# Human review and release record

Blank template. It records no review or release approval. Do not prefill reviewer identities or check items without actual evidence. A check report is not an acceptance decision.

## Exact package under review

- Submission ID, track and version:
- Repository commit or immutable artifact version:
- Manifest SHA-256:
- Records SHA-256:
- Checker version and rerun date:
- Automated report and actual records checked:
- Coverage: complete review package / sample only (sample is not a full pass):
- Unchecked source data, external artifacts, protected evaluation or other limits:

The reviewer reruns the current trusted checker on the actual files. Both automated stages must pass before formal package review. Resolve sample-only coverage first. A resource-only pass enters resource review, not RSI task validation. Any package change requires a fresh check and renewed review of affected evidence; approval is bound to these exact hashes and version.

## Scientific review (independent of the contributor)

- Actual reviewer, role, date and conflicts; pending if unassigned:
- Source authenticity, original citations and scientific provenance:
- Units, labels, exclusions, missing/failed data, physical/statistical validity:
- Scientific use, value and documented limitations; benchmark difficulty where applicable:
- For benchmark tasks: whether the scoring rule, checker and feedback measure the intended scientific behavior:
- Family relationships, contamination/exposure and independence of any proposed tests; explicitly record if no reserved evaluation set exists:
- For RSI tasks: replayable baseline, allowed edits, improver installation and subsequent-use mechanism, fixed judge, meaningful controls and full budget accounting:
- Actual replay or scientific tests performed, evidence and limitations (planned checks remain planned):
- Decision: pending / needs revision / reject / approve scientific scope:

## Engineering and release review

- Actual reviewer, role, date and conflicts:
- Reproduction instructions and supported environment verified:
- Resource usage card: immutable access route/checksum, readable record and loading example, expected output and citation. For benchmark tasks, also verify prediction/checker interfaces and feedback boundaries:
- Clean-download replay by someone other than the contributor: environment, exact commands, actual result and limits (pending until run):
- Contributed executables reviewed before any isolated execution; no automatic execution by the static gate:
- Original resource license/access verified; redistribution versus link-only conditions:
- Dataset host, versioned landing/download or access URL, last actual access check and outcome:
- Contributor consent for public artifacts, names and affiliations:
- Public/private artifact list; if protected assessment data exists, verify its storage and access owner. Otherwise record that no reserved evaluation set exists; a resource need not propose future custody:
- Citation instructions for original paper/resources and adapted package:
- Original-source citations and confirmed contribution roles:
- Exact public artifacts or source links, version and restrictions:
- Decision: pending / needs revision / reject / approve release scope:

## Evidence for a future quality profile

For each of the five criteria below, record status (`not_assessed`, `contributor_documented`, `maintainer_checked`, `independently_verified`, or `not_applicable` with a reason), exact version, evidence location, reviewer, date and checked scope. An automated result must identify the rule and files covered. Leave unperformed checks unassessed.

- Source, version and reuse/access terms:
- Scientific validity and data integrity (labels, units, missingness and uncertainty):
- Reproducible access and use:
- Available checks and useful development feedback:
- Independent evaluation and leakage/exposure limits:

These findings support a planned quality profile; this template assigns no numeric score. A documented limit can coexist with a useful, approved scope. Ordinary datasets do not require an RSI runner or a reserved test set. Unresolved release permissions still block redistribution regardless of other findings.

## Maintainer decision

- Actual maintainer and dated decision:
- Both reviews complete, conflicts addressed and outstanding issues resolved:
- Approved immutable version and the two file hashes confirmed:
- Decision: pending / approved for a specified release / not approved:
- Public page and release/version link, filled only after publication:

Record scientific-resource acceptance, RSI-task validation and public release separately. A candidate resource cannot be relabeled as RSI-ready on the strength of its metadata check. Valid negative research results are eligible; no recursive gain is required. The automatic checker never fills or approves this record. A merged PR alone does not establish scientific acceptance or publication.