Skip to content

Latest commit

 

History

History
83 lines (77 loc) · 11.2 KB

File metadata and controls

83 lines (77 loc) · 11.2 KB

Provenance Ledger — freeze-before-run custody for every pre-registered claim

Every pre-registration in this repository is graded below by the strength of its timestamp custody, honestly. Three grades:

  • GIT-SEPARATED: the frozen specification was committed AND pushed to this public repository in a commit that predates the commit containing its results. Push events are server-timestamped by GitHub and cannot be backdated; verify any row with git log --pretty="%h %cI %s" -- PREREG_FRENCH49.md and the GitHub commit page (or the GitHub Events API archive).
  • SELF-ATTESTED: the specification document states it was written before the runs, but spec and results entered the repository in the same commit — the claim rests on the document's own narrative, not on custody. We say so plainly.
  • FORWARD (unresolved): the specification is committed and pushed and its test has not yet run — the strongest grade, pending resolution and pending third-party (OSF) mirroring.
Registration Frozen in commit (pushed) Results in commit Custody grade Verdict record
PREREG A (French 49, L1/L2 gates + dissociation) 69f007a 2026-07-15 (with results) 69f007a SELF-ATTESTED Addendum A: P1 confirmed, P3 refuted, P4 triggered
PREREG B (L3 drawdown labeler @15%) 69f007a (with results) 69f007a SELF-ATTESTED Addendum B: P5/P7 refuted, P6 confirmed
PREREG C (threshold sweep) 69f007a (with results) 69f007a SELF-ATTESTED Addendum C: P8 refuted
Addendum D (replay D1/D2, costs D3, A3′ D4, CVaR D5, seeds D6) c4345e9 2026-07-15T12:22-04:00 ac90d7e / 49a640e 14:15 GIT-SEPARATED Addendum F: D1 refuted-favorably, D2/D3 confirmed, D6 confirmed. Addendum I (2026-07-18): D4 fourth outcome (A3′ strictly worse than A3), D5 tail-reading refuted → CVaR interpretation demoted
Addendum E (X1/X2/X3/X4, withheld era E-F, regional register E-R) 5c96c35 2026-07-15T13:09-04:00 e1f804b / 49a640e 14:15 GIT-SEPARATED Addendum F: X1 neither-branch, X3 anti-branch, X4 confirmed + divergence, E-F 5/6 scorecard. X2 RESOLVED 2026-07-21 (results/e_french49_L3_x2.json + prewar era extension per commit 1155f3c): modern verdict WINDOW (13-17%@h0 through 16-22%@h6; registered 15/h0 inside; island clause dead), prewar verdict WINDOW at h4/h6 only with PX2e anchor-in-window MISS (10% isolated at h0, cell solid on tier-2 re-read); cross-era 13-16% overlap under hysteresis supersedes the relocation narrative (descriptive). E-R Japan resolved separately (Addendum L row)
CRSP forward test (osf_package/PREREG_CRSP.md) 40b69ad 2026-07-15T12:24-04:00; OSF-mirrored 2026-07-15 22:47 ET (osf.io/nsx4e) FORWARD, third-party custody (data not yet obtained; WRDS pending) open
NBER-labeler forward test (osf_package/PREREG_NBER_FRENCH49.md) 40b69ad 2026-07-15T12:24-04:00; OSF-mirrored 22:47 ET (osf.io/nsx4e) run 22:55–23:10 ET same day; results committed RESOLVED under third-party custody — the program's first temporal cell Addendum G: PN1 HIT (weak+strong, causal primary); PN2 confirmed; PN3 granularity branch; one strong-form miss in the non-causal robustness cell, disclosed. LORO/era supplement resolved 2026-07-16 (Addendum G supplement): drop-2020 and drop-08/09 positive but attenuated (drop-Jan-09 = 41% of headline), two-recession concentration and one negative-significant 2010s era block disclosed
PREREG H (banded-monolith control battery, adverse branch lodged) git 2026-07-16 (spec pushed before runs) results committed 2026-07-16 GIT-SEPARATED Addendum I: net edge stands under b≤2 bands (Holm ≤7.1e-17); registered ≥30% turnover-cut magnitude NOT met (4.1%/15.7%), disclosed; 62% L1 attenuation disclosed
Temporal-deployment test (osf_package/PREREG_TEMPORAL_DEPLOYMENT.md) git 2026-07-16; OSF-mirrored 2026-07-15 23:35 ET (osf.io/nsx4e) frozen diagnostics committed on lodging; outcomes scored 2027-04-07 FORWARD, third-party custody — measurement-split (the first cell whose diagnostic is computed on data ending before the outcome window begins) open
CRSP T-split amendment (osf_package/PREREG_CRSP_AMENDMENT_TSPLIT.md) git 2026-07-16; OSF-mirrored 2026-07-15 23:35 ET (osf.io/nsx4e) FORWARD, third-party custody (T-split primary design; net register and window estimator fixed pre-data; WRDS expected September) open. Window-estimator κ calibration attempted 2026-07-20 (results/window_estimator_calibration.json, calibration-not-claim): INFEASIBLE — modern needs κ∈(0.129,0.143], prewar κ∈(0.471,0.486]; disjoint, and directionally so (denser prewar crisis occupancy would push θ coarser while the actual paying window moved finer, 15%→10%). Per the amendment's own clause the infeasibility (and any successor rule) must be lodged at osf.io/nsx4e BEFORE CRSP data access; until then only the fixed-15% robustness column has standing
PREREG J (expanding-window baseline D7 + probe/dormancy sensitivity D8) c742c5c 2026-07-20T01:35-04:00 (spec pushed alone, before implementation existed) results committed 2026-07-20 (same day, after spec push) GIT-SEPARATED Addendum J verdicts: D7 middle branch PJ2 (p=0.30) won — A1e collects 73.0% of Γ, A1e−A9 n.s., A6>A1e Holm 1.8e-3; ICAIF class sentence narrowed to recency-driven policies (adverse consequence taken; PJ1 trigger scored honestly as not met). D8 PJ4 confirmed — Γ positive-significant in all 6 probe/dormancy cells; cumulative-saturation observation labeled post-hoc
Addendum L + supplement (E-R Japan operationalization; occupancy/audit clauses; X2 era extension) L: dfd9fe7 2026-07-20T16:04-04:00 (alone, before any run touched Japanese data); supplement: 1155f3c 16:14 (alone; custody stated inside: post-hoc for Japan, ex-ante for remaining regions and prewar-X2) results committed 2026-07-20 (after spec pushes) GIT-SEPARATED (Japan battery); supplement's remaining-region and prewar-X2 clauses FORWARD Addendum L verdicts: E-R Japan sign prediction HIT (Γ=−0.000602±0.000162, negative-significant; PL2 branch p=0.25); inversion Holm 3.9e-8; sign rule consistent in 4/4 Japan cells (running scorecard 9/10); cell OCCUPANCY-ANOMALOUS (67.2% crisis occupancy at 15%); structure screen descriptive fail (z=−0.90), fourth screen/pool-value dissociation
PREREG K (event-level inference K1, expwin 100-seed K2, top-k perturbation K3) a02b9f8 2026-07-20 (spec pushed alone, before the three scripts existed) results committed 2026-07-20 (after spec push) GIT-SEPARATED (K1 honestly scoped in the spec as analysis-plan-only: its inputs were already-released LORO γ values) Addendum K verdicts: K1 PK1a MISS (L3 sign test p=0.061) → binding adverse branch TAKEN, L3 seed CI relabeled implementation-precision-only, paper edit owed; L3 Wilcoxon and both NBER tests pass (PK1b/PK1c HIT). K2 PK2a/PK2b HIT — share 58.7%, A1e−A9 residual positive-significant at 100 seeds (+0.000366±0.000128), small-but-real-residual branch adopted. K3 PK3a/PK3b HIT — Γ sign and A6<A1 direction survive k=3 and k=10; no fragility clause triggered
Addendum O (pooled cross-history event-level register; dollar register economically primary) 851d060 2026-07-21T00:10-04:00 (spec alone, before the script existed; pooling rule fixed before any remaining region downloaded) results committed 2026-07-21 (after spec push) GIT-SEPARATED; future-region refreshes FORWARD Addendum O verdicts (checkpoint a, n=109): PO1 sign test p=0.0625 FAIL → frequency claim RETIRED program-wide (binding); PO2 Wilcoxon p=0.0086 pass (against 0.35 odds); PO3 dollar CI [−1881,+9977] covers 0 → economic headline restated region-conditional (binding). Japan re-read sanity exact
Addendum N (prewar event-level N1, prewar net-cost N2, κ-fallback N3) 7add739 2026-07-20T16:21-04:00 (spec alone, before the scripts existed) results committed 2026-07-20 (after spec push) GIT-SEPARATED; N3 fallback FORWARD for CRSP (θ estimator retired, fixed-15% primary; OSF mirror owed before WRDS access) Addendum N verdicts: N1 PN1a/PN1b BOTH MISS — prewar-10% sign test fails (Wilcoxon passes), 1958–89 fails both tests; binding adverse branch taken, both withheld-era positive cells relabeled implementation-precision-only at event level. N2 partial-miss branch in both cells: Γ_net@25bps positive-significant but below gross — D3 amplification refuted out-of-sample; deficit survives the adjudicating register

Notes:

  • The withheld-era battery (E-F) has git-separated custody for its specification (addendum E pushed before the runs began), but the 1926–1989 data itself has been publicly available for decades; the custody proves the spec was frozen before these runs, not that the era was unexaminable in principle. Both statements appear in the paper.
  • OSF custody established 2026-07-15 22:47 ET: project osf.io/nsx4e holds PREREG_FRENCH49_snapshot_2026-07-15.md, PREREG_CRSP.md, and PREREG_NBER_FRENCH49.md under third-party timestamps. All future registrations are OSF-first. (No self-hosted record can prove the absence of unreported frozen specs; the OSF mirror is the custody answer to that file-drawer question from this date forward.)
  • Independent check of any push time: GitHub's public event archive (GH Archive) records push events for public repositories.
  • 4H disjoint-era replication: attempted 2026-07-20, NOT RUN — data unfetchable from this environment, no addendum lodged. The vendored 4H panel ends 2024-12-31, so 2025–26 would be a clean disjoint span for re-testing the E0 4H screen pass (whose confirmation reused the screen's span — selection-on-maximum bias, disclosed). The source (Bybit v5 kline API, all official domains incl. api.bytick.com) is CloudFront-geo-blocked from this location; the free archive (public.bybit.com) carries 2025–26 only as tick-level trading files (~100 MB/day for BTCUSDT alone; tens of GB for the span) and its MT4 kline mirror is 15-minute bars ending 2024 with no DOGEUSDT. Per the standing rule (no silent substitution, no sketchy sources), nothing was lodged and nothing ran; the replication remains open for an environment with API access, and the lodgeable prediction sketch (screen pass replicates p≈0.55; Γ stays ≈0 p≈0.8) is recorded here as UNLODGED — it has no registration standing until committed alone before the data exists locally.

Erratum (2026-07-21) — E-X2 verdict text, prewar 9% neighbor

The E-X2 verdict entry in PREREG_FRENCH49.md ("E-X2 verdicts, both eras") describes the prewar 10%/h0 cell's 9% neighbor as "negative-significant". The authoritative result file (results/e_french49_prewar_L3_x2.json, tier-1 th9_h0: Γ = −0.000110 ± 0.000130) shows the cell is negative but not significant — its CI covers zero. Registration and verdict text in the append-only ledger cannot be edited post-hoc, so this note records the transcription slip instead; the paper follows the JSON (paper/main.tex states "both 1pp neighbors n.s."). No adjudicated verdict changes: under the lodged criterion the 9% cell enters no window either way, and the PX2e anchor-miss verdict rests on the "isolated positive" clause — neither 1pp neighbor is positive-significant — which holds under either reading.