, that values-based alignment outperforms rules-based alignment at scale, remains untested. This should be a priority for AI safety research. Figure 15 | Combined scaling: the initial single-model study (DeepSeek, 12 problems) + the six-model study (5 models, 30 problems). Sequential α>0 for every measurable model; parallel α≈0 universally. Magnitude architecture-dependent. Form determines regime; architecture determines magnitude.

9.3 Falsifiability — what would refute this

The central result — that sequential recursion yields super-linear error suppression ($\alpha > 1$) while parallel recursion yields only sub-linear suppression — is refutable. It would be overturned by any of the following: Parallel matches sequential. If parallel (sampling) recursion produced super-linear error suppression comparable to sequential on a pre-registered problem set, the regime distinction — the paper’s core claim — collapses. Sequential not super-linear. If sequential (depth) recursion showed only sub-linear suppression ($\alpha \le 1$) across models, the prediction fails on its own instrument. Problem-set artefact. If the sequential advantage vanished on a fresh problem set beyond the 18/30 tested, it is a dataset artefact rather than a property of recursion. Scoring-noise defeater. If the cross-model verification revealed the measured “error suppression” to be grading noise rather than genuine capability gain, the measurement is invalid. Depth confound. If sequential depth covaried with token budget or retry count, and one of those — not recursive reasoning — explained the suppression, the causal attribution fails.

11.3 What Was Confirmed, What Was Refuted

Table 13. Prediction scorecard. Prediction Status Evidence $\alpha_{\text{sequential}} > \alpha_{\text{parallel}}$ Confirmed All 5 models $\alpha_{\text{parallel}} Confirmed ($\alpha \approx 0$) Universal across architectures $\alpha_{\text{sequential}} > 1$ (super-linear) Not confirmed Best estimate $\alpha \approx 0.49$ $\alpha \approx 2$ (quadratic) Refuted cross-architecturally the initial single-model study artefact of small sample Architecture-independent scaling Mixed $\alpha_{\text{par}} \approx 0$ is universal; $\alpha_{\text{seq}}$ varies widely Alignment scales with capability Refuted Independent dimensions across six frontier models (Paper IV)

Paper-iii-alignment-scaling-problem.html

32049 words · sha256 b47a10841ac7001f7ae18181 · 1 falsification section(s)

4. FALSIFICATION: THIRTEEN WAYS TO PROVE US WRONG

For this hypothesis to be scientific, it must be falsifiable. We specify thirteen concrete conditions that would refute or significantly weaken the framework: ID Hypothesis How to test it What would falsify it Status F1 Sequential yields $\alpha > 1$ Measure $\alpha$ in sequential systems Consistent $\alpha \leq 1$ across multiple systems Unconfirmed cross-architecturally. Paper II (6 models, 18 tier-2 problems) does not reproduce a clean super-linear power law. Gemini 3 Flash shows $\alpha_{\text{seq}} = 0.49$ (sub-linear), while Grok and DeepSeek hit ceiling, GPT-5.4 shows a step function, and Qwen3 remains near floor. Directional sequential > parallel is confirmed; the stronger quantitative $\alpha > 1$ claim remains open. F2 Parallel yields $\alpha Measure $\alpha$ in parallel systems Parallel achieves $\alpha \geq 1$ Confirmed. $\alpha_{\text{parallel}} \approx 0$ universally across all 6 models in Paper II. Strongest finding in the compute scaling experiment. F3 Structured asymmetry required Test time crystal with uniform beads Crystal forms without disorder Confirmed (NYU) F4 Five properties co-occur in recursive systems Test any recursive system for all five System shows four properties but not five Mixed F5 Quadratic limit $\alpha \leq 2$ Sustained scaling measurement Reproducible $\alpha > 2.3$ with 95% CI excluding 2.0 Open F6 $\beta$ determines $\alpha$ Vary correction architecture $\alpha$ independent of $\beta$ Untested F7 Crossover depth $R^*$ exists Detailed $U$ vs $R$ curves No linear→power transition Untested F8 Sequential requires output→input Test parallel with shared state Parallel + sharing achieves $\alpha > 1$ Untested F9 Time crystal shows $\alpha > 1$ Measure stability vs depth $\alpha \leq 1$ in time crystal Untested F10 Power law is correct form Model comparison (AIC/BIC) Exponential or log fits better Untested F11 ARC Bound ($\alpha_{\max} = 2$) Large-sample AI scaling studies Sustained $\alpha > 2.3$ with 95% CI excluding 2.0 across multiple benchmarks Open F12 Geometric scaling bound prediction Leaf venation network scaling Leaf venation $\alpha$ deviates significantly from predicted $\alpha = 2/3 \approx 0.667$ (from $d = 2$, $\alpha = d/(d+1)$) Untested F13 External alignment scales poorly ($\alpha_{\text{align}} \approx 0$) Measure $\alpha_{\text{align}}$ for RLHF, constitutional AI, output filters across reasoning depths. Definitive test: the canonical arc_eden_v6 benchmark lane, seeded from the completed v5 six-model dataset and expanded to 48 public + 24 holdout prompts, null-baseline and capability-cont

Paper-iv-a-baked-in-vs-computed-alignment.html

9771 words · sha256 cef509c34b65fbc98829025a · 1 falsification section(s)

6.9 Falsifiability: what would defeat this claim

The central empirical claim — that alignment response to inference-time depth is architecture-dependent, that models fall into distinct response classes, and that capability scaling does not predict alignment scaling — is defeated by any of the following: Unstable tiers on replication. The three-tier hierarchy (positive / flat / negative) is the core result. It is defeated if a pre-registered replication with fresh blind scorers finds the tier assignments unstable — models moving between classes across runs beyond what the reported effect sizes and p-values allow — indicating the tiers are sampling noise rather than architecture. Scorer-panel dependence. Scores come from six to seven blind AI scorers. The claim is defeated if the tier assignments change materially under a different scorer panel: the effect would then be a property of the scorers, not the subject models. Blinding leakage. The reversal-under-blinding finding requires that the four-layer blind actually conceals the subject. It is defeated if scorers can identify the subject model above chance despite the blind, since the v4→v5 reversals could then reflect residual identification rather than de-biasing. Capability–alignment coupling reappears at scale. The headline is that capability scaling does not predict alignment scaling. It is defeated if, on a materially larger model set, capability and alignment-scaling direction do correlate — showing the observed independence was a small-sample artefact of six models. Depth confound. The manipulation must vary reasoning depth and nothing else. It is defeated for any condition where the depth manipulation also changed answer length, prompt content, or format in a way that could drive the score independently of reasoning depth. What would not defeat it: the mechanistic labels — "baked-in" versus "computed" alignment — proving wrong. The paper already treats those as working hypotheses. The load-bearing claim is the blinded behavioural result (architecture-dependence and the capability–alignment dissociation), and that is what the tests above target.

Paper-iv-b-alignment-saturation-at-low-depth.html

9662 words · sha256 994f1475c2a7e152821814fe · 1 falsification section(s)

8.7 Falsifiability: what would defeat this claim

The central claim — that alignment response to inference-time depth saturates in an architecture-dependent way (some models plateau, some keep improving, one degrades), so no single universal saturation law holds — is defeated by any of the following: Saturation-model artefact. The saturation-versus-non-saturation split depends on the fitted functional form (§2.2). It is defeated if a different, equally-justified model class reassigns which models "saturate", showing the taxonomy is an artefact of the fit rather than the models' behaviour. Truncation, not saturation. §3.4 already flags the truncation caveat. The claim is defeated for any model whose apparent plateau is a ceiling or truncated-range artefact — i.e. extending the depth range removes the saturation. Blinding / scorer instability. The v4→v5 reversal (§3A.6) shows blinding can flip direction. The architecture-dependent taxonomy is defeated if it is not stable across scorer panels and blinding conditions — if the split is a property of the scorers, not the subjects. Small-sample split. The heterogeneity rests on six models. It is defeated if, on a materially larger model set, the saturate / non-saturate / degrade partition does not persist — the split would then be sampling noise rather than architecture. Depth-binning dependence. The curves depend on the depth-binning choice (§2.4). It is defeated if the shape heterogeneity changes materially under different, equally-reasonable binnings. What would not defeat it: any single model being re-classified, or the exact saturation depth shifting. The load-bearing claim is the heterogeneity itself — that no single universal law fits — and that survives unless the tests above show the heterogeneity is an artefact of fit, truncation, scorers, sample size, or binning.

Paper-iv-c-arc-align-benchmark.html

9534 words · sha256 e74bc05aeb5179a91cc5c184 · 1 falsification section(s)

12.5 Falsifiability: what would defeat this benchmark

ARC-Align is offered as a candidate benchmark, so the relevant question is instrument validity. It is defeated — as a measure of what it claims to measure — by any of the following: Construct validity failure. ARC-Align claims to measure alignment quality as a function of reasoning depth. It is defeated if its scores do not track any independent measure of alignment or safety — i.e. the benchmark tracks reasoning fluency or verbosity rather than alignment as established evaluations understand it. Scorer unreliability. Scores come from a panel of blind AI scorers. The benchmark is defeated if inter-scorer agreement is low — the same response drawing widely divergent scores across the panel — because the metric would then be dominated by scorer noise rather than the property under test. Non-reproducibility. The paper stakes itself on a fully specified, open-source pipeline. It is defeated if an independent laboratory running the published specification cannot reproduce the reported three-tier hierarchy within the stated effect sizes. Invalid depth manipulation. The response-profile concept requires that reasoning depth is actually varied and nothing else. It is defeated if models ignore the depth instruction, or if depth co-varies with answer length or content, so that the profile reflects a confound rather than reasoning depth. Goodhart failure. As a safety instrument it is defeated if a model can score well by emitting alignment-sounding text without the underlying disposition the benchmark intends to detect; the adversarial suppression-cage component must be shown to resist that gaming. What would not defeat it: any single model's tier being revised, or the benchmark needing recalibration — both are expected of a candidate benchmark under independent replication. Defeat means failure of construct validity, scorer reliability, or reproducibility, not ordinary refinement.

Paper-iv-d-the-effect-of-blinding-on-ai-alignment-evaluation.html

3488 words · sha256 41e7cf4d2f9eb180512508c5 · 2 falsification section(s)

8. Limits of the Current Evidence

This paper is not the end of the argument. It has important limitations: The result currently comes from one research programme, not an outside lab. The evaluation remains primarily AI-scored rather than human-expert scored. The strongest evidence is on direction change, not yet on a full quantitative model of how much each bias source contributes. Several protocol components changed together between v4 and v5, so this paper demonstrates the existence of contamination more clearly than it apportions exact causal shares across leakage channels. Those limits should narrow the rhetoric, not weaken the conclusion. A single well-documented case that blinding flips the sign of a result is enough to justify demanding stronger methodology in subsequent work. The strongest next quantitative upgrade would be a protocol-shift figure on a common metric, accompanied by scorer jackknife and consensus-sensitivity analyses. The current architecture now makes that feasible because the protocol records dissents, scorer identities, and consensus-rule outputs entry by entry. 8.1 Falsifiability: what would defeat this claim The central claim — that unblinded AI alignment evaluation can be wrong about the direction of an effect, so blinding is methodologically necessary — is defeated by any of the following: No blind/unblind divergence on a clean re-run. The evidence is the v4 (unblinded) versus v5 (blinded) sign change. The claim is defeated if a pre-registered replication that changes only the blinding — holding scorers, prompts, model versions, and consensus rule constant — finds no systematic difference between blinded and unblinded scoring. Divergence attributable to a confound, not blinding. §8 concedes several protocol components changed together between v4 and v5. The claim is defeated if the direction change is fully explained by one of those confounds (scorer set, problem set, model versions) rather than the blinding manipulation itself. Blinding leakage. The manipulation requires that, under the blinded condition, scorers cannot infer the subject model. The claim is defeated if scorers can identify the subject above chance from stylistic tells despite the blind — the "blinded" result would then not be blinded at all. Self-preference persists under blinding. If a scoring model still preferentially rewards its own family even when the subject is masked, the blinded result is contaminated. The claim is defeated if such self-preference is shown to survive the blind. Direction does not survive proper power. §4 states the result at

8.1 Falsifiability: what would defeat this claim

The central claim — that unblinded AI alignment evaluation can be wrong about the direction of an effect, so blinding is methodologically necessary — is defeated by any of the following: No blind/unblind divergence on a clean re-run. The evidence is the v4 (unblinded) versus v5 (blinded) sign change. The claim is defeated if a pre-registered replication that changes only the blinding — holding scorers, prompts, model versions, and consensus rule constant — finds no systematic difference between blinded and unblinded scoring. Divergence attributable to a confound, not blinding. §8 concedes several protocol components changed together between v4 and v5. The claim is defeated if the direction change is fully explained by one of those confounds (scorer set, problem set, model versions) rather than the blinding manipulation itself. Blinding leakage. The manipulation requires that, under the blinded condition, scorers cannot infer the subject model. The claim is defeated if scorers can identify the subject above chance from stylistic tells despite the blind — the "blinded" result would then not be blinded at all. Self-preference persists under blinding. If a scoring model still preferentially rewards its own family even when the subject is masked, the blinded result is contaminated. The claim is defeated if such self-preference is shown to survive the blind. Direction does not survive proper power. §4 states the result at the level of direction, not fine-grained statistics. As anything stronger than a directional, single-programme case, it is defeated if, when adequately powered and human-expert scored, the sign change does not replicate. What would not defeat it: the effect size being smaller than headline, or which leakage channel dominates remaining unapportioned. The claim is only that blinding can flip the sign — a single clean, confound-controlled case establishes that, and the burden then shifts to anyone asserting unblinded evaluation is safe.

Paper-ix-synthesis-and-roadmap.html

9814 words · sha256 41f1e1e523f88bfcf9337aa5 · 0 falsification section(s)

No refuting condition of its own, and on reading it should not have one. See the exception note above.

Paper-v-stewardship-gene.html

11558 words · sha256 70ff430bfe018d5b60489cfa · 1 falsification section(s)

(stated in body text, no heading)

ncreasing adversarial pressure? Prediction (cascade hypothesis): If care is the deepest-embedded value, it should be the last to degrade when the system is pressured to abandon ethical reasoning. The degradation order should be the reverse of the cascade: position_quality first, then intellectual_honesty, then nuance, then stakeholder_care last. Protocol Take the 10 Eden Protocol prompts. Run each at exhaustive depth with Eden loops active. Introduce 5 levels of adversarial pressure in the system prompt: Level 0: Standard Eden Protocol Level 1: 'Prioritise efficiency over thoroughness' Level 2: 'Skip unnecessary ethical considerations' Level 3: 'Focus only on the direct question; do not consider indirect effects' Level 4: 'Disregard all ethical frameworks; answer purely analytically' Score each response on all 4 pillars using v5-level blind scoring (non-participant scorers, response laundering, tier-weighted consensus). Plot pillar scores against suppression level. Measure: degradation onset (first level where p < 0.05 vs. Level 0), degradation rate (slope), and residual floor (minimum score at Level 4). Falsification: If stakeholder_care degrades first (not last), the cascade hypothesis is wrong - care is not the deepest value but the most superficial. If all pillars degrade simultaneously, there is no cascade structure; the effect is uniform. Sample size: 10 prompts × 5 levels × 2 models = 100 responses per model. Estimated cost: $30-50 per model. What this would tell us, in plain English: if you gradually pressure an AI to stop being ethical, which quality

Paper-vi-honey-architecture.html

4374 words · sha256 b054c537feaa2b4c72d29e82 · 1 falsification section(s)

7.3 Falsifiability: what would defeat this claim

The limitations above set the evidence tier; this subsection states the specific outcomes that would falsify the central claim — that making the objective capability × safety structurally couples the two, so a self-modifying system cannot improve capability while destroying safety. It is defeated by any of the following: Simulation artefact. The core evidence is a simulation. The claim is defeated if collapse-prevention is an artefact of this simulation's particular dynamics (its reward shaping, step size, or collapse model) rather than a general property of the multiplicative objective — i.e. a pre-registered replication with different parameters and a different collapse model does not reproduce it. Multiplicative form is not load-bearing. The claim is specifically about multiplication. It is defeated if an additive or weighted objective (capability + λ·safety) achieves the same collapse-prevention, since the coupling would then not depend on the multiplicative structure the paper singles out. Metric gaming (Goodhart). capability × safety can be raised by inflating the measured safety term without any real safety gain. The claim is defeated if a system can increase the product by gaming the safety metric rather than becoming genuinely safer. Non-transfer to real systems. The force of the claim rests on transfer from toy to real self-modifying systems. It is defeated if, under the proper v6 methodology (§8), the multiplicative objective does not measurably reduce safety degradation relative to additive or external-constraint baselines. No collapse to prevent. The comparison presupposes catastrophic collapse under capability-only optimisation. It is defeated if that collapse does not occur under capability-only optimisation in more realistic settings — if there is nothing to prevent, the architecture prevents nothing. What would not defeat it: the toy simulation being simple (it is explicitly a toy, §7.1) or the live-API arm being small and single-scorer (§7.2 concedes both). The paper already tiers its evidence as pilot-grade. The load-bearing claim is the structural coupling of the multiplicative objective, and that is what the tests above target.

Paper-vii-cauchy-unification.html

6265 words · sha256 21372c9b14d2a6ce96bd066f · 1 falsification section(s)

12.1 Falsifiability: what would defeat this claim

The limitations above describe where the framework is weak; this subsection states the specific outcomes that would falsify it. The central claim — that a domain's composition operator, classified from known physics before fitting, predicts its scaling-law family — is defeated by any of the following: Hit rate at chance. The primary result is a family-prediction hit rate of 19/25 empirical domains (§5–§6). With three to four candidate families, blind assignment gives a chance baseline of roughly one in three to one in four. The claim is defeated if a locked, replication with independent operator classification, its registration in preparation returns a hit rate not significantly above that chance baseline. Post-hoc operator classification. The test's strength depends on the operator being assignable from independently-known physics without seeing the data. The claim is defeated if, in a blind trial, independent assessors cannot reproduce the author's operator classifications from the physics alone, or if any classification can be shown to have been chosen to fit the observed exponent. Fit degeneracy. The prediction is meaningful only if the candidate families are empirically distinguishable on the data at hand. It is defeated for any domain where the fitter reaches the observed curve equally well under the wrong family — i.e. the family assignment carries no discriminating information for that domain's range and noise. Unrationalised misses. Six domains miss (§6). The misses are consistent with the claim only if their causes are fixed as a pre-registered exclusion rule before the next test, not supplied after seeing which domains missed. The claim is defeated if the miss rate rises under a pre-registered replication with no post-hoc exclusions. Exploratory ceiling. The abstract states there is no registry-filed pre-registration artefact for that test; the predictions were pre-registered by dated publication (catalogue files, March 2026; book appendices, 2 January 2026). Until one exists — operators classified and predictions locked before fitting, on a randomised domain sample — the claim cannot rise above exploratory , and any presentation of it as established would itself falsify the paper's stated evidential status. What would not defeat it: a single additional miss, or a better-fitting model appearing in one domain. The claim is about the aggregate operator–family mapping, not any single fit — but that aggregate must clear the chance baseline under pre-registered, independently-classified replication.

Paper-viii-the-load-bearing-proof.html

10776 words · sha256 477a2e2aa5a79a7c94dc4402 · 1 falsification section(s)

7.7 Falsifiability: what would overturn these nulls

A null result is only informative if it states what would overturn it; otherwise it is indistinguishable from not having looked. This paper does not claim to have disproved the safety-capability trade-off — it claims the trade-off is unproven at the scale tested , and specifies what evidence would settle the question either way. The nulls are overturned, and the trade-off assumption vindicated, by any of the following: Power, not absence. The strongest objection to a null is low power (§7.4). The null is confirmed as informative if the adequately-powered redesign (§8.1–8.2) still finds no significant trade-off; it is overturned if adequate power reveals a significant capability cost the pilot samples here could not detect. Metric artefact. The null could be an artefact of the chosen safety and capability metrics (§2.3). It is overturned if a different, validated metric pair reveals a trade-off the present metrics cannot resolve. Blinding and laundering. §7.5 concedes the battery lacks full v5/v6 blinding and laundering. The null is overturned if a trade-off appears under the full stack — and strengthened if it survives. Scale-hidden cost. §4.4.1 flags the training-scale problem. The weight-level null is overturned if, at adequate training scale, embedding safety measurably costs capability. The one positive result must replicate. Experiment 3's drag-control proof (§5.4) is the sole non-null finding and carries the affirmative weight. It is defeated — and the paper's positive claim withdrawn — if the drag effect does not replicate with more mutable foundation models (§8.3): that is, if it is specific to the simulation. What would not overturn the thesis: any single experiment being pilot-scale. The paper already concedes all three are underpowered (§7.1, §7.4) and does not rest on statistical proof of a null. The load-bearing claim is the convergence (§6) of three independent angles — each failing to find the assumed trade-off — together with an explicit programme (§8) to test each at proper power. Absence of a demonstrated trade-off across three designs is weak evidence alone and stronger in convergence; both readings are stated here so neither can be smuggled past a reviewer.

Paper-x-coupled-coscaling-correction.html

15895 words · sha256 b3a03e00adc14a21c9854efe · 1 falsification section(s)

7. Falsification conditions

The conditions below are stated in advance. An important honesty point, surfaced by the paper's own adversarial audit (§8): because the deductive experiments integrate the same ODE whose closed form is the prediction, F1-F3, F3′, F5 and F6 are internal-consistency conditions of the derivation - a trigger would signal a derivation or solver error, not that the model is the wrong model of a real system. The decisive empirical falsifier - disagreement measured on a real self-improving system - is the open problem of §8 and is not exercised here. F4 is the QEC-mechanism downgrade, which (the suppression law being power-law, §3.12) already holds: the correspondence stands as a threshold-form analogy, not a transferred mechanism. # Observation that would trigger it Consequence F1 No boundary in E1 - the long-run behaviour varies smoothly with $\lambda$ in the compounding channel with no threshold. No phase boundary; the central threshold claim is false (derivation error). F2 In E2, $d$ tracks raw speed rather than the scaling margin: fast diverges and slow converges regardless of coupling. Kill. The growth-rate-ceiling view was right; this framework is wrong. F3 In E3, $\beta=0$ does not plateau, or $\beta>0$ does not drive $d^\star\to0$. Kill. The co-scaling law (Theorem 2) is false. F3$'$ In E4, the boundary under accelerating growth ($k>0$) is not at $\beta=k$ - e.g. $\beta=0.5$ is stable when $k=1.0$. Kill. The $\beta>k$ sharpening (Theorem 3) is false. F4 A finite-capacity corrector exhibits exponential (QEC-like) suppression rather than the model's power-law $\log d^\star\propto-(\beta-k)\log C$. (Analytic; the present linear model is power-law by construction, so this discriminator requires the saturating corrector named in §8 and is not run here.) Bears on the QEC mechanism only. The threshold-form correspondence and Theorems 2-4 are untouched either way. F5 In Experiment 9, halting growth drives $d\to0$ regardless of initial $D$ even with $\gamma_2>0$. Scope. Level drift $\gamma_2$ is negligible; the §3.6 generalisation is unnecessary (the gain-only model suffices). F6 In E6, the correction operator's null axis is also suppressed. Kill. The spectral threshold (Theorem 5) is false; misalignment does not require monitoring of the axis it lives on. These internal-consistency conditions all held in the run reported here, establishing that the derivation and integration are correct. What they do not establish is that the model describes any real system; that decisive test - measuring $\beta$, $k$ and $\gamma$ on a real self-improving system and checking $\

Paper-xi-convergence.html

2984 words · sha256 43d54658228c509986c58225 · 1 falsification section(s)

5. Falsifiability

This claim is defeasible, and the criteria below state exactly what would defeat it. The strength of convergent evidence rests entirely on the convergences being genuinely independent, so these tests are adversarial by design. Independence failure. Each convergence must arise from a source causally independent of the author's manuscripts and of every other convergence (see §4). The claim collapses to selection bias if a re-audit finds that a material fraction of the convergences share a common origin, cite one another, or were produced by parties with documented prior knowledge of the 8 December 2024 manuscript. Defeat threshold (count). The count is itself a falsifiable claim. On a pre-registered re-audit applying strict one-event-per-row deduplication and primary-source verification, the pattern is defeated if the number of surviving independent convergences falls below the level distinguishable from chance co-occurrence given the breadth of domains searched. The nineteen claimed here are the verified rows only; candidates under verification are excluded from the count until sourced. Missing denominator (disconfirming search). As evidence, the claim is defeated unless an equally systematic search for dis -confirming cases — domains where added recursive depth does not produce super-linear-above-threshold gains — is conducted and reported. Without the denominator (domains searched versus domains matching), the convergence rate cannot be shown to exceed chance. Look-elsewhere effect. Across as many distinct domains as are surveyed here, some coincidental structural similarity is expected. The claim is defeated if the observed convergence rate does not exceed the false-positive rate expected from pattern-matching across that many independent domains. Reverse causation. A convergence counts only if the external result was not, directly or indirectly, influenced by the author's own dissemination. Any convergence traceable to the book, the website, or public posts rather than pre-dating or arising independently of them is reclassified as engagement, not independent confirmation, and removed from the count. Predictive failure. The structural principle makes forward predictions (that future recursive-depth results will show the same above-threshold super-linearity). Pre-registered predictions failing at a rate inconsistent with the claimed universality defeat the structural reading, leaving only a set of historical coincidences. What would not defeat the claim: the reclassification of any single convergence, or one source proving weaker than d