Experiment manifests and artifact hashes
Identify exact input and output bytes without claiming unseen history or independent financial proof.
One manifest must name several producing objects
A reproducible run needs data, source, settings, engine environment, output identity and the policy linking them. A short model name cannot establish those objects. The public checksum map lists the synthetic bars, canonical settings, source launcher, full receipt and transaction CSV with digest and byte size. The receipt separately records data, source and settings hashes. These links make this fixture traceable but do not imply that an external auditor ran it. All inputs remain invented, and the replay carries explicit false fields for economic validation and production promotion.
Byte equality is narrower than semantic similarity
SHA-256 identifies the selected bytes. A change in line endings, whitespace or column order can change the digest even when a parser later produces similar values. Conversely, a matching file can still contain an exposed holdout, wrong financial premise or invented observations. The browser inspector reads raw bytes and compares the selected public reference artifact, not a guessed filename or a reserialized object. The checksum result should name its scope: equality with a publication reference, not broker authenticity, causal market evidence or a new engine execution.
A completed file comparison
An unchanged copy of the published CSV matches the declared bars digest. Adding one byte, altering a price or selecting the source-code reference for the same CSV produces a mismatch. The comparison is useful even when the file is not valid market data, because it answers an identity question. An unreadable, empty or overlarge local file has no valid comparison record and clears export. A mismatch can still be exported as a bounded observed digest result; it is not silently converted to a match or treated as a verified financial rejection.
Research exposure needs another history
A hash cannot prove when a researcher viewed a period or selected an exit. Access records, candidate decisions and effective policy versions are separate evidence. A correction changes artifact identities and should retain the prior run, including failures, while acknowledging any already exposed evaluation. The teaching packet has no untouched economic holdout, so its complete file map does not invent one. Actual model certification and managed-cloud route evidence also remain separate. A publicly hosted checksum map and downloadable launcher support method reproduction without creating independent consensus, a profitable forecast or an authority to transact.
Filled invented system case
A fictional reader compares the public bars file to the bars reference and obtains equal bytes. A second copy with an appended newline has a different digest even though a forgiving CSV reader might ignore it. The source-code reference selected for the original CSV also mismatches. The conclusion is raw-file identity only. No private file is uploaded, no engine runs in the browser, and no unseen holdout or actual financial result is inferred from either outcome.
| Object | Record | Bounded finding |
|---|---|---|
| Input identity | Synthetic CSV SHA-256 and size | Raw bytes |
| Run links | Source / data / settings IDs | Declared linkage |
| Altered copy | One added byte | Mismatch |
| Financial validation | Different evidence required | Not established by hash |

Primary references and scope
Expanded source review: 13 October 2026. Documentation supports the stated mechanism, not a tested account, hypothetical fill or provider outcome.