spb/anomaly-atlas Public License

expH (validation 2016-2021, opened once; holdout stays sealed):
- negative finding REPLICATES: 0 net survivors at full half-spread OOS
- CKX fails pre-registered sub-period clause (kappa* 0.445->0.263)
- ES->SPY flips sign OOS (+0.003, p=0.68) — regime-fragile
- daily reversal family: VR30 excess -0.143 -> +0.023 (crisis-regime)
- figure: alphas collapse OOS, the SPX staleness artifact persists

Atlas (via new_finding.py, Level 2 each):
- F001 nothing in the searched universe survives the full ladder
- F002 SPX->SPY lead is index content staleness (Fisher), survives SPA

Publications: new tab + P001 'The Artifact Frontier, Part I' — formal
write-up of expA->expG with live {{figure:...}} tokens rendered from
committed results.json

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

- net = gross - kappa*(EDGE/2)*turnover swept over kappa in {0,.1,.25,.5,1,2}
- median breakeven kappa* = 0.0114 (max 0.28): the median rule captures ~1%
  of one half-spread per trade; ES->SPY kappa*=0.0028 splice-invariant
- survivors: 3 at kappa=0.1 (all sparse names), 1 at 0.25 (CKX 1day,
  skeptical prior), 0 at 0.5+; zero intraday at kappa=1 — pre-registered
  falsification did not trigger; three-layer doctrine closes
- results sanitized (NaN -> null: Python json emits bare NaN, invalid for
  every other consumer); kappa* log-scale figure with cost reference lines

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

- web/lib/charts.js: expB staleness scatters, expC VR30-excess strip plot
  (FDR emphasis), expD raw-vs-fresh dumbbells; palette validated
  (validate_palette.js all-pairs PASS on card surface), native tooltips,
  ink-token text, selective direct labels
- figures injected on experiment pages + latest figure on home
- rendered and eyeballed (rsvg): fixed tick-label collision

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

- pre-specified EDGE+Roll intraday bounce null predicts impossible AC1
  (-3..-27): discarded its positive-excess family as null artifact;
  variance-consistency rule adopted (taxonomy T1 amended)
- corrected MA(1)-consistent VR triage: 14 Level-0 cells (JPM 1min -0.243,
  MSFT 5min 2000-07...) + daily 2008-15 mega-cap/index reversal family
  (XOM -0.133 FDR, SPY -0.055 FDR)
- H01 confirmed (liquid intraday lag-1 reversion = 0 net of bounce);
  decay clause partially falsified — recorded as-is
- new: stats/multiple_testing.py (BH-FDR, bootstrap p), EDGE wrapper,
  pre-specified splits in data/universe.py, 6 new gate tests

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

- synthetic generators with known ground truth + 11-test gate; the gate
  caught a real VR estimator bug (double division by q) before real data
- stats: reversion (VR/AC1/half-life), leadlag, block bootstrap;
  validation: Roll spread, excess reversion, staleness, LOCF
- expB (pre-specified protocol): 31 tickers Q1 2024 RTH 1min —
  bounce AC1 -0.009/-0.051/-0.232 by staleness tercile, VR30 down to 0.55
  with zero economics, SPY spuriously leads stale names (+0.047 at +1min,
  Spearman +0.43), SPX lags SPY by 1min (+0.065)
- artifact_taxonomy.md: 7 entries T1-T7 with measured magnitudes

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

- hf_client.py: cache-first single API client (throttle, 429/5xx back-off,
  50k-row pagination, committed data-manifest index) + 5 offline unit tests
- expA benchmark: 8 probe families, results with embedded hardware manifest
- research/data_source_profile.md (reviewed): coverage, timestamps ET,
  session structure, daily-vs-1min semantics, adjustment re-basing, limits
- LOG entry; expA hypothesis/analysis finalized

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>