Validation dossier (#25)

[NingyuSUN]

The last item of #25: docs/VALIDATION_DOSSIER.md, structured after the seven steps of FDA's draft guidance Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products (January 2025). It borrows the structure only: it is not a regulatory submission and not for clinical use.

The seven steps:

1. Question of interest: can LLM- and agent-produced curation records (single-cell cell types, literature evidence) be admitted to a research knowledge base without expert review, and which must an expert see?

2. Context of use: the model and bioevidence together; research summaries and a research knowledge base. Clinical and regulatory use are excluded.

3. Model risk: influence is high for admitted records and consequence is moderate, so the risk is medium.

4. Credibility plan: pinned sources, reference standards, splits, protocol freezes with their commit hashes, metrics, negative controls and reproducibility. It records honestly that numerical acceptance criteria were not set in advance.

5. Execution: the runs.

6. Results, with Wilson intervals and deviations:

   • integrity: 0/721 admitted with an invalid identifier or citation (upper bound 0.5%);
   • agreement: 19.6% [15.2, 24.9] of admitted single-cell annotations disagree with the authors' term;
   • the refuted definition check;
   • the funnel figure.

7. Adequacy:

   • integrity is established;
   • semantic correctness is not;
   • literature direction is indicative;
   • clinical use is not adequate.

   It also proposes acceptance criteria for the next evaluation.

Life-cycle maintenance covers model changes, reference updates and audit sampling.

Linked from the README, the docs homepage and the site navigation. mkdocs build --strict passes, and every number was checked against the committed summaries.