Some documents under docs/ are part of the preregistration record and are
covered by the preregister-tier0-v2.2 freeze manifest. They describe the state
of the world as it was committed before the reportable campaign executed, and
they still point at results/final/ — the v1 output path — in places.
They have deliberately not been updated, and they will not be. Rewriting a preregistered hypothesis or interpretation document after seeing the results would destroy exactly the evidence the freeze exists to provide. A preregistration you edit afterwards is not a preregistration.
This note is not frozen, and exists so a reviewer can read the frozen documents correctly without them being altered.
| Document | Freeze role |
|---|---|
docs/EXPERIMENT_PROTOCOL.md |
hypotheses_and_primary_outcomes |
docs/CLAIM_TO_ARTIFACT_MATRIX.md |
provenance_and_interpretation |
docs/RESULT_INTERPRETATION.md |
provenance_and_interpretation |
docs/PAPER_REQUIREMENTS.md |
provenance_and_interpretation |
docs/FIXTURE_PROVENANCE.md |
provenance_and_interpretation |
docs/CASE_B_LDREA_TRACEABILITY.md |
case_b_ldrea_mapping |
docs/ARTIFACT_RUNS_COMPLIANCE.md |
artifact_runs_compliance |
docs/REPRODUCIBILITY.md is not frozen and has been kept current.
results/final/ in a frozen document means results/final_v2/The frozen documents were written when results/final/ was the reportable output
path. The authoritative campaign is now results/final_v2/. Every result file
named in those documents exists under results/final_v2/ with the same name.
results/final/ still exists, holds the superseded v1 campaign, and is
retained unedited for provenance. It is not reportable: its freeze tag existed only
locally when that campaign ran. See results/V1_V2_COMPARISON.md.
EXPERIMENT_PROTOCOL.md and CLAIM_TO_ARTIFACT_MATRIX.md specify “≥ 30 runs per
case, ≥ 60 total”. That is a preregistered minimum, not a description of what
ran. The v2 campaign ran 31 per case, 62 total, which satisfies it.
The specific figure TD = 1.000 (60/60) in RESULT_INTERPRETATION.md §on
cross-platform reproduction is the v1 measurement. The reportable value is
TD = 1.000 (62/62), in results/final_v2/determinism_summary.json.
preregister-tier0-v1 in a frozen documentWhere a frozen document names the freeze tag or FREEZE_MANIFEST.sha256, the
governing freeze for the reportable campaign is preregister-tier0-v2.2 at
commit c44f25d6fdb67e0bc4ac73a6217125dec8da1c0e, with manifest
preregistration/FREEZE_MANIFEST_V2.sha256 (133 files).
Their substance is unaffected by any of the above. In particular:
EXPERIMENT_PROTOCOL.md’s instruction that the supportable wording is
“deterministic across the tested supported environments” and never
“platform independent” is the rule the published claim was held to, and it
still governs.RESULT_INTERPRETATION.md’s per-metric statements of what each number means and
does not mean apply unchanged to the v2 values.FIXTURE_PROVENANCE.md’s 29 fixture entries — including FP-014 (the three Case A
non-runtime ratings) and FP-020 (the author-specified Monte Carlo distributions)
— describe the same committed inputs, which are byte-identical in v1 and v2.CASE_B_LDREA_TRACEABILITY.md describes the same machine-verified mapping that
freeze_check.py re-verifies on every run.For the current, machine-verified mapping from every reported number to its
evidence file and exact JSON field, use
artifact_review/PAPER_TO_ARTIFACT_RESULTS.md. It is generated from
results/final_v2/ by python -m tools.build_paper_result_map, and
tools/verify_reported_results.py checks it on every run.
That the frozen documents are unmodified is checkable:
python tools/freeze_check.py --final-v2 # includes frozen-file verification
experiments/verify_freeze.py hashes each frozen file at the freeze commit and
compares against FREEZE_MANIFEST_V2.sha256, then confirms no frozen file changed
between the freeze and the campaign. results/final_v2/PROVENANCE.json records
frozen_files_changed_since_freeze: {} and results_commits: ["c44f25d6…"] — the
campaign ran at the freeze commit itself.