Patient-evidence verifier boundary

Problem

If retrieved features can reach the verifier, a similar patient can vouch for a finding the current patient does not show. The boundary has to be enforced by the interface, not by convention.

Contribution

Relationship to research: Independent implementation. Isolates the verification boundary described in a manuscript shared with the account owner. No learned verifier is trained here.

Account owner’s role: Commissioned, reviewed and published the implementation. The code was produced with AI-assisted generation on commission and reviewed before release; it is not the original code of any cited publication.

What was built

• PatientEvidenceBundle and CandidateAttributes as frozen records with schema-level rejection of reference content
• Rules-based fixture verifier, trainable adapter interface, calibration wrapper fitted on validation only
• Reference-independence and prior-sensitivity tests

Method

Support scores are compared under permuted reference-side inputs (must be identical) and under changed prior features (may differ).

Data and access

Three hundred synthetic candidate/evidence pairs with grouped patient splits.

Evaluation protocol

Precision, recall, F1 at 0.6; expected calibration error; coverage under abstention. All synthetic.

Results

No measured result is published for this project. Demonstration outputs from the repository’s own synthetic fixtures are labelled as such in the repository and are not presented here as results.

Limitations

• A learned reproduction needs the full verifier specification and authorised data, neither of which is available here.

Reproducibility

Offline; deterministic fixture backend.

Repository metadata: Python, MIT, last push 2026-09-13, 0 stars.

Source publication and links

No publication is attached to this project.