Checklist and unproven

19 series, lags audited), sector ETFs (validated, inception respected), minute bars (A6), EDGAR delisted-exit registry (9,250 exits 1996-2026; data/cache/delisted_*). UNPROVEN - prices for dead US names before ~2015 are not obtainable from free sources (Yahoo 404s them, Stooq now requires JavaScript; only ~250 mostly post-2020 names recovered), so survivorship cannot be removed from free data; it is measured and quarantined (T20: analytic drag -4 bps/week vs +32 bps/week model edge; pre-1983 windows THIN). Needs a paid/CRSP-type source to close.

canon C54/C55): harness VALID (self-check: generaliser/memoriser/non-learner/leak all classified correctly). FIRST REAL RESULT (ld1, 12 windows, memory-bank learner): same-year disguised effect +0.23%/wk (CI +0.01..+0.54) but TRANSFER -0.03%/wk (CI -0.09..+0.02) -> verdict MEMORISATION (the memory recognises a year by its market-context fingerprint); in-band share HARMED -1.4 pt. Under the C56 past-only rule the loop's memory cannot see a year's own first run, so production is protected; the honest headline (transfer) is: the learning system does NOT yet generalise. Making transfer > 0 is now the main research objective.

>= 50% of weeks in the 5-10% band): UNPROVEN - target not reached after controlled testing; research continues (Bible Phase 45). Exact reason: the real system lands 22-23% of weeks in band (39 revealed windows, 2,049 weeks); no 1-25 stock volatility basket reaches 30% (proxy); the timeline dial (14%), volatility targeting (no gain), the basis search (nothing adopted) and six learners (band_cfg buys ~2 pt only with significantly worse risk) all fail to raise it without extra risk.

engine/edgar.fix_activist, resumable) finished 2026-09-29: kept 31,417 of 53,865 activist events; 2014+ verified (about 78% kept - the rest were filer-side misattributions), but verification found almost no matches for 2001-2013 (about 1-3 per year kept vs ~1,100-1,700 before), which points to a verification failure on older SEC filing pages, not to those events being wrong. The fixed file keeps only verified events (pre-2014 activist data effectively absent; backup data/cache/events_before_13dfix.parquet). Impact: nil today - ev_activist is excluded from the miner and the evidence weights; it must stay excluded until pre-2014 verification is fixed.

Bible Phase 45). Exact reason: movement IS predictable - picks touch +-10% within the week 57.9% end-to-end (CI 55.8-60.2%) vs 14.6% base, 72.6% among top picks (83-92% since 2020; a naive ATR ranking matches it); weeks with 10 candidates above a 95% threshold: 0.1%. With free daily data the 95% bar is not reachable at 10 picks/week.

Bible Phase 45). Exact reason: no direction signal on movers exists in the available data - B07 (stand-in inputs) 51.8%, Brier = base rate; V5 run: gate open 0/716 weeks; B25 (real literature-backed inputs on PREDICTED movers: post-earnings drift, insider clusters, reversal/momentum, 52-week high, miner score, event type, regime; ~970 accuracy/coverage cells, 2018-2026): 49-53% at full coverage vs 51.7% always-up, no cell reaches 80%, best lower bounds 0.55-0.65 (adjusted 0.45-0.59); planted-signal control recovered at 92-95% and shuffled labels read ~50%, so the pipeline could see a signal if one existed.

Bible Phase 45). Exact reason: it requires direction (V2), which has no measurable signal; the full pipeline therefore takes no bets (share10 lower bound 0.00). Movers-only: 15% of picks finish >= +10%.

owner decision). Also for the owner: live decides at 15:42 ET and fills before the close, while the simulations decide at the close and fill at the next open (execution parity).

scripts/livesim_loop2.py), Phase 25, self-tuning, Live retrain (engine/train.py), Live promotion path (engine/improve.py: challenger/promoted/rejected/rolled_back/confirmed), heavy tests, antimemo, exits, stops; log_experiment stamps code_hash/code_files/code_mixed, git, canon/Bible hashes, config hash, data snapshot, sha256 ids, and warns on any missing field. 8 superseded scripts archived in scripts/legacy (documented as not provenance-complete). Historic records stay as written (append-only).

one bundle per round), Phase 25 and self-tuning runners. First real bundle: state/checkpoints/loop2_round003_20260929T081401 - verify() ok (2026-09-29).

Live champion + Test basis v1, 19 registry metrics, whole-engine code hash, data snapshot); verify() ok, no drift. Unmeasured at freeze (no registry source): mover/pattern/analog/missed-winner/direction metrics.

point-in-time truncation with a NaN-aware control); real-cache parity 60 tickers x 12 dates, candles family strict-vs-batch + future-invariance + gappy data: 0 failures (state/research/parity/latest.json, 2026-09-29).

A13); bit-reproducible across runs and hash seeds (B11); ticker-name invariant (tests/integration disguise test); 154 miner tests.

B14); P(real)>0.9 bin 96-97% truly real in Phase 25 (exact truth).

B04; 677k rows 1973-2026, 2 seeds, rolling origin): verdict PASS on all 10 gates - rank IC 0.0317, Newey-West t 12.95, positive in every era (worst era 0.0035), beats the shuffled-outcome control, net spread > 0 after costs, seed agreement 1.0. Earlier single-cut results kept in the history (movement IC 0.38; direction ~0).

B03, 53 tests: name-keyed, hash-chained versions, earlier-windows-only reads, integrity verify). Real run (1,500 tickers, 458k weekly rows, six eras): verify ok, 0 audit problems; active patterns +0.10%/wk out of sample (t 2.31, optimistic), discarded ~0.

B04; state/research/pattern_movers/ablation_real_370892cce0386506.json): pattern inputs add mover AUC +0.0010..+0.0017 per era, but random/shuffled-pattern controls gain up to +0.0015, so deploy = False in every era; movers grid _ipat 85.8% vs _i 84.7% top-10 is within that control band. PatternMoverModel returns zeros unless its ablation passes.

after every weekday close + Saturday 60-day backfill) verified end to end 2026-09-29: GitHub run 36527933768 succeeded and uploaded 14 day files to release intraday-2026-09; scripts/sync_intraday.py pulled them back (7 new day files). Only completed sessions are written (6 tests; caught a time-comparison bug).

walk-forward coordinate search over the miner's own settings, fold-share + week-bootstrap adoption rule, untouched holdout; 7 tests incl. noise adopts nothing and a regime switch prefers short memory). Real run (movement target, 600 tickers, 2008+): EVALUATED - no change adopted (holdout IC 0.352 unchanged); fdr_q inert; shrink_k removed (inert under empirical Bayes).

macro lags fixed), heavy-tested on real data (B05: no-analog baseline beats every analog mode, skill -0.17/-0.18, p~1). EVALUATED - NOT ADOPTED: no edge.

+ pacing, admission, gaming guards; B09, 106 tests). Real replays (15 test windows): in-band 22.1% -> 14.3% with the dial, risk better; fails the tiered gate. EVALUATED - NOT ADOPTED; hook stays off (meta dial_on).

mistake kinds, hash-chained LessonStore; B06, 56 tests). Real data: lessons from other windows -0.19%/wk (90% CI -0.34..-0.02); 7 era triples beat the shuffled null in 2/7. EVALUATED - NOT ADOPTED.

45 windows): memorisation (same window vs renamed+shifted) = 0.000; positive control (RecallControl memoriser) +4.9%/wk on its window, 0 on the disguised rerun - flagged in 45/45, so the harness can see memorisation.

decay death rule + Bonferroni-corrected rescue at in-scope size, C43); regime patterns correctly rescoped; engine/pattern_lifecycle.py + bank integrity on real data (B03, 53 tests).

19 series, lags audited), sector ETFs (validated, inception respected), minute bars (A6), EDGAR delisted-exit registry (9,250 exits 1996-2026; data/cache/delisted_*). UNPROVEN - prices for dead US names before ~2015 are not obtainable from free sources (Yahoo 404s them, Stooq now requires JavaScript; only ~250 mostly post-2020 names recovered), so survivorship cannot be removed from free data; it is measured and quarantined (T20: analytic drag -4 bps/week vs +32 bps/week model edge; pre-1983 windows THIN). Needs a paid/CRSP-type source to close.

2026-09-29) - original seeds 100+ (FDR 6.7%, decaying 0% held, P(real)>0.9 bin 97% real) and fresh hold-out seeds 300+ (FDR 9.9%, decaying 0%, 96%); all 10 criteria pass.

Bible Phase 45). Exact reason: movement IS predictable - picks touch +-10% within the week 57.9% end-to-end (CI 55.8-60.2%) vs 14.6% base, 72.6% among top picks (83-92% since 2020; a naive ATR ranking matches it); weeks with 10 candidates above a 95% threshold: 0.1%. With free daily data the 95% bar is not reachable at 10 picks/week.

Bible Phase 45). Exact reason: no direction signal on movers exists in the available data - B07 (stand-in inputs) 51.8%, Brier = base rate; V5 run: gate open 0/716 weeks; B25 (real literature-backed inputs on PREDICTED movers: post-earnings drift, insider clusters, reversal/momentum, 52-week high, miner score, event type, regime; ~970 accuracy/coverage cells, 2018-2026): 49-53% at full coverage vs 51.7% always-up, no cell reaches 80%, best lower bounds 0.55-0.65 (adjusted 0.45-0.59); planted-signal control recovered at 92-95% and shuffled labels read ~50%, so the pipeline could see a signal if one existed.

versioned, drift monitor; B07, 44 tests). Real walk-forward: no type carries direction skill. EVALUATED - NOT ADOPTED (neutral weights).

B08). Real walk-forward (19,304 positions, 2005-2022): out-of-sample gain CIs straddle 0, chosen rule flips between folds; pattern-failure exits carry no information (z~1). EVALUATED - NOT ADOPTED: exits stay week-end.

B08, 51 tests). Stops cut P(loss>20%) 3.3% -> 1.2% but gaps leave a -73% worst loss; the conditional gap model is calibrated (5.2%/1.06% exceedance vs 5%/1%); holding the -20% cap at 1% needs dropping ~41% of positions (loss-cap filter).

Bible Phase 45). Exact reason: it requires direction (V2), which has no measurable signal; the full pipeline therefore takes no bets (share10 lower bound 0.00). Movers-only: 15% of picks finish >= +10%.

39 windows rebuilt with the move signal, insider data and opening prices; regen logs).

PID 26844, data/loop2_run5.log). Round 3 (2026-09-29): all gates OK on 3/3 windows, hardened feed, past-only basis lineage (3/3 on the untrained neutral basis - nothing trained precedes them yet), THIN window reported apart, reveal after the basis decision (1965, 1997, 1991), checkpoint bundle written. Performance targets are T3/T19.

>= 50% of weeks in the 5-10% band): UNPROVEN - target not reached after controlled testing; research continues (Bible Phase 45). Exact reason: the real system lands 22-23% of weeks in band (39 revealed windows, 2,049 weeks); no 1-25 stock volatility basket reaches 30% (proxy); the timeline dial (14%), volatility targeting (no gain), the basis search (nothing adopted) and six learners (band_cfg buys ~2 pt only with significantly worse risk) all fail to raise it without extra risk.

engine/adaptive.Adapter + memory; B15, 171 tests; deterministic across hash seeds). EVALUATED through the learning-delta harness: memory-bank learner transfer -0.03%/wk (memorisation), basis learner transfer -0.002%/wk - no demonstrated value on unseen years; B24 is searching for learners that transfer.

parity 0.0 on 4 windows × 8 days with insider data on).

engine/edgar.fix_activist, resumable) finished 2026-09-29: kept 31,417 of 53,865 activist events; 2014+ verified (about 78% kept - the rest were filer-side misattributions), but verification found almost no matches for 2001-2013 (about 1-3 per year kept vs ~1,100-1,700 before), which points to a verification failure on older SEC filing pages, not to those events being wrong. The fixed file keeps only verified events (pre-2014 activist data effectively absent; backup data/cache/events_before_13dfix.parquet). Impact: nil today - ev_activist is excluded from the miner and the evidence weights; it must stay excluded until pre-2014 verification is fixed.

B16, 29 tests); viewed live on GitHub Pages at phone width 390px on 2026-09-29 - renders, no horizontal page overflow (375<=390), no console errors except a missing favicon; now shows the long-term bank (11 hash-chained versions from the newest research bank until a production bank exists). Content refreshes each rebuild.

28 settings, 106 values; 2 pass |t|>=3.5; losing values shown); viewed live at phone width 390px on 2026-09-29 - renders, no horizontal overflow. Also dashboard.html (safety-audited: it refused to publish while planted test secrets leaked in via pytest's cache - fixed).

2026-09-28): re-tester OK on 3/3 windows under the provenance/stale-code gate; engine/retester.py (0.5% relative, code hash first).

3/3, 2026-09-28); tests/test_session.py peeking-rule control fails it as required; blind_gates look-ahead probe 6/6 honest pass + 6/6 peeker caught (state/research/blind_gates/results.json).

tested); pit.Guard with hash-chained audit log (55 tests); features.build future-invariant on real data (5 cuts, 107k rows) and real-cache parity 60x12 with future-invariance per family: 0 failures. (The WIDER future-knowledge audit of C56 is T20.

supervise, heartbeat, exclusion report); on real data it correctly EXCLUDED 3 windows whose results held NaN (cause fixed, 8ef3835).

B15); final report reproducibility audit PASS (state/reports/final_report.json).

VALIDATED on original and hold-out seeds).

C33): round 2 proved all 490 fills at the next open (164+142+184) via engine/fill_audit.gate -> pit.audit_fills; 5 planted defects caught in tests.

canon C54/C55): harness VALID (self-check: generaliser/memoriser/non-learner/leak all classified correctly). FIRST REAL RESULT (ld1, 12 windows, memory-bank learner): same-year disguised effect +0.23%/wk (CI +0.01..+0.54) but TRANSFER -0.03%/wk (CI -0.09..+0.02) -> verdict MEMORISATION (the memory recognises a year by its market-context fingerprint); in-band share HARMED -1.4 pt. Under the C56 past-only rule the loop's memory cannot see a year's own first run, so production is protected; the honest headline (transfer) is: the learning system does NOT yet generalise. Making transfer > 0 is now the main research objective.

canon C56): 12 channels audited on real caches (engine/leak_audit.py, 52 tests); fixes APPLIED and live: hardened feed, NetworkGuard, crypto removed before disguise, past-only BasisLineage from neutral defaults, split-invariant tradable filter (Test path), THIN windows apart, learning headline = transfer delta; 210 affected tests pass; round 3 ran clean on it. Quarantined (documented, cannot be removed from free data): survivorship (pre-1983 thin), year fingerprints from levels, macro revisions unmeasurable. Owner question pending: C54 same-year reruns vs C56.

owner decision). Also for the owner: live decides at 15:42 ET and fills before the close, while the simulations decide at the close and fill at the next open (execution parity).

engine/broker.py) and tick; tests/test_live_safety.py + audit: paper-only clients broker.py True, tick.py True; 0 credential reads outside the live side (state/research/data_live_safety/report.md).

holidays + half days) since 2026-09-29 - it had allowed 1,090 closed half-hour slots; tests/test_broker_hours.py (5) + tests/test_live_safety.py (the audit catches a planted weekday-only rule; real broker 0 unsafe slots); scheduler (engine/tick.py) screens non-sessions too.

state/research/data_live_safety/report.md): the one defect found (holidays/half days allowed) FIXED 2026-09-29 with tests/test_broker_hours.py; the one research->Live write path (improve.test_and_promote, Bonferroni-gated) recorded as a reviewed exception, audit fails on any second writer.

research never imports the live order path; engine/explain.py split out); tests/test_isolation.py pass; CI runs the gate on every push.