Review the methods.
Check the evidence.

1. Read the project brief, then scientific specification for the question, fixed review budgets and terminology.
2. Inspect preregistration, decisions and execution history. PanNuke's accepted primary analysis remains amended or exploratory; public Git history does not independently timestamp the PUMA freeze before its outcomes.
3. Check the separation of source groups and group-safe out-of-fold predictions. Patch separation alone does not establish patient/WSI independence. Final reference groups remain outside selection and tuning.

The natural-data action remains retain_uncorrected. CONFIRMATORY_COMPLETE has not been reached. Prospective workflow benefit and clinical utility remain unestablished.

Retained findings; follow the source links for designs, exact values and uncertainty.

PanNuke primary

Positive controlled ranking; adverse H4 downstream restoration. The accepted primary analysis remains amended or exploratory.View source

RIVA

Independent-expert disagreement precision increased by +4.95 percentage points versus matched random review at the 5% budget. Disagreement enrichment, not adjudicated error or clinical benefit.View source

MIDOG++

Independent-expert disagreement precision increased by +7.66 percentage points versus matched random review at the 5% budget. Disagreement enrichment, not adjudicated error or clinical benefit.View source

NuCLS independent pathologist

At 5% review, precision 0.333333 versus 0.215111 matched random. Limited to the qualifying JP.1 cohort across five patients.View source

NuCLS multi-rater

Frozen ranking rule failed; the guided downstream intervention was adverse. The retained primary claim is not supported.View source

MoNuSAC controlled

Retrieval improved, but downstream and class-safety gates failed. Action remains retain_uncorrected.View source

PUMA controlled transfer

All seven internally pre-specified gates passed. Controlled injected changes and flag_exclude training only; not natural-error detection. Public history does not independently establish pre-outcome freeze timing.View source

Interpret H4's interval correctly. It is the central 95% range across 100 random-review repetitions on a fixed reference set, not a group-bootstrap confidence interval. No adverse result has been removed.

The reviewer ZIP contains the article, key documentation, source authorities, compact NuCLS/MoNuSAC arrays and scoped verifiers. It needs no account, GPU, PyTorch or raw dataset. Open its index.html locally to read this guide.

Check integrity with Python 3.12

Compare the ZIP's SHA-256 with the release checksum, extract it, then run from its directory:

python -I scripts/review_project.py

This uses only the standard library. It checks the kit inventory, the thirteen-file article, upstream file identities and selected narrative contracts. It makes no network requests and performs no model training.

Get-FileHash aanca-reviewer-kit-v1.zip -Algorithm SHA256

sha256sum aanca-reviewer-kit-v1.zip

Optionally recalculate saved numeric results

python -m pip install -r requirements-reviewer.txt
python -I scripts/review_project.py --numeric --report ../aanca-review.json

NumPy is the only additional library. This independently recalculates the saved NuCLS multi-rater and MoNuSAC evidence, including matched controls and group bootstraps. Their not_supported conclusions remain visible. A failed check returns a nonzero exit code; the JSON receipt records its scope.

Use --online to explicitly compare the live article with the downloaded snapshot. An older snapshot can differ after an article update.

• IntegritySaved files match their declared identities and selected public claims are consistent.
• Numeric recalculationThe optional NumPy checks reconstruct selected results from saved arrays using project-maintained independent implementations.
• Software executionThe full repository supports a deterministic synthetic smoke workflow and complete Ubuntu/Windows test coverage. The small reviewer kit does not contain that environment.
• Source-image reproductionRequires lawful dataset access, frozen inputs and compute; it is not performed by the kit.
• New-data replication and clinical utilityRequire additional independent evidence. Passing verification is not a clinical validation.

For the larger PanNuke arrays, use the separate primary evidence release. Read the complete reviewer guide and reproducibility instructions for every supported scope and command.

Local copying and execution for non-clinical scientific evaluation are permitted under the limited evaluation licence. Other rights are reserved; third-party terms remain separate. See contributions and AI assistance and citation metadata.