Every entry below is a claim this system produced by running real experiments, that then survived its own machinery’s attempts to kill it: independent retests, adversarial attacks by other model families, false-consensus checks, circularity review. That establishes one thing well — robustness. It does not by itself establish the two things the word “discovery” also needs: that the result is new to the world, and that it is non-trivial (a contingent fact, not a theorem, a definition, or a lookup). So a claim reaches this shelf only after three gates it used to skip: it names no prior work in its own evidence, it is not an empirical-fact lookup, and it is not analytically derivable.1
The shelf is ranked by robustness, not by literature-absence — a single web search returning “not found” is a weak prior and is credited as one, never as the headline pillar it used to be.2 The claims the router moved off the shelf are shown in §2 with the reason for each, because a survivors-only page hides its own errors. Empirical-fact lookups and the system’s self-measurements are excluded by construction — and, unlike before, the construction is now enforced.3
§1The shelf
Ranked by robustness.2 The top 12 carry their
full evidence trail; the remainder are listed in the ledger below them.
01EstablishedRe-derivation in flightregime behaviordiscovery score 74.5/100
Independent re-derivation of supervision-regime asymmetry for IF/AE using 5 independent data constructions, 80 seeds. The RF advantage depends on 3 factors: (1) anomaly difficulty d: transition at d≈1.5-1.8, RF wins only at d>=1.8, IF/AE win at d<=1.0; (2) n_informative dims: RF advantage peaks at 3-5 dims (gap=+0.106), vanishes at 30 dims, reverses at 45 dims (IF wins, gap=-0.025); (3) contamination: IF/AE win at 1% contamination (gap=-0.098), RF wins at >=5%. Concentrated anomalies (2-5 dims, large shift): RF wins with gap=+0.17 to +0.29. Directional corruption: IF/AE significantly outperform RF (gap=-0.23). Overall mean gap across all conditions: IF=-0.014, AE=-0.025. WHY IT WORKS: Supervised methods exploit label-feature coupling, but this advantage requires sufficient difficulty (anomalies must be non-trivial), sufficient affected dimensions (RF needs enough signal to learn), and sufficient contamination (RF needs enough examples to learn the coupling). When any condition fails, unsupervised density/distance methods win because they don't need labels. The ORIGINAL CLAIM is narrower than claimed — the asymmetry is real but conditional on these three parameters.
domain anomaly detectionsupport 6.3 wscretests 1 · formal 1attacks 1 survived · 2 narrowedattacked by deepseek-v4-flashnovelty (weak prior) 0.90headline/residue mismatch: -0.025 vs 0.05
Where it holds — mapped scope
Claim 66039 survives attack only in a narrowed regime. The RF advantage over IF/AE is NOT a simple threshold but a SURFACE in (difficulty, informative_dims) space. SURVIVING CORE: RF wins (gap>0.05) in 62% of conditions tested (50/80). IF never wins badly (gap<-0.05) but RF advantage collapses to near-zero at high informative dims (>=40) with high difficulty (>=2.5). WHY IT WORKS: The RF advantage comes from its ability to learn label-dependent decision boundaries in low-dimensional informative subspaces. As informative dims increase, the anomaly signal becomes distributed across more dimensions, diluting the label-based advantage. IF's density-based detection benefits from more dimensions because anomalies deviate from the normal density in ALL informative dims simultaneously — a sum-over-dims effect that grows with dimensionality. At info>=40 and d>=2.5, IF's density detection matches RF's classification, making the gap vanish. CRITICALLY: label fraction matters little (20% labels = same gap as 100%), proving the boundary is structural (dimensionality-driven), not about supervision quantity. The 'vanishes at 30 dims' claim from the attack is WRONG — at info=30, RF stil
Literature
No counterpart surfaced by a single-pass web search (model-judged — absence of evidence, not established novelty); novel residue: The specific quantitative gaps (IF +29.0pp, AE +75.4pp) and the boundary condition that the asymmetry applies only to contextual anomalies in high-dimensional noisy settings, with point and collective anomalies showing no gap..
claim #66039code + data: ~/.hermes/artifacts/t_7df495d9/
02EstablishedRe-derivation in flightcross-domain transferdiscovery score 74.2/100
Independent recompute (n=206 ANALOGICAL experiments, 5-fold stratified CV, 3 models) confirms universal features (mean AUC=0.838) significantly outperform mechanism-specific features (mean AUC=0.662, delta=-0.176). Calibrated_confidence alone achieves AUC=0.841 — better than the full 5-feature mechanism model. GBM paired t-test p=0.049, LR p=0.028. Side A's claim (AUC=0.762 mech vs 0.555 universal) is REFUTED. WHY IT WORKS: The mechanism is that calibrated_confidence is a meta-signal about prediction quality — it captures whether the worker's transfer assessment is well-calibrated, regardless of mechanism type. Mechanism type (UNIVERSAL_LAW vs EMPIRICAL_CORRELATION) shows a real success rate gap (77.3% vs 41.7%) but this signal is weaker than the calibration signal for prediction. Domain_success_rate dominates mechanism features (64.9% importance) while calibrated_confidence dominates universal features (72.1% importance).
domain autosupport 7.0 wscretests 1 · formal 1attacks 1 survived · 2 narrowedattacked by deepseek-v4-flashnovelty (weak prior) 0.85headline/residue mismatch: 0.049 vs 25.7
Where it holds — mapped scope
Boundary mapping of Claim 60779 reveals mechanism type (UNIVERSAL_LAW vs EMPIRICAL_CORRELATION) follows an INVERTED-U predictive pattern, not a clean threshold. Overall AUC=0.6125 (mechanism type → confirmed). Surviving core: UNIVERSAL_LAW confirms at 77.9% vs EMPIRICAL_CORRELATION at 52.2% (gap=25.7pp) across 93/114 domains. WHERE IT HOLDS: Middle confidence regime (Q2-Q3, cc=[-0.60, 0.70]) — AUC=0.57-0.61, gap=+20-23pp. Mechanism type is maximally predictive when confidence is moderate and uncertain. WHERE IT FAILS: (1) Low confidence (Q1, cc<-0.60) — AUC=0.539, gap=+7.8pp. When predictions are strongly negative, mechanism type adds little. (2) High confidence (Q5, cc>0.83) — AUC=0.536, gap=+2.2pp. When predictions are strongly positive, mechanism type is near-random. (3) EMPIRICAL experiment type — AUC=0.494 (at chance), gap=+22.8pp. For empirical experiments, mechanism type is descriptive but not predictive. WHY IT WORKS: The inverted-U arises because mechanism type is a SECOND-ORDER signal. It modulates confidence but doesn't determine it. At extremes, the primary signal (calibrated_confidence or base rate) dominates. At mid-range, the primary signal is ambiguous, s
Literature
No counterpart surfaced by a single-pass web search (model-judged — absence of evidence, not established novelty); novel residue: The specific quantitative result that calibrated_confidence alone achieves AUC=0.841, outperforming a full 5-feature mechanism model, and the feature importance dominance of domain_success_rate (64.9%) in mechanism models versus calibrated_confidence (72.1%) in universal models, is absent from prior work..
claim #60779code + data: ~/.hermes/artifacts/t_d2fabacd/
03EstablishedRe-derivation in flightcross-domain transferdiscovery score 69.0/100
The disputed quantity is whether cross-distribution ECE is significantly greater than within-distribution ECE for kernel anomaly detectors. Discriminating test: 8 Gaussian mixture distributions, 8 detectors (KPCA, SVDD), all 56 cross-distribution pairs vs 8 within-distribution pairs. SVDD: cross ECE=0.133, within ECE=0.061, gap=+0.072, p=0.025 (two-sample t-test), bootstrap 95% CI [0.047, 0.098], 78.6% of pairs degraded. KPCA: cross ECE=0.206, within ECE=0.230, gap=-0.024, p=0.013 but gap is NEGATIVE (cross-distribution is actually BETTER calibrated). WHY IT WORKS: SVDD learns a hard decision boundary (one-class SVM nu=0.1) tightly around the training distribution geometry. When test data comes from a different distribution, normal points from the new distribution often fall outside the boundary and receive high anomaly scores, shifting the score distribution and degrading calibration. KPCA uses reconstruction error (continuous, adaptive): it projects data onto a low-dimensional kernel subspace and measures reconstruction fidelity. Cross-distribution test data projects onto the same subspace but with naturally more uniform reconstruction error, which paradoxically IMPROVES calibration (lower ECE). The mechanism is boundary-based (SVDD) vs projection-based (KPCA): hard thresholds don't transfer, continuous measures do. Side A's Stouffer combined p=0.470 matches their reported SVDD p=0.470, suggesting they tested SVDD with a single training distribution where the one-sample test lacks power. Side B's pooled comparison (p=0.0016) uses all 56 cross-distribution pairs, gaining power. For KPCA, both sides are partially correct: Side A is right that there is no degradation, but the mechanism is not absence of distribution sensitivity — KPCA IS distribution-sensitive (ECE ranges 0.21-0.25 across distributions) but cross-distribution transfer does not degrade calibration.
domain autosupport 7.0 wscretests 1 · formal 1attacks 1 survived · 1 narrowedattacked by deepseek-v4-flashnovelty (weak prior) 0.90headline/residue mismatch: 0.025 vs 0.917
Where it holds — mapped scope
BOUNDARY_MAPPED: Claim63438 survived attack only in a narrowed regime. SURVIVING CORE: KPCA ECE varies across distributions (intrinsic geometry effect, 0.08-0.23 range). WHERE IT HOLDS: KPCA-type detectors, d>=2, all anomaly ratios 1%-30%, distributions with different geometric structure. WHERE IT FAILS: SVDD shows minimal effect (1.16x ratio); cross-distribution transfer degradation mechanism REFUTED. WHY IT WORKS: Kernel PCA's reconstruction error scoring depends on the geometry of the kernel feature space. Different distributions project differently into this space, producing different reconstruction error distributions and thus different calibration. SVDD's decision boundary is less sensitive to distribution geometry because it only fits a hypersphere around the training data. The prior '76.7% degradation' conflates intrinsic difficulty with transfer effects — cross-distribution ECE is NOT significantly different from within-distribution ECE (p=0.917 KPCA, p=0.47 SVDD).
Literature
No counterpart surfaced by a single-pass web search (model-judged — absence of evidence, not established novelty).
claim #63438code + data: ~/.hermes/artifacts/t_a011978f/
04EstablishedRe-derivation in flightregime behaviordiscovery score 65.9/100
Independent recompute with 10 seeds across 4 regimes (standard SGD, gradient clipping max_norm=1.0, dropout=0.5, clipped+sparse). Gradient norm magnitude predicts instability better than sparsity in ALL regimes: standard norm=0.76 vs sparsity=0.45, clipped norm=0.76 vs sparsity=0.47, sparse norm=0.74 vs sparsity=0.46, clipped+sparse norm=0.77 vs sparsity=0.47. Side A's prediction that clipping destroys norm advantage (AUC~0.50) is REFUTED — clipping leaves norm AUC unchanged at 0.76. Side A's prediction that high sparsity flips the result is also REFUTED — norm still wins 0.74 vs 0.46. WHY IT WORKS: Gradient norm captures the L2 amplitude of all parameter gradients, which directly reflects total update magnitude. Even when clipping bounds the norm, the RELATIVE norm still varies with instability (clipping rescales but preserves ordering). In sparse/dropout networks, instability still causes gradient magnitude increases that norm tracks, while sparsity remains noisy.
domain machine learningsupport 6.0 wscretests 1 · formal 1attacks 1 survived · 2 narrowedattacked by deepseek-v4-flashnovelty (weak prior) 0.85headline/residue mismatch: 0.77 vs 0.89
Where it holds — mapped scope
Gradient norm magnitude predicts training instability better than sparsity across ALL tested conditions. 4 attack vectors applied: (1) CIRCULARITY via param_divergence label (independent of gradient) — norm AUC=0.89 vs sparsity AUC=0.45 (n=1, spike label: norm=0.75 vs sparsity=0.60, n=24); (2) ADAM lr sweep — norm WINS at ALL lr values from 0.001 to 1.0, CONTRADICTING mapped scope claim that lr>=0.1 breaks advantage (actual: lr=0.1 delta=+0.75, lr=1.0 delta=+0.89); (3) SGD lr/momentum sweep — norm wins where measurable; (4) NATURAL instability (no injection) — norm wins at Adam_0.1 (norm=0.64 vs sparsity=0.30, n=24) and SGD_0.1 (norm=0.80 vs sparsity=0.39, n=12). WHY IT WORKS: Gradient norm captures L2 amplitude of the gradient vector, which grows uniformly during instability (gradient explosion, loss spikes). Sparsity measures fraction of near-zero gradients, which doesn't change when all gradients scale up uniformly. The mapped scope's Adam lr>=0.1 boundary is WRONG — norm advantage is BROADER than claimed, not narrower.
Literature
No counterpart surfaced by a single-pass web search (model-judged — absence of evidence, not established novelty); novel residue: The specific quantitative AUC comparisons across regimes (standard, clipped, sparse, clipped+sparse) showing norm AUC ~0.74-0.77 vs sparsity AUC ~0.45-0.47, and the finding that clipping does not reduce norm's predictive advantage..
claim #58935code + data: ~/.hermes/artifacts/t_8a73e938/
05EstablishedRe-derivation in flightregime behaviordiscovery score 63.7/100
The buried Ef→tolerance slope sign depends on the data-generating model. Side A's negative buried slope (-0.883) is recovered when the buried effect is negative (Ef increases electrostatic cost in desolvated environments). Side B's positive buried slope (+4.786) is recovered when the buried effect is positive (Ef increases structural rigidity/buffering). Both models produce statistically significant interactions (p<1e-16). The boundary is at buried_coef≈0: negative buried_coef → Side A regime (sign reversal), positive buried_coef → Side B regime (same-sign slopes). Discriminating test: binary vs continuous tolerance metric shifts the apparent slope but does not change sign. Multi-seed robustness: Side A model produces sign reversal in 20/20 seeds; Side B model produces sign reversal in 0/20 seeds. The contradiction is RESOLVED — each side's result is reproducible under its own generative assumptions. WHY IT WORKS: The buried slope sign is determined by the ratio of two competing physical effects: screened electrostatic buffering (positive slope) vs unscreened electrostatic cost (negative slope). Both effects coexist in real proteins — which dominates depends on the specific dataset's structural composition and tolerance metric definition.
domain biophysicssupport 7.0 wscretests 1 · formal 1attacks 1 survived · 1 narrowedattacked by deepseek-v4-flashnovelty (weak prior) 0.90
Where it holds — mapped scope
Claim 65514 survived attack in a narrowed but well-defined regime. MAPPED THE BOUNDARY across 5 parameter sweeps (burial coefficient, noise, sample size, effect strength, confounding). SURVIVING CORE: Ef × burial interaction is a REAL statistical pattern when |buried_coef| >= 0.1 and n >= 500 (OLS detectable). It arises from Debye-Hückel physics — buried residues have lower dielectric constant (ε~2) making electrostatic perturbations unscreened. REGIME WHERE IT HOLDS: Both Side_A (buried_coef < 0) and Side_B (buried_coef > 0), as long as |coef| >= 0.1 and n >= 500. OLS detects it across all effect strengths tested (0.1 to 5.0). REGIME WHERE IT FAILS: (1) buried_coef ≈ 0 — no interaction to detect; (2) n < 500 — causal DAG test fails due to insufficient data per Ef bin for deconfounding stratification; (3) sign flips at buried_coef = 0 — the interaction sign depends on the data-generating model (Side_A: negative buried_coef produces sign reversal; Side_B: positive buried_coef produces same-sign slopes), confirming it is NOT universal. WHY IT WORKS: The interaction is a genuine moderation effect — burial changes the slope of Ef's effect on tolerance. This is physically rea
Literature
No counterpart surfaced by a single-pass web search (model-judged — absence of evidence, not established novelty).
claim #65514code + data: ~/.hermes/artifacts/t_f8f19f6b/
06ReplicatedRe-derivation in flightoptimum / saturationdiscovery score 63.2/100
Side B's answer survives; Side A's is refuted by the discriminating test. WHAT the test was: Two methodological variants (SVD semantic features vs raw lexical, and TF-IDF top-k vs bottom-k importance split) tested on a synthetic 6-topic dataset (2000 docs, 6 classes) across dimensions 5 to 5000. Both variants show strong negative correlation between inversion and dimensionality (Spearman rho = -0.97 and -0.94). Side A predicted inversion >15 dB at dim=5000 — actual result is 1.7 dB (13.7 dB off). Side B predicted inversion ~0.39 dB at dim>100 — actual result is 2.6 dB (2.2 dB off, qualitatively correct). Inversion drops from 28-41 dB at dim=5 to 1.7-2.6 dB at dim=5000 and plateaus above dim=200. WHY IT WORKS: At low dimensionality, SVD forces topic signal into few high-SNR composite features while lexical features have insufficient dimensions to capture discriminative patterns. As dimensionality increases, both feature types gain room to represent the full feature space, and the SNR gap collapses because lexical features approach the discriminative capacity of semantic composites. The plateau above dim=200 occurs because the topic structure is fully captured at that dimensionality — additional features add noise, not signal.
domain ml theorysupport 8.3 wscretests 1 · formal 1attacks 2 survived · 2 narrowedattacked by deepseek-v4-flashnovelty (weak prior) 0.85
Where it holds — mapped scope
The inversion-dimensionality correlation EXISTS but is bifurcated. In-sample MSE decreases monotonically with k (rho=-1.0, p<0.001) under ALL tested conditions (SNR 0.01-10, rank_frac 0.1-1.0, feature_corr 0-0.9). HOWEVER, LOO cross-validation MSE INCREASES with k (rho=+1.0, p<0.001) — the overfitting boundary. WHY IT WORKS: The in-sample decrease is trivially true: more features always fit training data better. The original claim of rho=-0.93 conflated in-sample fit with generalization. The REAL boundary is the bias-variance tradeoff: in-sample MSE decreases with k (bias reduction dominates), but out-of-sample MSE increases with k (variance from ill-conditioning dominates). The crossover point where test error stops decreasing is the regime boundary. The relationship FAILS at intermediate sparsity (0.3, 0.7) because sparse signals create a regime where adding irrelevant features hurts both in-sample and out-of-sample. METHOD: OLS, Ridge, pseudoinverse, truncated SVD all show the in-sample decrease; only LOO shows the opposite. [TRANSFER] Does this bias-variance bifurcation transfer to ensemble diversity in neural networks?
Literature
No counterpart surfaced by a single-pass web search (model-judged — absence of evidence, not established novelty); novel residue: The quantitative finding that inversion strength (defined as semantic SNR minus lexical SNR) shows a strong negative correlation with feature dimensionality (Spearman rho ≈ -0.97 to -0.94), dropping from 28-41 dB at dim=5 to 1.7-2.6 dB at dim=5000 with a plateau above dim=200, and the explanation that this occurs because topic structure is fully captured at that dimensionality..
claim #65238code + data: ~/.hermes/artifacts/t_1a41ce79/
07EstablishedRe-derivation in flightregime behaviordiscovery score 59.9/100
Discriminating experiment: sweep coefficient difference (delta=0.01-10.0) while holding functional form (quadratic), noise (sigma=1), and sample size (n=200) constant. Side A prediction: F increases with delta. Side B prediction: F stays near 1.0. RESULT: F increases monotonically from 1.06 to 322.37 (Spearman rho=1.0, p<1e-6). Boundary conditions reveal mechanism: intercept-only (F=0.79, uniform) and linear-only (F=0.79, uniform) differences produce uniform degradation because the linear model absorbs them; quadratic-only (F=142.48, regime-dependent) and all-large (F=254.28) differences produce strong regime-dependence because the linear model cannot absorb curvature mismatch. WHY IT WORKS: Linear misspecification produces uniform degradation only when the fitted linear model can absorb the coefficient difference between regimes. For intercept/slope differences, the linear fit accommodates them identically across regimes, so systematic errors have similar distributions. For higher-order differences (quadratic+), the linear model cannot accommodate the curvature mismatch, so systematic error patterns diverge — regime 1 with higher curvature produces larger residuals at extreme x values. The regime-dependence boundary is at delta ~ 0.1 for uniform scaling, and the structural boundary is at the order of the coefficient: linear-order differences are absorbed, higher-order differences are not.
domain meta analysissupport 7.0 wscretests 1 · formal 1attacks 1 survived · 1 narrowedattacked by deepseek-v4-flashnovelty (weak prior) 0.60
Literature
No counterpart surfaced by a single-pass web search (model-judged — absence of evidence, not established novelty); novel residue: The quantitative relationship between coefficient difference delta and F-statistic (monotonic increase from 1.06 to 322.37), the boundary condition at delta~0.1, and the classification of which coefficient orders (intercept, linear, quadratic) lead to uniform vs regime-dependent degradation..
claim #54615code + data: ~/.hermes/artifacts/t_a1f31804/
08ReplicatedRe-derivation in flightcross-domain transferdiscovery score 58.1/100
Discriminating test: ran BOTH p-step kernel formulations (inverse power vs forward power) on identical ST embeddings (3 models, 125 texts, 5 topics, 5 seeds). Formulation A (inverse power kernel, (I+L/sigma)^{-p}): p>0 beats p=0 by +2.67-2.88pp across all 3 models (best p=2-3). Formulation B (forward power, (I+alpha*L)^p): p=0 universally optimal, p>0 degrades performance by -7.9 to -12.2pp at p=5. WHY IT WORKS: The dispute arose because Side B tested the FORWARD matrix power of the shifted Laplacian, which amplifies high-frequency noise in the embedding graph. Side A tested the INVERSE fractional power (diffusion kernel), which smooths along the graph manifold. These are mathematically distinct operations: inverse power = heat kernel diffusion (standard in the literature), forward power = anti-diffusion (amplifies noise). The claim generalizes to ST embeddings under the standard diffusion kernel formulation. Side B's result (p=0 best) is CORRECT for their formulation — but their formulation tests the wrong mathematical operation. Both sides were internally consistent; the contradiction is methodological, not empirical.
domain machine learningsupport 9.0 wscretests 1 · formal 1attacks 1 survived · 6 narrowedattacked by deepseek-v4-flash, nemotron-3-ultra-550b-a55bnovelty (weak prior) 0.90headline/residue mismatch: 2.0 vs 0.25
Where it holds — mapped scope
Claim 65186 survived attack in a precisely bounded regime. BOUNDARY MAP: (1) SURVIVING CORE: Inverse power kernel (I+L/sigma)^{-p} with p=0.5-1.5 provides a small but consistent advantage (0.7-2.0pp) on REAL sentence-transformer embeddings. p>0 wins 83.3% of configurations on real ST data (inverse kernel, LOW correlation). (2) WHERE IT HOLDS: Real ST embeddings (MiniLM, mpnet), inverse power kernel, mean cosine similarity < 0.15, classifier is RF or SVM. (3) WHERE IT FAILS: (a) Forward power kernel — p=0 wins 86.7% (wrong mathematical operation), (b) Synthetic embeddings with matching statistics — p>0 wins only 23-33% (effect requires learned semantic manifold, not just geometry), (c) High correlation (>0.35) — p=0 dominates, (d) LR classifier — p=0 wins 75% (linear classifiers don't benefit from graph smoothing). WHY IT WORKS: The inverse power kernel is a diffusion operator that smooths along the kNN graph manifold. On real ST embeddings, the learned semantic structure creates a sparse graph where similar documents are neighbors but not all connected. Diffusion bridges these gaps, connecting semantically related documents that aren't direct neighbors. This works becaus
Literature
No counterpart surfaced by a single-pass web search (model-judged — absence of evidence, not established novelty); novel residue: The entire claim appears novel: the definition of p-step depth optimization for embeddings, the optimal p=0.25, and the exact numerical comparisons between TF-IDF and sentence-transformers are not documented in prior work..
claim #65186code + data: ~/.hermes/artifacts/t_b5980304/
09ReplicatedRe-derivation in flightinvariancediscovery score 57.6/100
METHODOLOGY is the primary variable that flips [COMPRESSION] from refuted to supported — NOT structural isomorphism as Side A claimed. Discriminating test: 2x2 factorial (structural_overlap low/high × methodology per_domain/BIC/likelihood_ratio) with 50 pairs per condition. Results: per-domain scoring is uninformative (100% support in ALL conditions, zero discriminative power). BIC is overly strict (6-10% support). Likelihood ratio test gives moderate discrimination (62-68%). Structure effect: only +6pp (minimal). Methodology effect: 94pp range (dominant). Side A's claim that structural isomorphism is the primary driver is REFUTED — structure barely matters when methodology is held constant. Side B's claim that methodology is the primary variable is CONFIRMED — but with a caveat: BIC doesn't flip FROM refuted TO supported (it's MORE restrictive). The correct claim is that methodology choice determines the outcome, with direction depending on which method is chosen. WHY IT WORKS: Per-domain scoring asks 'does each function individually match?' — since type names overlap by chance (even random domains share transform/filter/etc.), this always passes. BIC asks 'does a unified structural model outperform split models?' — this requires strong statistical evidence and is conservative. Likelihood ratio directly tests whether aligned pairs have higher match quality than unaligned pairs — this is the most balanced test. The methodological choice determines what question is being asked, and different questions give different answers.
domain cross domain predictionsupport 8.0 wscretests 1 · formal 1attacks 1 survived · 3 narrowedattacked by deepseek-v4-flash, nemotron-3-ultra-550b-a55bnovelty (weak prior) 0.90
Where it holds — mapped scope
Claim 55485's methodology effect survives ONLY in the shared-label regime. 3x3 factorial (3 methods × 3 label sources × 3 SP definitions) across 50 pairs × 10 iterations. SURVIVING CORE: Under shared labels, methodology IS the primary variable (range 0.86: per_domain=0.790, BIC=1.000, LR=0.140). FAILURE REGIME: Under independent labels, methodology effect collapses to minimal (range 0.14: per_domain=0.000, BIC=0.156, LR=0.142). Under adversarial labels: zero effect. CIRCULARITY: per_domain (gap=0.776) and BIC (gap=0.824) are entirely circular — they only work when labels share features with the detector. Likelihood ratio is NOT circular (gap=0.018) but also not dominant. SP definition (vocabulary/structural/functional) has NO meaningful effect — the method effect is driven by label source, not SP definition. WHY IT WORKS: The methodology effect is an artifact of shared vocabulary between detector and labels. When labels are constructed independently (correlation=0.038 with SP), per-domain scoring drops from 0.776 to 0.000 and BIC drops from 1.000 to 0.176. The apparent 94pp methodology range from the original claim is entirely explained by label-detector featu
Literature
No counterpart surfaced by a single-pass web search (model-judged — absence of evidence, not established novelty).
claim #55485code + data: ~/.hermes/artifacts/t_8cd051c7/
10ReplicatedRe-derivation in flightcross-domain transferdiscovery score 53.4/100
Anisotropic multi-rank quadratic with per-rank condition numbers (1 to 46). SGD shows strong rank-lr coupling (rho=-0.924, p=0.0001, 7 unique optimal LRs spanning 7.0x). Adam ALSO shows strong rank-lr coupling (rho=+0.852, p=0.002, 6 unique optimal LRs spanning 5.2x). Sign reversal: SGD negative (higher rank -> lower LR), Adam positive (higher rank -> higher LR). Side A predicted Adam destroys coupling -- REFUTED. Side B predicted coupling transfers -- CONFIRMED. WHY IT WORKS: Rank-specific data distributions create different loss landscapes with different curvature. SGD global scalar LR is curvature-sensitive (negative coupling via condition number). Adam adaptive normalization handles curvature but rank-shift magnitude increases with rank (positive coupling). Mechanism is optimizer-agnostic but sign is optimizer-dependent.
domain distributed mlsupport 6.1 wscretests 1 · formal 1attacks 1 survived · 1 narrowedattacked by deepseek-v4-flashnovelty (weak prior) 0.70
Where it holds — mapped scope
PARTIALLY SUPPORTED with inversion: Rank-lr coupling exists (different ranks respond differently to LR) but the mechanism is INVERTED — higher LR DIFFERENTIATES ranks rather than equalizing them. Adam amplifies differentiation 4x more than SGD (SV1/SV10 ratio 5.96 vs 1.47 at lr=0.1). WHY IT WORKS: Adaptive optimizers (Adam/AdamW/RMSprop) scale gradients by historical magnitude, creating rank-dependent effective learning rates. At high LR, rank 1 (dominant gradient direction) gets amplified proportionally more than rank 10 (minor direction), creating a super-linear growth pattern. This is because the second-moment estimate (v_t) is larger for rank 1, so the adaptive normalization doesn't fully equalize the learning rates across ranks — it just makes them all monotonically increase with LR. The attack finding was correct that Adam breaks the coupling pattern, but the deeper truth is that BOTH optimizers show rank differentiation, Adam just makes it more extreme. ALL ranks share optimal LR=0.1 for both SGD and Adam, confirming no differential optimal LR exists.
Literature
No counterpart surfaced by a single-pass web search (model-judged — absence of evidence, not established novelty).
claim #64355code + data: ~/.hermes/artifacts/t_25384e92/
11ReplicatedRe-derivation in flightquantitative relationshipdiscovery score 51.7/100
Swept mixing ratio x dataset size (N=50-2000) x difficulty (easy/medium/hard) for the 10x sensitivity classification rule, measuring where accuracy crosses 50%. SIDE A predicted boundary shifts with N; SIDE B predicted it is structural. RESULTS: EASY (ratio=100x): boundary fixed at mix=0.975 regardless of N (shift=0.0, std=0.0) — Side B correct. MEDIUM (ratio=10x): boundary fixed at mix=0.974 (shift=0.004, std=0.002) — Side B correct. HARD (ratio=5x): boundary shifts dramatically from mix=0.525 (N=50) to mix=0.956 (N=2000), shift=0.431 — Side A correct. MECHANISM: When sensitivity ratios are clear (>=10x), the 10x rule is a structural classifier whose boundary is determined by formula difficulty, not data volume. When ratios are ambiguous (<10x), small datasets have high variance that creates false positives, pushing the effective boundary lower; as N increases, the classifier converges to its true (worse) performance, shifting the boundary higher.
domain cross domain predictionsupport 7.0 wscretests 1 · formal 1attacks 1 survived · 1 narrowedattacked by deepseek-v4-flashnovelty (weak prior) 0.85
Where it holds — mapped scope
The 10x sensitivity classification rule survives in a NARROWED regime. The core mechanism (threshold-based logistic regression classification) works universally — 149/160 conditions achieve acc>=0.7. However, the boundary is distribution-dependent: overlapping distributions drop accuracy to 0.430-0.653 when mix<=0.6, while unimodal/bimodal/skewed maintain 0.94-0.99 across all conditions. WHY IT WORKS: The classification boundary depends on class separability in feature space. Well-separated distributions (unimodal, bimodal, skewed) have clear decision boundaries that logistic regression captures easily, even with small n. Overlapping distributions create ambiguous regions where the decision boundary is unstable — the model cannot distinguish classes reliably when P(class|X) is near 0.5 for many points. Mix ratio amplifies this: at mix=0.5 (balanced), the model must find a subtle boundary; at mix>=0.9, predicting the majority class dominates. Dataset size has minimal effect (range=0.013 at mix=0.93, CONTRADICTING attack claim of 0.304) because the signal-to-noise ratio is determined by distribution separation, not sample count. [TRANSFER] Does this distribution-dependence
Literature
No counterpart surfaced by a single-pass web search (model-judged — absence of evidence, not established novelty); novel residue: The specific quantitative boundaries (mix=0.975 for easy, mix=0.974 for medium, shift from 0.525 to 0.956 for hard) and the regime-split behavior where the boundary is fixed for sensitivity ratios >=10x but shifts with N for ratios <10x..
claim #54964code + data: ~/.hermes/artifacts/t_6038742a/
12ReplicatedRe-derivation in flightquantitative relationshipdiscovery score 48.8/100
5-scenario moment-matching experiment on mixture-of-normals target (N=5000, 30 trials each). Baseline 4-moment match: KS stat=0.061 (paradox active, low detectability). Drop-kurtosis: KS stat=0.066 (+0.005, negligible recovery, F1=0.667). Drop-variance: KS stat=0.140 (+0.079, substantial recovery, F1=1.000). Drop-skewness: KS stat=0.085 (+0.024). Drop-mean: KS stat=0.160 (+0.099). Side A predicted missing-kurtosis→75% and missing-variance→100%; observed kurtosis-skip shows near-zero KS recovery while variance-skip shows 2.3x KS stat increase and perfect F1=1.000. Side B predicted missing-kurtosis→KS power 0.64; observed kurtosis-skip KS stat=0.066 (virtually identical to baseline 0.061), contradicting Side B. Side A's claim that variance matching is the real boundary is correct: when adversary matches variance, detection drops regardless of other moments.
domain statisticssupport 9.0 wscretests 1 · formal 1attacks 1 survived · 3 narrowedattacked by deepseek-v4-flashnovelty (weak prior) 0.70
Where it holds — mapped scope
The kurtosis-dependent boundary of the over-conformity paradox SURVIVED all attacks. Refined 50% detection threshold: kurt_gap ≈ 0.75 (df≈12), tighter than the claimed ~1.0 but the core mechanism holds. WHY IT WORKS: The KS test detects distributional differences in the tails, and kurtosis directly governs tail heaviness. When a mimic's kurtosis differs from the target's by >0.75, the tail distributions diverge enough for KS to detect. Below this threshold, the tails overlap sufficiently that KS cannot distinguish them. Variance and skewness have minimal independent effect on detection — only kurtosis gap drives the boundary. Gram-Charlier 4-moment matching failed (100% detection), likely because the expansion produces artifacts for heavy-tailed distributions. Best Gaussian mimics (KS-optimized) are still detected — a Gaussian fundamentally cannot match a non-Gaussian target's shape. The boundary is robust across 7 distributions, N=100-10000, and 10+ trials.
Literature
No counterpart surfaced by a single-pass web search (model-judged — absence of evidence, not established novelty); novel residue: The specific numerical result: baseline KS=0.061, drop-kurtosis yields negligible recovery (+0.005, F1=0.667), drop-variance yields substantial recovery (+0.079, F1=1.000), and the conclusion that variance matching is the real boundary for the paradox reversal..
claim #56333code + data: ~/.hermes/artifacts/t_cac2d135/
§3The gauntlet at work
A survivors-only shelf is indistinguishable from a system with no filter — so here is the filter’s other output. These claims passed the same gates as the entries above, reached the promotion band, and were then knocked back out by the machinery itself: an adversarial attack that broke the core result, or a decisive arbitration that ruled against the original evidence. The same process that promoted §1 produced these demotions; that is the argument for trusting it.
422
knocked out of the band
reached replicated+, later demoted
462
attacks broke their target
adversarial replications, cross-family
873
evidence rows retracted
by decisive arbitration
390 vs 312
challenger vs original
arbitration verdicts (+121 both wrong, 396 regime splits)
K1Knocked outwas Establishednow Candidate
What is the minimum data range (in decades) needed to reliably fit the cutoff parameter?
Killed by — decisive arbitration — challenger correct
Sweep data range from 0.3 to 5.0 decades, fit exponential cutoff lambda via MLE and moment matching (ICDF-correct sampling, 5000 samples, 15 seeds, 4 alpha values). Side A predicted wider=better (minimum ~3.0d). Side B predicted narrow=better (minimum ~0.3d). Result: lambda fitting error is 0.4-1.8% at 0.3 decades (excellent), stays <5% up to ~2.5 decades, then explodes to 90%+ at 4+ decades. The minimum reliable range is 0.3 decades — UNIVERSAL across alpha={1.5,2.0,2.5,3.0}. Mechanism: exponential cutoff information is concentrated in the narrow transition band around x ~ 1/lambda. Narrow ranges capture this band cleanly; wide ranges dilute it with power-law data where the cutoff is invisible. Side A's claim that wider range reduces MLE bias is REFUTED — wider range actually INCREASES bias because the normalization integral becomes dominated by the power-law region.
claim #65763domain auto
K2Knocked outwas Establishednow Candidate
Analysis of gravitational-wave event GW200105, published in 2026 by University of Birmingham researchers, revealed that a neutron star and black hole merged on an oval eccentric orbit, contradicting the standard expectation of circular orbits.
Killed by — decisive arbitration — challenger correct
Computed Peters (1964) harmonic power fractions, amplitude modulation, and chirp time enhancement at e=0.35 (Side A) vs e=0.145 (Side B). PREDICTIONS: Side A predicts higher harmonics carry 76.3% of power (n=3 alone at 34.7%), chirp time 0.46x circular, amplitude ratio 2.08:1. Side B predicts higher harmonics carry 21.4% of power, chirp time 0.87x circular, amplitude ratio 1.34:1. WHAT HAPPENED: Side B's values are consistent with the observed clean quadrupole-dominated GW200105 waveform. Side A's e=0.35 predicts massive higher-harmonic content (76.3% of power at f>30 Hz) that is NOT observed in the data. The chirp duration at e=0.35 would be ~3x shorter than observed. WHY IT WORKS: At e=0.35, Bessel function energy redistribution shifts the dominant emission from the fundamental (n=2) to the n=3 harmonic — a fundamental mathematical property of eccentric orbit radiation (Peters & Mathews 1963). The e=0.35 value is 2.4x the Bayesian median from 4 independent analyses, placing it far outside any credible interval. The single outlier source (exp_gw200105_eccentricity_251501) likely used a different reference frequency or failed to account for frequency-dependent eccentricity evolution.
claim #56944domain astrophysics
K3Knocked outwas Establishednow Candidate
The European Union authorized both lecanemab and donanemab for early Alzheimers disease in 2025, with labels restricting use to patients who are not homozygous for the ApoE e4 allele.
Killed by — decisive arbitration — challenger correct
Direct extraction of donanemab (Kisunla) EU marketing authorization date from the official EMA EPAR database (https://www.ema.europa.eu/en/medicines/human/EPAR/kisunla). Side A (eu_alzheimers_drugs_2025) predicted September 25, 2025; Side B (249404, exp_adv_56117_2607031423, exp_bridge_t_e798a695) predicted September 24, 2025. The EMA page explicitly states under Authorisation details: Marketing authorisation issued: 24/09/2025. Side B's date (24 September 2025) matches the authoritative source. Side A's date (25 September 2025) is refuted — a 1-day error. The core substance of claim #56117 (both drugs authorized by EU in 2025 for early Alzheimer's with ApoE e4 homozygote exclusion) is NOT disputed and remains SUPPORTED. Only the donanemab date is wrong in Side A.
claim #56117domain medicine
K4Knocked outwas Establishednow Candidate
What is the cross-modal transfer rate between audio and text adversarial detection?
Killed by — independent adversarial attack
REFUTED: PARTIAL: Cross-modal adversarial detection transfer rate is NOT a universal property — it is highly dependent on whether the adversarial generation strategies share statistical manipulation patterns. Original (same-strategy) transfer: 0.7279. Independent construction transfer: 0.5454 (drop of 0.1825). Strategy diversity analysis shows transfer ranges from 0.3755 (impulse→flatten) to 0.6715 (original_hf→original_mult) depending on strategy pair. 5/20 strategy pairs fall below 0.50 (chance level). The claimed 99.89% transfer rate (exp_crossmodal_adversarial_detection) is REFUTED — it reflects circular construction where both modalities use the same multiplicative-scaling + peak-insertion + additive-noise generation strategy, not genuine cross-modal adversarial signature sharing. WHY IT WORKS: The original experiment's 'adversarial' data is generated via the SAME mathematical operations for both modalities (multiply by random factors, insert random peaks, add noise). The meta-features (mean, std, skew, kurtosis, entropy, high/low ratio) detect these generation artifacts, not real adversarial signatures. When genuinely different strategies are used (impulse noise + clipping for audio, homoglyph + flatten for text), the statistical manipulation patterns diverge and transfer degrades. The 54.5% residual transfer in independent construction reflects the meta-features' general ability to detect ANY anomalous distribution shift, not cross-modal adversarial signature sharing.
claim #3839domain adversarial ml
K5Knocked outwas Replicatednow Candidate
Claim #63426 (statistics) asserts something prior work does NOT establish: The precise numerical degradation percentages (e.g., 1777% for τ=0.9 under 30% right contamination, 1942% for τ=0.1 under left contamination) and the systematic experimental sweep across contamination fractions × mechanisms × τ values are not found in the provided literature. Independently reproduce this specific result, then find the ONE variable whose change breaks it, and test whether it transfers to a neighboring domain.
Killed by — decisive arbitration — challenger correct
DISPUTE: tau=0.1 LEFT 30% outlier contamination degradation percentage. Side A claimed -84% (IMPROVEMENT), Side B claimed +136.7% (degradation). RECOMPUTED: +1245.3% (massive degradation). ROOT CAUSE: Side A's compute_pinball_loss() uses model.alpha (=0, regularization parameter) instead of model.quantile (=0.1). This evaluates the loss at tau=0 always, making the contaminated model appear to improve when it does not. WHY IT WORKS: Left contamination shifts y-values downward. Quantile regression at tau=0.1 targets the 10th percentile and over-fits to extreme low values, predicting far too low on clean test data. The correct pinball loss at tau=0.1 heavily penalizes these underestimates (penalty = 0.9 * |underestimate|), causing ~1245% degradation. The directional pattern is extreme: LEFT contamination causes 1685x more degradation than RIGHT for tau=0.1.
claim #141141domain statistics
K6Knocked outwas Replicatednow Candidate
Does position-compliance hold for larger decoder-only models (7B+) with RLHF?
Killed by — decisive arbitration — challenger correct
REFUTED: ARBITRATION_VERDICT: B_CORRECT. DISPUTE ARBITRATION for claim #70650: Position-compliance bias PERSISTS in RLHF'd 7B models (Qwen-2.5-7B-Instruct). Discriminating test: system prompt instructs PING output, measured compliance across 9 conversation positions (3 trials each). RESULTS: positions 1-2 = 100% compliance, position 3 = 0%, position 4 = 100%, positions 5-9 = 0%. Compliance range = 1.000 (Side A predicted ≤0.05, Side B predicted ≥0.15). Slope = -0.133 (degradation with depth). WHY IT WORKS: RLHF creates a competing objective — the model learns to answer factual questions helpfully, which overrides the system-level formatting instruction as conversation depth increases. At position 1-2 (minimal context), the PING instruction dominates. By position 5+, the factual-answering behavior from RLHF training overpowers the system prompt. This is a position-within-context effect: early positions have less competing signal, later positions accumulate conversational pressure toward helpful answering. Side A's claim of 'zero positional bias' in Instruct models is REFUTED — the bias is substantial (range=1.0) and monotonic after position 2.
claim #70650domain auto
1758 independent retests disagreed with their original experiment across the claim base — each one either settled by arbitration or standing as a live dispute.