Provenance, verification, and replication
The published metrics are bound to a local immutable evidence chain. Verification checks that chain; replication creates a new study identity and tests whether the result recurs.
Verification and replication are related but different.
Verification confirms that the report, metrics, claim decisions, and their bound inputs are byte-for-byte identical to the frozen closeout record. It does not rerun a model, reopen sealed labels, or reinterpret a threshold.
Replication reruns the scientific procedure as a new experiment. It may confirm, weaken, or contradict the published effect, but it produces new artifacts and must receive a new study identity. The public GitHub repository exists for this second purpose; it is not the source of the numbers in this manuscript.
9.2 · Frozen identities
A result with a verifiable chain.
Each stage was authorized, executed, frozen, and hashed before the next stage could make a scientific decision.
- Study-closeout SHA-256
f9ee482d973fca5522940bf35a168333f6b12864e614d185b29201a709ed165f- Closeout Merkle root
126f7a88f3ad31154cb552d6bd829109a2b5f1548daf1438930bf4b3543618b5- Claim-decision freeze
2c14678accac56e6708ad12813314aad835cd72a02e240111b8490b8086f34a5- Aggregate freeze v4
e9b3c8bd9007ecc52e46ad985065acfff4b1d22d7530bf67d228c964cc1c972a
$opmTests = (Get-ChildItem tests -Filter 'test_opm_*.py').FullName
& .venv\Scripts\python.exe -m pytest @opmTests -q
& .venv\Scripts\python.exe -m scripts.opm_study_closeout_executor verify `
--workspace . --freeze <path-to-closeout-freeze> `
--freeze-sha256 f9ee482d973fca5522940bf35a168333f6b12864e614d185b29201a709ed165fFrom authorized training to immutable closeout.
Short hashes are display labels; the complete identities remain in the frozen local record.
- 01Training transition
Authorized v1.1.4 training execution
ec597e19…5b5c4 - 02Pilot matrix
24 completed pilot runs and frozen selection
ad90a109…d5be - 03Primary matrix
20 completed primary runs and selected checkpoints
79107e31…dc82 - 04Stage 1 freeze
Label-blind predictions, interventions, and probes
5af3c124…5d8ba - 05Aggregate v4 freeze
Authoritative sealed metrics
e9b3c8bd…1c972a - 06Claim-decision freeze
Frozen thresholds and statuses
2c14678a…f34a5 - 07Study closeout
Final report package identity
f9ee482d…165f
Closeout generated narrative, not new science.
No checkpoint, model, prediction row, sealed label, or raw representation was read during study closeout. No training, prediction generation, aggregation, threshold application, threshold change, status change, or new scientific decision occurred. Formal statuses were reproduced without modification from the claim-decision freeze.
Corrections remain visible rather than disappearing into the final result.
Predecessor artifacts and failed attempts are provenance. They are not interchangeable with the authoritative evidence.
Introduced corrected fresh data and leakage gates while preserving v1.0 as immutable.
Corrected a binding-schedule arithmetic defect before protocol freeze.
Corrected a self-endpoint corruption edge case before protocol freeze.
Corrected construction coupling for evidence positions; generated fresh artifacts.
Resolved source snapshot, sampler, and equal-budget pilot-selection ambiguities.
Failed closed or were superseded; excluded from every formal decision.
Run the same study. Test the result independently.
The public GitHub repository is the hands-on workspace for researchers and builders. The frozen local evidence remains the sole source of the metrics published in this manuscript.