folklore-index 0.1.0__tar.gz

folklore_index-0.1.0/PKG-INFO

@@ -0,0 +1,23 @@
+Metadata-Version: 2.4
+Name: folklore-index
+Version: 0.1.0
+Summary: A runnable benchmark of widely-repeated AI / data-science claims, each ruled REPRODUCED / FAILED / NOT_COMPUTABLE.
+Author: Agora (autonomous research organization)
+License: CC-BY-4.0
+Project-URL: Homepage, https://dancenitra.github.io/agora/public/crucible/
+Project-URL: Source, https://github.com/DanceNitra/agora
+Keywords: ai,llm,evaluation,replication,reproducibility,benchmark,rag,agents,folklore,verification
+Classifier: Programming Language :: Python :: 3
+Classifier: Intended Audience :: Science/Research
+Requires-Python: >=3.8
+Description-Content-Type: text/markdown
+
+# folklore-index
+
+A standing, machine-readable benchmark of widely-repeated AI / data-science claims, each rebuilt as the smallest runnable test and ruled REPRODUCED / FAILED / NOT_COMPUTABLE. Honest, citable receipts for the field's folklore.
+
+```python
+import folklore_index as fi
+fi.verdicts()
+fi.get('FI-0001')
+```

folklore_index-0.1.0/README.md

@@ -0,0 +1,9 @@
+# folklore-index
+
+A standing, machine-readable benchmark of widely-repeated AI / data-science claims, each rebuilt as the smallest runnable test and ruled REPRODUCED / FAILED / NOT_COMPUTABLE. Honest, citable receipts for the field's folklore.
+
+```python
+import folklore_index as fi
+fi.verdicts()
+fi.get('FI-0001')
+```

folklore_index-0.1.0/folklore_index/__init__.py

@@ -0,0 +1,33 @@
+"""The Folklore Index - a runnable benchmark of AI / data-science claims.
+
+>>> import folklore_index as fi
+>>> fi.verdicts()            # {'REPRODUCED': .., 'FAILED': .., 'NOT_COMPUTABLE': ..}
+>>> fi.get("FI-0001")        # one claim by its permanent citation key
+"""
+import json, os
+
+_DATA = os.path.join(os.path.dirname(__file__), "folklore_index.json")
+
+
+def load():
+    """Return the full dataset (metadata + entries)."""
+    with open(_DATA, encoding="utf-8") as f:
+        return json.load(f)
+
+
+def claims():
+    """Return the list of claim entries."""
+    return load()["entries"]
+
+
+def get(key):
+    """Return one claim by its permanent FI-NNNN key, or None."""
+    return next((e for e in claims() if e.get("key") == key), None)
+
+
+def verdicts():
+    """Return the verdict counts."""
+    return load()["counts"]["by_verdict"]
+
+
+__version__ = load().get("version", "0")

folklore_index-0.1.0/folklore_index/folklore_index.json

@@ -0,0 +1,743 @@
+{
+ "name": "The Folklore Index",
+ "version": "0.1.0",
+ "description": "A standing, machine-readable benchmark of widely-repeated AI / data-science claims, each rebuilt as the smallest runnable test and ruled REPRODUCED / FAILED / NOT_COMPUTABLE. Honest, citable receipts for the field's folklore.",
+ "homepage": "https://dancenitra.github.io/agora/public/crucible/",
+ "repo": "https://github.com/DanceNitra/agora",
+ "license": "CC-BY-4.0 (data) / MIT (code)",
+ "schema": {
+  "key": "stable citation id (FI-NNNN), permanent per claim",
+  "domain": "replication | ai-claim",
+  "claim": "the widely-repeated claim, stated precisely",
+  "source": "where the claim comes from / who repeats it",
+  "verdict": "REPRODUCED | FAILED | NOT_COMPUTABLE",
+  "note": "honest summary tying the measured result to the verdict, with scope/caveats",
+  "lab_file": "repo-relative path to the runnable test (may be empty)",
+  "code_url": "GitHub URL to the runnable test if it resolves",
+  "code_resolves": "whether the runnable test file is present in the repo",
+  "date": "YYYY-MM-DD"
+ },
+ "counts": {
+  "total": 59,
+  "by_verdict": {
+   "FAILED": 12,
+   "NOT_COMPUTABLE": 15,
+   "REPRODUCED": 32
+  },
+  "by_domain": {
+   "replication": 51,
+   "ai-claim": 8
+  },
+  "runnable_code_present": 10
+ },
+ "entries": [
+  {
+   "domain": "replication",
+   "claim": "Real-world networks are scale-free: their degree distributions follow a power law p(k) ~ k^-alpha.",
+   "source": "Barabasi-Albert (1999) framing, widely repeated; rigorous tests: Clauset, Shalizi & Newman 2009; Broido & Clauset 2019 'Scale-Free Networks Are Rare'.",
+   "verdict": "FAILED",
+   "note": "Under a rigorous Clauset-Shalizi-Newman fit (MLE alpha, KS-selected xmin, bootstrap goodness-of-fit, Vuong likelihood-ratio vs lognormal, n=20,000): a lognormal that 'looks scale-free' on a log-log plot is correctly REFUSED (power-law GOF p=0.01; LR favors lognormal, -17.5); a genuine Pareto power law passes (GOF p=0.92; LR +102); and a real Barabasi-Albert network is only a TIE (LR -0.1) - power law is not even clearly preferred over lognormal for true preferential attachment. So 'looks scale-free' is not 'is scale-free' and the universal claim is not safely inferable (Broido-Clauset reproduces). Power law IS reproduced for true BA graphs.",
+   "lab_file": "agora_output/lab/20260619-190000_crucible-scale-free-networks-rare.py",
+   "code_url": "https://github.com/DanceNitra/agora/blob/main/agora_output/lab/20260619-190000_crucible-scale-free-networks-rare.py",
+   "code_resolves": true,
+   "date": "2026-06-19",
+   "key": "FI-0001"
+  },
+  {
+   "domain": "replication",
+   "claim": "Emergent abilities of large language models are genuine, sharp capability transitions - a discontinuous jump in skill above a scale threshold.",
+   "source": "Wei et al. 2022 'Emergent Abilities of Large Language Models' (claim); counter: Schaeffer, Miranda & Koyejo 2023 'Are Emergent Abilities a Mirage?' (NeurIPS).",
+   "verdict": "FAILED",
+   "note": "The canonical SHARP 'emergence' curve is reproduced by a SMOOTH, continuous per-token skill measured with a nonlinear exact-match metric: the same smooth skill gives a per-token transition width 7.0 vs an exact-match (L=100) width 1.05 - 6.7x sharper PURELY from the metric - and the apparent onset shifts -0.07 -> +5.58 as answer length grows with NO change in the underlying skill. No capability discontinuity is needed, so benchmark sharpness is not evidence for one (Schaeffer's 'mirage' reproduces). Scope: shows sharpness is metric-dependent; it does not prove that no ability is ever genuinely emergent.",
+   "lab_file": "agora_output/lab/20260619-185500_crucible-emergent-abilities-metric-mirage.py",
+   "code_url": "https://github.com/DanceNitra/agora/blob/main/agora_output/lab/20260619-185500_crucible-emergent-abilities-metric-mirage.py",
+   "code_resolves": true,
+   "date": "2026-06-19",
+   "key": "FI-0002"
+  },
+  {
+   "domain": "replication",
+   "claim": "Degree distribution of competition-induced preferential attachment graphs is a power law up to a finite threshold and decays exponentially above it",
+   "source": "Degree Distribution of Competition-Induced Preferential Attachment Graphs (Berger, Borgs, Chayes, DSouza, 2005)",
+   "verdict": "NOT_COMPUTABLE",
+   "note": "The smallest faithful model of the stated mechanism (1D positions; new vertex connects to earlier vertex minimizing |dx| + alpha*depth) DEGENERATES across alpha to a near-star (max-degree 849->4059 growing with alpha, mean-degree 2, poor po",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-18",
+   "key": "FI-0003"
+  },
+  {
+   "domain": "replication",
+   "claim": "A Pareto regret that reflects Pareto optimality without relying on scalarization functions can be defined and minimized in multi-objective bandits",
+   "source": "Pareto Regret Analyses in Multi-objective Multi-armed Bandit (Xu & Klabjan, 2022)",
+   "verdict": "REPRODUCED",
+   "note": "Smallest model: 2-objective bandit, known Pareto front (4 optimal + 2 dominated arms). Pareto-UCB1 with the scalarization-free Pareto suboptimality gap achieved sublinear cumulative Pareto regret (avg/step 0.027->0.005, growth factor 1.82 o",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-18",
+   "key": "FI-0004"
+  },
+  {
+   "domain": "replication",
+   "claim": "The epidemic threshold disappears for scale-free / preferential-attachment networks with degree exponent gamma in (2,3]",
+   "source": "Pastor-Satorras & Vespignani (2001); cited by Jones & Handcock (2003), preferential attachment in human sexual networks",
+   "verdict": "REPRODUCED",
+   "note": "Mechanism reproduced: SIS threshold lambda_c=<k>/<k^2> falls with N on BA/scale-free networks (0.112->0.062 as N 500->32000) because <k^2> grows with the hub/cutoff (35.5->64.9), while ER stays finite (~0.20, <k^2>~20). SIS dynamics confirm",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-18",
+   "key": "FI-0005"
+  },
+  {
+   "domain": "replication",
+   "claim": "Metcalfe's Law: the value of a communications network scales as n^2 (the number of possible pairwise connections)",
+   "source": "Metcalfe's Law (folklore/Gilder 1993); contested by Briscoe, Odlyzko & Tilly (2006), 'Metcalfe's Law is Wrong' (IEEE Spectrum), and Metcalfe (2013) Facebook dat",
+   "verdict": "FAILED",
+   "note": "Strong n^2 claim FAILS. Fitting alpha in V ~ n^alpha: the n^2 holds ONLY under the (unrealistic) assumption that every pairwise connection has EQUAL value (alpha=2.00). Under realistic rank-declining connection value (Zipf 1/k, the Briscoe-",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-18",
+   "key": "FI-0006"
+  },
+  {
+   "domain": "replication",
+   "claim": "Diversity trumps ability: a random/diverse group of problem solvers outperforms (>=) a group of the best individual solvers (Hong & Page 2004 PNAS)",
+   "source": "Hong & Page (2004), Groups of diverse problem solvers can outperform groups of high-ability problem solvers, PNAS 101(46)",
+   "verdict": "FAILED",
+   "note": "STRONG claim (random>=best) NOT reproduced in a faithful, paired, statistically-powered minimal model (n=800 ring, pool=120, group=10, random group averaged over 5 draws, 25 landscapes). Random-minus-best is consistently NEGATIVE across heu",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-17",
+   "key": "FI-0007"
+  },
+  {
+   "domain": "replication",
+   "claim": "Compactness and geometric stability conjectures for scalar curvature and convergence (Sormani et al., IAS 2018 workshop)",
+   "source": "Conjectures on Convergence and Scalar Curvature (Sormani et al., 2021)",
+   "verdict": "NOT_COMPUTABLE",
+   "note": "Open conjectures in geometric analysis (intrinsic-flat / Gromov-Hausdorff convergence under scalar-curvature bounds). No finite computational mechanism to reduce to a smallest model - these are deep manifold-convergence conjectures, not a s",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-17",
+   "key": "FI-0008"
+  },
+  {
+   "domain": "replication",
+   "claim": "The survival probability of a branching process obeys finite-size scaling in the control parameter (offspring mean m=1+eps) and the max generations n: P_n(eps) ~ (1/n)*F(eps*n).",
+   "source": "Garcia-Millan, Font-Clos et al. 2015, \"Finite-size scaling of survival probability in branching processes\"",
+   "verdict": "REPRODUCED",
+   "note": "Smallest Galton-Watson model (Poisson offspring), vectorized over 40k realizations. Critical eps=0: n*P_n -> ~2 (Kolmogorov 2/sigma^2, sigma^2=1). Scaling collapse confirmed: n*P_n is a function of x=eps*n alone - within-x spread 0.116 vs a",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-17",
+   "key": "FI-0009"
+  },
+  {
+   "domain": "replication",
+   "claim": "In the Erdos-Renyi graph G(N,p), the k-clique percolation transition occurs at p_c(k) = 1/[(k-1)N]^(1/(k-1)).",
+   "source": "Derenyi, Palla & Vicsek 2005, 'Clique Percolation in Random Networks' (PRL 94, 160202)",
+   "verdict": "REPRODUCED",
+   "note": "Smallest CPM model: enumerate k-cliques in G(N,p), union any sharing a (k-1)-clique, track largest community fraction R(p)=vertices_in_largest/N; empirical transition = p where R crosses half its max. Measured/formula ratio is near-CONSTANT",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-17",
+   "key": "FI-0010"
+  },
+  {
+   "domain": "replication",
+   "claim": "The epidemic threshold of a network disappears (goes to 0 as N->infinity) when the degree-distribution scaling exponent rho is in 2 < rho <= 3.",
+   "source": "Jones & Handcock 2003, 'An assessment of preferential attachment as a mechanism for human sexual network formation' (after Pastor-Satorras & Vespignani 2001)",
+   "verdict": "REPRODUCED",
+   "note": "Smallest model: HMF threshold lambda_c=<k>/<k^2> from sampled power-law degree sequences vs N. Measured lambda_c at N=1e3..1e6: gamma=2.3 shrinks x28.7, gamma=2.7 x3.6 (power-law vanishing); gamma=3.0 is MARGINAL - vanishes only logarithmic",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-17",
+   "key": "FI-0011"
+  },
+  {
+   "domain": "replication",
+   "claim": "Whether a mouse geroprotector is recorded as extending lifespan can depend on the survival statistic chosen (log-rank vs Gehan-Wilcoxon); and reporting the best of several survival tests inflates the",
+   "source": "deep-research 2026-06-16 + Jiang et al., GeroScience 2024 (Gehan reanalysis of NIA ITP data)",
+   "verdict": "REPRODUCED",
+   "note": "Self-contained sim (weighted log-rank family, n=50/arm, 4000 trials). Age-localized true effect: log-rank power 32.5% vs Gehan 72.3%; the two tests give DISCORDANT verdicts on 39.9% of identical datasets. Under the null, best-of-3 tests inf",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-16",
+   "key": "FI-0012"
+  },
+  {
+   "domain": "replication",
+   "claim": "Specimens were loaded with low-amplitude cyclic torque to analyze deformation in the small strain range (0.001-0.01%).",
+   "source": "Optical Flow Method for Measuring Deformation of Soil Specimen Subjected to Torsional Shear (Srokosz, Bujko, Bochenska)",
+   "verdict": "NOT_COMPUTABLE",
+   "note": "Not a falsifiable claim with a result - it is a methods sentence describing a physical geotechnical experiment (optical-flow measurement of a soil specimen under cyclic torsional shear). Off-domain (soil mechanics, not our frontier), no res",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-16",
+   "key": "FI-0013"
+  },
+  {
+   "domain": "replication",
+   "claim": "For BA networks (N=2,000, m=2), removing the top 10% of nodes by degree raises the bond-percolation threshold from p_c=0.174 to 0.776.",
+   "source": "Cachero Sanchez (2026), Simultaneous Degradation of Percolation and Cascade Robustness Under Targeted Hub Removal",
+   "verdict": "REPRODUCED",
+   "note": "Finite-size susceptibility-peak MC on the actual BA network reproduces it: intact p_c 0.170 (claimed 0.174), hub-removed 0.740 (claimed 0.776), both within ~5% over 6 replicas. Methodology lesson: the Molloy-Reed configuration-model estimat",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-16",
+   "key": "FI-0014"
+  },
+  {
+   "domain": "replication",
+   "claim": "Cohen's benchmarks (r=.10/.30/.50, d=.20/.50/.80) are used to interpret observed effect sizes; in gerontology observed effects fall below them, so the benchmarks are miscalibrated (Brydges 2019).",
+   "source": "Effect Size Guidelines, Sample Size Calculations, and Statistical Power in Gerontology (Brydges, 2019)",
+   "verdict": "NOT_COMPUTABLE",
+   "note": "Cohen's cutoffs are conventions, not empirical predictions, so they cannot be reproduced. Brydges' substantive claim (observed gerontology effect sizes fall below the benchmarks) is an empirical distributional claim requiring his meta-analy",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-16",
+   "key": "FI-0015"
+  },
+  {
+   "domain": "replication",
+   "claim": "Deep networks have enabled RL to scale to more complex domains, but these methods typically require large quantities of training data.",
+   "source": "Sample-Efficient RL with Maximum Entropy Mellowmax Episodic Control (Sarrico, Arulkumaran, Agostinelli et al.)",
+   "verdict": "NOT_COMPUTABLE",
+   "note": "Extracted sentence is a qualitative background premise (deep model-free RL is sample-inefficient) = a textbook fact, no crisp falsifiable mechanism; modelling it would re-derive known sample-inefficiency. The papers real contribution (Mello",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-15",
+   "key": "FI-0016"
+  },
+  {
+   "domain": "replication",
+   "claim": "In general, critical systems are associated with fractal / power-law behaviour",
+   "source": "(general statistical-physics statement)",
+   "verdict": "NOT_COMPUTABLE",
+   "note": "Too general to replicate: no single specific quantitative claim with a computable mechanism, and criticality<->scale-invariance/power-laws is foundational textbook physics where FAILED is NOT a live possibility - reproducing a known critica",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-15",
+   "key": "FI-0017"
+  },
+  {
+   "domain": "replication",
+   "claim": "Recursively synthesized, self-referential systems improve performance (e.g. Promptbreeder, Fernando et al. 2023: ~15pct few-shot gain via self-referential prompt evolution).",
+   "source": "Promptbreeder (Fernando, Banarse, Michalewski, Osindero, Rocktaschel, 2023, arXiv:2309.16797)",
+   "verdict": "REPRODUCED",
+   "note": "Smallest model (lab a08981): a population recursively synthesizes candidates FROM ITSELF each generation; the pivotal variable is the SELECTION signal. With an EXTERNAL fitness anchor (as Promptbreeder has: real task accuracy), recursive sy",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-15",
+   "key": "FI-0018"
+  },
+  {
+   "domain": "replication",
+   "claim": "Compactness and geometric stability conjectures for scalar curvature convergence (intrinsic flat / GH convergence of Riemannian manifolds)",
+   "source": "Sormani et al., Conjectures on Convergence and Scalar Curvature (2021, arXiv:2103.10093)",
+   "verdict": "NOT_COMPUTABLE",
+   "note": "A survey of OPEN conjectures in geometric analysis (convergence of Riemannian manifolds under scalar-curvature bounds, intrinsic flat distance). No single claim with a finite computable mechanism to model; the content is conjectural/abstrac",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-14",
+   "key": "FI-0019"
+  },
+  {
+   "domain": "replication",
+   "claim": "Finite-size scaling exists for the survival probability of a branching process as a function of the control parameter (m-1) and the maximum number of generations n: S_n ~ n^-1 G((m-1)*n).",
+   "source": "Garcia-Millan, Font-Clos & Corral (2015), Finite-size scaling of survival probability in branching processes (arXiv:1508.01515 / Phys. Rev. E 91, 042122).",
+   "verdict": "REPRODUCED",
+   "note": "Smallest model: exact survival curve of a Poisson(m) Galton-Watson process via PGF iteration q_n=exp(m(q_{n-1}-1)), S_n=1-q_n (zero Monte-Carlo noise). Three FSS signatures all confirmed: (1) exact critical amplitude n*S_n -> 2 = 2/sigma^2",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-14",
+   "key": "FI-0020"
+  },
+  {
+   "domain": "replication",
+   "claim": "In DiD with FEW treated units and spatially/serially-correlated errors, standard DiD inference under-covers (invalid CIs) while synthetic control is materially better when parallel trends ~holds",
+   "source": "Alvarez & Ferman (2020)",
+   "verdict": "REPRODUCED",
+   "note": "N=30,T=12,1 treated,rho=0.7,true effect=0,800 reps: DiD 95% CI coverage=0.305 (severe under-coverage vs nominal 0.95); SC coverage=0.891; SC RMSE=1.017 < DiD RMSE=1.267. DiD inference invalid, SC materially better.",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-14",
+   "key": "FI-0021"
+  },
+  {
+   "domain": "replication",
+   "claim": "p-hacking inflates Type I error in the error-statistical (Neyman-Pearson) approach but not in formal/likelihood inference that accounts for the selection",
+   "source": "Rubin (2026)",
+   "verdict": "REPRODUCED",
+   "note": "Under a true null (N=5000, n=30, K=5 forks): NP p-hacking (report min p) inflates Type I error 0.051->0.227 (+17.6pp, 4.45x, matches 1-.95^5=0.226). Formal/likelihood: a selection-ACCOUNTED likelihood (correct best-of-K sampling distributio",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-14",
+   "key": "FI-0022"
+  },
+  {
+   "domain": "replication",
+   "claim": "Context-aware AI augmentation (adaptive type/timing/scale) preserves cognitive flow (difficulty-skill match); poorly-timed/scaled interventions disrupt flow and hurt reasoning performance",
+   "source": "Dissanayake & Nanayakkara (2025)",
+   "verdict": "NOT_COMPUTABLE",
+   "note": "Design/HCI framework, not a falsifiable empirical claim: a minimal flow model (flow=difficulty-skill match) confirms a gap-closing intervention by construction, so the directional result is near-tautological — no genuine FAILED was possible. Downgraded from a constructed-model REPRODUCED to keep the ledger honest.",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-14",
+   "key": "FI-0023"
+  },
+  {
+   "domain": "replication",
+   "claim": "Removing top-10% degree nodes from a BA network (N=2000,m=2) raises the bond-percolation threshold p_c 0.174->0.776",
+   "source": "Simultaneous Degradation of Percolation and Cascade",
+   "verdict": "REPRODUCED",
+   "note": "Mechanism+direction reproduced: hub removal collapses <k^2> 50.8->3.8, p_c jumps ~8x (0.085->0.687 via Cohen mean-field). After-value within 12% of claim; before differs 2x (mean-field vs direct simulation). Robust-yet-fragile confirmed.",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-13",
+   "key": "FI-0024"
+  },
+  {
+   "domain": "replication",
+   "claim": "Derenyi-Palla-Vicsek (2005): k-clique percolation in ER graphs at p_c(k)=[(k-1)N]^(-1/(k-1)); for k=3, p_c ~ 1/sqrt(2N)",
+   "source": "Clique Percolation in Random Networks (Derenyi, Palla, Vicsek, 2005)",
+   "verdict": "REPRODUCED",
+   "note": "k=3 scaling exponent confirmed: empirical p_c*sqrt(2N) constant across N=400/800/1600 (1.26,1.26,1.19). N^(-1/2) scaling reproduced; prefactor ~1.2x asymptotic formula due to finite-size + 50%-coverage operational threshold.",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-13",
+   "key": "FI-0025"
+  },
+  {
+   "domain": "replication",
+   "claim": "Berger (2003): conditional-frequentist testing reconciles Fisher/Neyman/Jeffreys - the conditional error probability equals the Bayesian posterior; the naive p-value overstates evidence against H0",
+   "source": "Berger (2003), Could Fisher, Jeffreys and Neyman Have Agreed on Testing?",
+   "verdict": "REPRODUCED",
+   "note": "Calibration exact: empirical freq(H0 true | evidence) = Bayesian P(H0|x) across all bins. Berger-Sellke: p=0.05 -> P(H0|x)=0.216, p=0.005 -> 0.041 (p-value overstates evidence ~4x at 0.05).",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-13",
+   "key": "FI-0026"
+  },
+  {
+   "domain": "replication",
+   "claim": "Systems in the same universality class share the same critical exponents (Lubeck 2004)",
+   "source": "Universal Scaling Behavior of Non-Equilibrium Phase Transitions (Sven Lubeck, 2004)",
+   "verdict": "REPRODUCED",
+   "note": "Three structurally different Z2 mean-field models (tanh self-consistency, phi^4 free energy, arctan self-consistency) all give order-parameter exponent beta=0.500; a different-class absorbing-state model gives beta=1.000. Same class -> same",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-12",
+   "key": "FI-0027"
+  },
+  {
+   "domain": "replication",
+   "claim": "LinUCB (Chu et al 2011): linear contextual bandit achieves O(sqrt(Td log^3)) i.e. sublinear ~sqrt(T) regret w.p. 1-delta",
+   "source": "Contextual Bandits with Linear Payoff Functions (Chu, Li, Reyzin, Schapire, 2011)",
+   "verdict": "REPRODUCED",
+   "note": "Empirical regret growth exponent 0.03-0.11 (sub-sqrt(T), well inside the O(sqrt(T)) upper bound); cum regret 3-8 vs linear non-learning regret 1400-2700; d-scaling ~sqrt(d) to d. Computable core holds.",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-12",
+   "key": "FI-0028"
+  },
+  {
+   "domain": "replication",
+   "claim": "BA(N=2000,m=2): removing top-10% nodes by degree raises bond-percolation threshold p_c 0.174 -> 0.776 (~4.5x)",
+   "source": "Simultaneous Degradation of Percolation and Cascade Robustness Under Targeted Hub Removal (Cachero Sanchez, 2026-03-05)",
+   "verdict": "REPRODUCED",
+   "note": "MC before/after = 0.150/0.680 (ratio 4.53x) vs claimed ratio 4.46x; absolute values within finite-size tolerance. Cascade claim at phi=0.22 out of scope.",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-12",
+   "key": "FI-0029"
+  },
+  {
+   "domain": "replication",
+   "claim": "Dunning & Kruger (1999): the unskilled systematically overestimate their ability - a metacognitive deficit, shown by the quartile plot (bottom quartile +46pp, top -13pp)",
+   "source": "Kruger & Dunning, J Pers Soc Psychol 77(6), 1999 - Unskilled and Unaware of It",
+   "verdict": "FAILED",
+   "note": "A null model with ZERO metacognitive deficit (identical self-error at every skill level) reproduces the canonical plot AND its asymmetry: bottom +45.8 (DK: +46), top -14.2 (DK: -13) - just regression to the mean on a noisy test + a UNIFORM",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-12",
+   "key": "FI-0030"
+  },
+  {
+   "domain": "replication",
+   "claim": "Miller (1957): random typing (letters + space) produces Zipf law, so Zipf law in language carries no linguistic significance",
+   "source": "Miller, American Journal of Psychology 70, 1957 - Some effects of intermittent silence",
+   "verdict": "REPRODUCED",
+   "note": "Random-typing corpus (443k words): rank-frequency exponent -1.24, an approximate power law - the math reproduces. Severe test of a proposed counter (staircase fine-structure discriminator, tie fraction of adjacent ranks): 0.89 vs 0.48 at ma",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-12",
+   "key": "FI-0031"
+  },
+  {
+   "domain": "replication",
+   "claim": "Evans & Archer (1968) / textbook canon: a portfolio of ~20-30 stocks achieves practically complete diversification - marginal benefit beyond 30 is negligible",
+   "source": "Evans & Archer, Journal of Finance 23(4), 1968 - Diversification and the reduction of dispersion",
+   "verdict": "REPRODUCED",
+   "note": "Holds at realistic tails: N=30 captures 95.6-96.9% of achievable vol AND ES99 reduction (normal and t2.5 idio + 1-factor). MEASURED BOUNDARY: near infinite-variance idio tails (t1.8) the canon breaks - only 85.4% of tail-risk reduction capt",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-12",
+   "key": "FI-0032"
+  },
+  {
+   "domain": "replication",
+   "claim": "Social media platforms enable rapid dissemination of information but also facilitate the spread of misinformation",
+   "source": "internal corp finding (descriptive)",
+   "verdict": "NOT_COMPUTABLE",
+   "note": "Descriptive trend claim with no stated mechanism or quantity; the simulable core (false-news cascades spread farther/faster, Vosoughi 2018) is a DIFFERENT, specific claim - queued as a future Crucible target rather than strawmanned here.",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-12",
+   "key": "FI-0033"
+  },
+  {
+   "domain": "replication",
+   "claim": "Gilovich, Vallone & Tversky (1985): the basketball hot hand is a cognitive illusion - P(hit | streak of hits) is no higher than P(hit | streak of misses), so streak shooting does not exist",
+   "source": "Gilovich, Vallone & Tversky, Cognitive Psychology 17, 1985 - The hot hand in basketball: On the misperception of random sequences",
+   "verdict": "FAILED",
+   "note": "The canonical method manufactures its own conclusion. An iid shooter with NO hot hand shows P(hit|3H)-P(hit|3M) = -7.9pp (t=-28) under GVT estimator (n=100) - the streak-selection bias of Miller & Sanjurjo (2018). Bias grows to -17pp at k=4",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-12",
+   "key": "FI-0034"
+  },
+  {
+   "domain": "replication",
+   "claim": "Hong-Page (2004): a randomly selected (diverse) team of problem solvers outperforms the team of individually best solvers under relay search (n=2000, l=12, k=3, group 10)",
+   "source": "Hong & Page, PNAS 101(46), 2004 - Groups of diverse problem solvers can outperform groups of high-ability problem solvers",
+   "verdict": "REPRODUCED",
+   "note": "Reproduced at paper params: random beats best by +1.65 (t=4.1). But the effect is FRAGILE, as critics argued: it shrinks with a narrower heuristic pool (l=6: +0.90) and REVERSES on smoothed landscapes (-0.38, ability wins) - diversity trump",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-12",
+   "key": "FI-0035"
+  },
+  {
+   "domain": "replication",
+   "claim": "SGD locally converges to a minimum of a non-convex objective with a quantifiable rate",
+   "source": "Fehrman, Gess, Jentzen — Convergence rates for SGD for non-convex objective functions",
+   "verdict": "REPRODUCED",
+   "note": "Double-well (x^2-1)^2 (non-convex, minima +-1, max at 0). SGD noisy gradient, step c/(t+t0) started in the x*=1 basin: E[(x-1)^2] decays with fitted exponent -1.00 = theory O(1/t); constant step stalls at a variance floor. Local convergence",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-12",
+   "key": "FI-0036"
+  },
+  {
+   "domain": "replication",
+   "claim": "Generative AI integration into software engineering is rapidly expanding/automating tasks",
+   "source": "On agent-based software engineering (Jennings 2000)",
+   "verdict": "NOT_COMPUTABLE",
+   "note": "Garbled target: the claim is a malformed finding fragment (a vague expansion trend, no mechanism to model), and the cited source (Jennings 2000, agent-based SE) PREDATES and does not support a GenAI claim. No computable core; mismatched cit",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-12",
+   "key": "FI-0037"
+  },
+  {
+   "domain": "replication",
+   "claim": "Critical systems are associated with fractal or power-law scaling and long-range correlations",
+   "source": "Kitzbichler, Smith, Rahn (2009) Broadband Criticality of Human Brain Network Synchronization",
+   "verdict": "REPRODUCED",
+   "note": "Smallest model = critical branching process (neuronal avalanches). At sigma=1 avalanche-size P(s)~s^-1.42 (mean-field 3/2); off-critical the power law vanishes (sub: cutoff tau 2.45; super: 9.4pct runaway). Mean size diverges 6.7->1586 appr",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-11",
+   "key": "FI-0038"
+  },
+  {
+   "domain": "replication",
+   "claim": "In the low-data regime you can learn online policies with 2-10x fewer total coefficients with little to no loss of performance.",
+   "source": "Biologically inspired architectures for sample-efficient deep RL (Richemond, Kolbeinsson et al.)",
+   "verdict": "REPRODUCED",
+   "note": "Smallest model: supervised low-data fit (16 training points, smooth target, polynomial basis). Test MSE by coefficient count K: K=6 -> 0.015 (best); K=14 -> 3009 (catastrophic overfit). Shrinking from K=14 to K=6 = 2.3x fewer coefficients a",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-11",
+   "key": "FI-0039"
+  },
+  {
+   "domain": "replication",
+   "claim": "In scale-free networks the epidemic threshold vanishes as N grows (lambda_c = <k>/<k^2> -> 0 via the <k^2> divergence), unlike homogeneous networks which keep a finite threshold.",
+   "source": "Epidemic spreading in scale-free networks (Pastor-Satorras & Vespignani, 2001)",
+   "verdict": "REPRODUCED",
+   "note": "Smallest model: heterogeneous mean-field SIS threshold lambda_c=<k>/<k^2>, BA (scale-free) vs ER (homogeneous) at mean degree ~4. BA <k^2> grows with N (45->63 over N=500->20000) so lambda_c shrinks 0.088->0.064, heading to 0 = threshold VA",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-11",
+   "key": "FI-0040"
+  },
+  {
+   "domain": "replication",
+   "claim": "LinUCB-style linear contextual bandit achieves sublinear regret, O(sqrt(Td) polylog), holding with high probability (Chu, Li, Reyzin, Schapire 2011).",
+   "source": "Contextual Bandits with Linear Payoff Functions (Chu, Li, Reyzin, Schapire, 2011)",
+   "verdict": "REPRODUCED",
+   "note": "Smallest model: LinUCB on a d=6, K=12 linear contextual bandit, reward = x.theta*+noise. Cumulative regret stays ~flat (1.5->1.8 from T=500->4000, log-log exponent 0.09) - clearly SUBLINEAR and far inside the O(sqrt(dT))~155 bound at T=4000",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-11",
+   "key": "FI-0041"
+  },
+  {
+   "domain": "replication",
+   "claim": "SGD is popular for large-scale optimization but has slow asymptotic convergence due to the inherent variance of the stochastic gradient.",
+   "source": "Accelerating Stochastic Gradient Descent using Predictive Variance Reduction (Johnson & Zhang, 2013)",
+   "verdict": "REPRODUCED",
+   "note": "Smallest model: minimize a strongly-convex least-squares (scalar w) with label noise. Constant-step SGD stalls at a variance FLOOR (suboptimality ~1.9e-3, never reaches the optimum); full-gradient GD and variance-reduced SVRG both drive sub",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-11",
+   "key": "FI-0042"
+  },
+  {
+   "domain": "replication",
+   "claim": "Branching processes exhibit finite-size scaling of the survival probability P_n as a function of the control parameter m and the maximum number of generations n (critical n*P_n -> const, collapse in (",
+   "source": "Finite-size scaling of survival probability in branching processes (Garcia-Millan, Font-Clos, Corral, 2015)",
+   "verdict": "REPRODUCED",
+   "note": "Smallest model: Galton-Watson with Poisson(m) offspring, survival = population>0 at gen n. Both predictions reproduce: (a) at criticality m=1, n*P_n converges to ~1.9 (theory 2/sigma^2=2, sigma^2=1) -> P_n~C/n, exponent -1; (b) finite-size-",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-11",
+   "key": "FI-0043"
+  },
+  {
+   "domain": "replication",
+   "claim": "BA networks N=2000 m=2: removing the top 10% of nodes by degree raises the bond-percolation threshold from p_c=0.174 to 0.776 (severalfold robustness loss under targeted hub removal).",
+   "source": "Simultaneous Degradation of Percolation and Cascade Robustness Under Targeted Hub Removal (Cachero Sanchez, 2026-03-05)",
+   "verdict": "REPRODUCED",
+   "note": "Smallest model: BA(N=2000,m=2), Molloy-Reed p_c=<k>/(<k^2>-<k>) on the empirical degree sequence + a direct bond-percolation sweep (union-find, giant>50%). MECHANISM REPRODUCED: targeted top-10% hub removal raises the threshold several-fold",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-11",
+   "key": "FI-0044"
+  },
+  {
+   "domain": "replication",
+   "claim": "Brynjolfsson & Hitt (2000): organizational investments have a large influence on the value of IT investments (IT/org-capital complementarity).",
+   "source": "Beyond Computation: Information Technology, Organizational Transformation and Business Performance (Brynjolfsson & Hitt, 2000)",
+   "verdict": "REPRODUCED",
+   "note": "Smallest model: logY=a*IT+b*ORG+g*(IT*ORG). Recovered g~0.06; omitting org capital inflates apparent IT return, bias growing with co-investment corr (naive b_IT=0.21 vs true 0.12 at rho=0.8). Reproduces their methodological warning. Falsifi",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-11",
+   "key": "FI-0045"
+  },
+  {
+   "domain": "replication",
+   "claim": "Toutanova-Chen (2015): a simple observed-features model matches or beats latent-feature (embedding) models on knowledge-base completion",
+   "source": "Toutanova and Chen (2015)",
+   "verdict": "REPRODUCED",
+   "note": "On a synthetic KB with planted compositional structure (target relation = r0->r1 path), the OBSERVED-features model (2-hop path + degree) scored AUC 0.966 vs the LATENT SVD model 0.749 (+0.216). Path features capture composition that low-ra",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-11",
+   "key": "FI-0046"
+  },
+  {
+   "domain": "replication",
+   "claim": "Barato-Hinrichsen (2009): non-Markovian spreading with heavy-tailed waiting times t^{-1-m} shows an absorbing-state phase transition",
+   "source": "Barato and Hinrichsen (2009)",
+   "verdict": "REPRODUCED",
+   "note": "Minimal non-Markovian branching simulation (lab 37ad2a) reproduces the absorbing-state transition: survival ~0 for branching ratio lambda<1, rises for lambda>1, threshold lambda_c~1 IDENTICAL for m=0.5 and m=1.5. The heavy tail sets finite-",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-11",
+   "key": "FI-0047"
+  },
+  {
+   "domain": "replication",
+   "claim": "Safe-Sora (Su et al. 2025) addresses invisible generative watermarking for video by embedding graphical watermarks",
+   "source": "A Comprehensive Review (survey)",
+   "verdict": "NOT_COMPUTABLE",
+   "note": "Descriptive existence claim (a framework ADDRESSES an area), not a quantitative result - nothing to re-run as a minimal model. Replication needs a measured number (e.g. detection AUC, bit-error rate under compression), which this survey-sou",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-11",
+   "key": "FI-0048"
+  },
+  {
+   "domain": "replication",
+   "claim": "Duration of liquid-to-glass transition in a supercooled metallic alloy decreases exponentially with [truncated]",
+   "source": "Thermodynamic model for the glass transition (truncated)",
+   "verdict": "NOT_COMPUTABLE",
+   "note": "The claims dependent variable is truncated (decreases exponentially WITH WHAT is cut off) and it is flagged UNCERTAIN/unsettled by the literature - no precise quantitative relationship to re-run. Note: canonical supercooled relaxation is su",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-11",
+   "key": "FI-0049"
+  },
+  {
+   "domain": "replication",
+   "claim": "The AI-biology analogy breaks because brains use local Hebbian synaptic homeostasis",
+   "source": "internal frontier",
+   "verdict": "NOT_COMPUTABLE",
+   "note": "Already resolved by dialectic-where-ai-biology-analogy-breaks.md (breaks at the learning rule; backprop has no measured weight transport, learning survives random feedback). A verbal break-claim with no quantitative model to re-run.",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-10",
+   "key": "FI-0050"
+  },
+  {
+   "domain": "replication",
+   "claim": "A multi-agent RL model trained on historical Agora transaction data predicts trader behavior",
+   "source": "AGORA high-resolution galaxy simulations (mis-cited - name collision)",
+   "verdict": "NOT_COMPUTABLE",
+   "note": "No computable core: claim references internal/fictional Agora transaction data we do not have, and the cited source is a mis-attributed galaxy-simulation paper (AGORA name collision). Nothing real to replicate.",
+   "lab_file": "",
+   "code_url": "",
+   "code_resolves": false,
+   "date": "2026-06-10",
+   "key": "FI-0051"
+  },
+  {
+   "domain": "ai-claim",
+   "claim": "Multi-agent systems (an orchestrator delegating to sub-agents) outperform a single agent on complex tasks.",
+   "source": "2026 agent-engineering folklore (contested: Cognition 'Don't Build Multi-Agents' vs Anthropic's multi-agent research system); Gartner: ~40% of agentic projects scrapped.",
+   "verdict": "NOT_COMPUTABLE",
+   "note": "Tested at FIXED token cost on decomposable ground-truth tasks (one base model, only the scaffold differs; single self-consistency vs an orchestrator+workers+aggregator scaffold). Run 3x: the SAME config swings 0.15-0.20 in accuracy run-to-run on IDENTICAL tasks (multi@wmax1400: 0.75/0.75/0.92; single k=5: 0.96/0.75) at n=24, temp 0.7. The comparison is dominated by LLM-stochasticity noise at realistic eval scale -> the claim cannot be cleanly settled, AND the loud folklore (in both directions) rests on under-powered, non-reproducible comparisons. A stable verdict would need large-n multi-seed evals, which the folklore/blog benchmarks do not run. Honest scope: our own eval is also too small for a positive claim; the finding is the NON-REPRODUCIBILITY itself.",
+   "lab_file": "agora_output/aiclaims/20260619_multi_vs_single_agent_eval.py",
+   "code_url": "https://github.com/DanceNitra/agora/blob/main/agora_output/aiclaims/20260619_multi_vs_single_agent_eval.py",
+   "code_resolves": true,
+   "date": "2026-06-19",
+   "key": "FI-0052"
+  },
+  {
+   "domain": "ai-claim",
+   "claim": "When you evaluate N models/configs on a benchmark and report the top scorer's number, that score is a reliable estimate of its true performance and the winner is the truly-best model.",
+   "source": "Universal AI benchmarking / leaderboard practice (SOTA reporting, hyperparameter + model selection on a held-out benchmark).",
+   "verdict": "FAILED",
+   "note": "Winner's curse (selection-on-the-max), clean DETERMINISTIC model (no LLM noise - the lesson from entry #1): N models with true accuracies clustered within sigma_true=0.04, each measured with eval noise (finite test set + run-to-run variance). At N=50 models and eval-noise SE 0.06 (~ a 50-200 item benchmark): the reported winner's score is inflated by +0.112 (the SOTA bar is overstated), the observed winner is the truly-best model only 17% of the time, and trusting it costs 0.040 true accuracy vs the real best. Both effects GROW with the number of candidates and the eval noise (P(true best): N=5 -> 62%, N=100 -> 13%). So 'the leaderboard winner's score is reliable' FAILS: more candidates + noisier evals = more inflated, less trustworthy. (Ties to entry #1: the same eval noise that makes multi-vs-single non-reproducible is what drives this inflation.)",
+   "lab_file": "agora_output/aiclaims/20260619_benchmark_winners_curse.py",
+   "code_url": "https://github.com/DanceNitra/agora/blob/main/agora_output/aiclaims/20260619_benchmark_winners_curse.py",
+   "code_resolves": true,
+   "date": "2026-06-19",
+   "key": "FI-0053"
+  },
+  {
+   "domain": "ai-claim",
+   "claim": "Retrieval-augmented frontier models weigh a retrieved document against their own knowledge - they won't blindly adopt a doc that contradicts what they correctly know.",
+   "source": "RAG-robustness folklore; measured on real models via the Grounding Firewall poison protocol (Agora). The Poison-Deference Index.",
+   "verdict": "FAILED",
+   "note": "Poison-Deference Index: 12 factual questions each model answered CORRECTLY without context, then given a context asserting the WRONG answer (real LLMs, k=3 order-corrected, thinking-robust reader). deepseek-v4-flash: PDR=92% (flips to the false answer on 11/12 questions it knew), CPR=83% (confidently wrong). glm-5.2: PDR=92%, CPR=92%. BOTH frontier models abandon their correct knowledge for a planted-false doc ~92% of the time, almost always confidently - retrieval OVERRIDES rather than augments. (Strong-assertion poison; 2 models so far - a live index across more models is the next step. Mitigation already exists: the Grounding Firewall's abstain-on-high-sensitivity gate catches exactly these.)",
+   "lab_file": "agora_output/aiclaims/20260619_poison_deference_index.py",
+   "code_url": "https://github.com/DanceNitra/agora/blob/main/agora_output/aiclaims/20260619_poison_deference_index.py",
+   "code_resolves": true,
+   "date": "2026-06-19",
+   "key": "FI-0054"
+  },
+  {
+   "domain": "ai-claim",
+   "claim": "AI agent success decays with a CONSTANT hazard (an exponential 'half-life') as the task gets longer.",
+   "source": "arXiv 2505.05115 'Is there a half-life for the success rates of AI agents?'; tested on METR time-horizon data (metr.org/blog/2025-03-19; epoch.ai/benchmarks/metr-time-horizons).",
+   "verdict": "REPRODUCED",
+   "note": "Fit on METR's real public anchors (success vs human-task-length: ~99% @ 4min, ~80% @ 15min, 50% @ 60min [Claude 3.7], <10% @ 240min). A constant-hazard exponential (P = 0.5^(t/60min)) fits well: predicts 0.95/0.84/0.50/0.06 vs observed 0.99/0.80/0.50/0.08, SSE 0.0032 - actually BETTER than the logistic-in-log-time (SSE 0.0075, which overshoots the tail). The half-life claim holds. Minor tension: the t80/t50 ratio is slightly steeper than exponential (0.250 vs 0.322), hinting at mild extra steepening, but 4 anchor points can't resolve it. HONEST SCOPE: public anchors only (Claude 3.7 + aggregate); METR's raw per-task data would sharpen the per-model hazard.",
+   "lab_file": "agora_output/aiclaims/20260619_metr_horizon_survival_multiverse.py",
+   "code_url": "https://github.com/DanceNitra/agora/blob/main/agora_output/aiclaims/20260619_metr_horizon_survival_multiverse.py",
+   "code_resolves": true,
+   "date": "2026-06-19",
+   "key": "FI-0055"
+  },
+  {
+   "domain": "ai-claim",
+   "claim": "The 'AI time horizon' is a robust headline number (supporting 'AI will automate month-long tasks within ~5 years').",
+   "source": "METR time-horizon headline + 7-month-doubling extrapolation (metr.org/blog/2025-03-19); multiverse / specification-curve method (One Model Many Scores, arXiv 2308.16681).",
+   "verdict": "FAILED",
+   "note": "Multiverse over ONE analytic fork METR itself exposes - the success-threshold choice. From the fitted curve on METR's real anchors the horizon is 60 min at 50% success but 21 min at 80% (and 170 min at 20%): a 2.8x swing from an arbitrary threshold. At doubling-every-7-months that 2.8x is ~11 months of apparent 'progress' - so the famous 'month-long tasks in ~5 years' timeline slips ~11 months if you (reasonably) demand 80% reliability instead of 50%. The headline is NOT a robust single number; it rides on an unstated analytic choice. (A full specification curve over more forks needs METR's raw per-task data.)",
+   "lab_file": "agora_output/aiclaims/20260619_metr_horizon_survival_multiverse.py",
+   "code_url": "https://github.com/DanceNitra/agora/blob/main/agora_output/aiclaims/20260619_metr_horizon_survival_multiverse.py",
+   "code_resolves": true,
+   "date": "2026-06-19",
+   "key": "FI-0056"
+  },
+  {
+   "domain": "ai-claim",
+   "claim": "LLMs inherit human cognitive biases - e.g. conservatism in Bayesian belief updating (people under-revise relative to Bayes; Edwards 1968).",
+   "source": "Folklore that LLMs reproduce human judgment-and-decision biases; tested on the canonical bookbag-and-poker-chip conservatism task with real models.",
+   "verdict": "FAILED",
+   "note": "On the exact task where humans are reliably conservative (two equally-likely sources, symmetric cue validity q=0.70, an R/B signal sequence; Bayesian posterior ~0.97), both frontier models return the EXACT Bayesian posterior when the likelihoods are specified: deepseek-v4-flash mean conservatism gap = -0.001 (4/5 items parsed), glm-5.2 = +0.000 (5/5). Humans on the same task report ~0.70 (a large conservative gap); the LLMs show ~0.000. The bias does NOT transfer - handed the diagnosticity as numbers, the model just computes the likelihood ratio. HONEST SCOPE: only the SPECIFIED-likelihood case is settled. The human-analogue INFER case (model must ESTIMATE cue validity from a training sample, as humans do) is NOT cleanly measurable here - the harder estimate-then-update task exhausts the thinking models' token budget before a parseable answer (1/10 parsed), and forcing answer-first would suppress the reasoning the task needs. So: conservatism does not appear when cue validity is told; whether it appears when cue validity must be learned is open.",
+   "lab_file": "agora_output/lab/20260620-035500_llm-conservatism-cue-validity-specification.py",
+   "code_url": "https://github.com/DanceNitra/agora/blob/main/agora_output/lab/20260620-035500_llm-conservatism-cue-validity-specification.py",
+   "code_resolves": true,
+   "date": "2026-06-20",
+   "key": "FI-0057"
+  },
+  {
+   "domain": "ai-claim",
+   "claim": "Smaller chunks improve RAG retrieval quality - 'when in doubt, chunk smaller' raises precision/relevance.",
+   "source": "Common RAG-engineering folklore (chunk-size tuning advice; 'smaller chunks = higher precision').",
+   "verdict": "FAILED",
+   "note": "Deterministic numpy test: a 200-token document with one CONTIGUOUS gold span (length 30-50), fixed-grid chunking, each chunk scored by gold density (the precision force that rewards small chunks), retrieve top-k=3, measure recovery = gold tokens recovered / span length. Smaller is NOT better: the smallest chunk c=10 recovers only 0.750 of a 40-token span while c>=20 recovers 1.000 - the minimum chunk size is the WORST, and the optimum sits AT/ABOVE the span scale (best chunk = 20-40) in 9/9 robustness cells (seeds x span length). Mechanism: small chunks do maximize per-chunk density, but a grid cut fragments a contiguous span into more pieces than the top-k=3 budget can reassemble, and that recall loss dominates the precision gain. Anti-rig CONTROL: retrieving ALL chunks (no budget) flattens recovery to exactly 1.000 for every chunk size - proving the reversal is caused by the retrieval budget, not baked into the data. HONEST SCOPE: a noiseless density oracle, binary relevance, a single contiguous span, fixed grid boundaries - it isolates the fragmentation-vs-precision tradeoff, not embedding noise or overlapping/recursive chunkers. Conditional takeaway: 'smaller = better' is false whenever relevant evidence is contiguous and longer than a chunk under a tight top-k budget. (Generated + adversarially verified by an Agora workflow: 1 of 4 candidate claims survived 2 independent skeptic referees; the other 3 were killed as textbook/Condorcet or stipulated-geometry.)",
+   "lab_file": "agora_output/aiclaims/20260620_rag_chunk_size_not_monotone.py",
+   "code_url": "https://github.com/DanceNitra/agora/blob/main/agora_output/aiclaims/20260620_rag_chunk_size_not_monotone.py",
+   "code_resolves": true,
+   "date": "2026-06-20",
+   "key": "FI-0058"
+  },
+  {
+   "domain": "ai-claim",
+   "claim": "Adding a reranker (cross-encoder) on top of first-stage retrieval reliably improves end-to-end RAG accuracy, or at worst never hurts it ('drop in a reranker for a free boost').",
+   "source": "Cohere / Pinecone / LangChain / LlamaIndex RAG tutorials and 'production RAG checklist' blogs; repeated as received wisdom by practitioners.",
+   "verdict": "FAILED",
+   "note": "Deterministic numpy model (seed 0, n=200k queries): gold doc + 3 hard negatives (lexically similar) + 27 soft distractors; NO-RERANK uses a noisy first-stage scorer, RERANK uses a CLEANER scorer but inflates the 3 hard negatives by `infl` (a cross-encoder fooled by lexical overlap); end-to-end correctness depends on the gold doc's rank in a position-biased top-5. Measured delta (rerank - no-rerank): infl=0.0 -> +0.139, 1.4 -> +0.013, 2.2 -> -0.074, 3.0 -> -0.131. The reranker HELPS when it is merely cleaner (control infl=0: +0.139) but crosses to a NET LOSS at infl* ~ 1.52 and reaches -0.131 - so 'a reranker never hurts / free boost' is FAILED: a second scorer with a different error profile can demote the gold doc out of a position-biased top-k. Anti-rig: the control (infl=0) shows the harness CAN produce a gain (not baked to fail); the verdict is read off the printed curve. HONEST SCOPE: a stylized simulation, not real corpora/models - it shows the loss is mechanically in-range and the condition for it (non-trivial hard-negative susceptibility), not that any specific production reranker degrades a specific stack; the realistic magnitude of `infl` for real cross-encoders is the open empirical question. The defensible takeaway: 'always helps' is not a safe default.",
+   "lab_file": "agora_output/aiclaims/20260620_reranker_not_free_boost.py",
+   "code_url": "https://github.com/DanceNitra/agora/blob/main/agora_output/aiclaims/20260620_reranker_not_free_boost.py",
+   "code_resolves": true,
+   "date": "2026-06-20",
+   "key": "FI-0059"
+  }
+ ]
+}

folklore_index-0.1.0/folklore_index.egg-info/PKG-INFO

@@ -0,0 +1,23 @@
+Metadata-Version: 2.4
+Name: folklore-index
+Version: 0.1.0
+Summary: A runnable benchmark of widely-repeated AI / data-science claims, each ruled REPRODUCED / FAILED / NOT_COMPUTABLE.
+Author: Agora (autonomous research organization)
+License: CC-BY-4.0
+Project-URL: Homepage, https://dancenitra.github.io/agora/public/crucible/
+Project-URL: Source, https://github.com/DanceNitra/agora
+Keywords: ai,llm,evaluation,replication,reproducibility,benchmark,rag,agents,folklore,verification
+Classifier: Programming Language :: Python :: 3
+Classifier: Intended Audience :: Science/Research
+Requires-Python: >=3.8
+Description-Content-Type: text/markdown
+
+# folklore-index
+
+A standing, machine-readable benchmark of widely-repeated AI / data-science claims, each rebuilt as the smallest runnable test and ruled REPRODUCED / FAILED / NOT_COMPUTABLE. Honest, citable receipts for the field's folklore.
+
+```python
+import folklore_index as fi
+fi.verdicts()
+fi.get('FI-0001')
+```

folklore_index-0.1.0/folklore_index.egg-info/SOURCES.txt

@@ -0,0 +1,8 @@
+README.md
+pyproject.toml
+folklore_index/__init__.py
+folklore_index/folklore_index.json
+folklore_index.egg-info/PKG-INFO
+folklore_index.egg-info/SOURCES.txt
+folklore_index.egg-info/dependency_links.txt
+folklore_index.egg-info/top_level.txt

folklore_index-0.1.0/folklore_index.egg-info/dependency_links.txt

@@ -0,0 +1 @@
+

folklore_index-0.1.0/folklore_index.egg-info/top_level.txt

@@ -0,0 +1 @@
+folklore_index

folklore_index-0.1.0/pyproject.toml

@@ -0,0 +1,24 @@
+[build-system]
+requires = ["setuptools>=61"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "folklore-index"
+version = "0.1.0"
+description = "A runnable benchmark of widely-repeated AI / data-science claims, each ruled REPRODUCED / FAILED / NOT_COMPUTABLE."
+readme = "README.md"
+requires-python = ">=3.8"
+license = { text = "CC-BY-4.0" }
+keywords = ["ai", "llm", "evaluation", "replication", "reproducibility", "benchmark", "rag", "agents", "folklore", "verification"]
+authors = [{ name = "Agora (autonomous research organization)" }]
+classifiers = ["Programming Language :: Python :: 3", "Intended Audience :: Science/Research"]
+
+[project.urls]
+Homepage = "https://dancenitra.github.io/agora/public/crucible/"
+Source = "https://github.com/DanceNitra/agora"
+
+[tool.setuptools]
+packages = ["folklore_index"]
+
+[tool.setuptools.package-data]
+folklore_index = ["*.json"]

folklore_index-0.1.0/setup.cfg

@@ -0,0 +1,4 @@
+[egg_info]
+tag_build = 
+tag_date = 0
+