BioGuard AI — Refusal demonstration

Build a request. See what the service actually did with it.

Six choices, sixty-four combinations, every one posted to the packaged BioGuard AI service and recorded. Change anything on the left and the panel on the right shows the real response — which conditions fired, the detail the service gave, and its own explanation for each.

What this page is, and what it is not

Every response was produced by service 0.14.5 and recorded on 8 October 2026. This page holds those recordings and selects one; it evaluates nothing in your browser. Evaluating the request here would be a second implementation of the abstention rules, and the moment it disagreed with the service this page would be wrong while looking authoritative.

Inputs are synthetic fixtures with known ground truth — no biological specimen and no deposited study is represented. No successful interpretation is shown, and that is deliberate: the model is fitted to a cohort this project generated, so a value computed from it says nothing about cancer. Where the service did compute, the number is replaced by a visible marker rather than quietly dropped.

RequestPOST /interpretation/interpret

Analytical QC verdict

Panel completeness

PSA concentration

Specimen matrix

Measurement system

PSA unit reported

Presets

How these refusals are held in place

554Tests passing 0 failing.

100%Mutation score 2 of 2 injected faults caught.

96.3%Line coverage Of 3,134 statements. Reported, not gated.

all pass Quality gatesCoverage, module floor, mutation, failing tests.

Mutation score is the gate: faults are injected into the governance and interpretation code and the suite must catch them. A refusal no test asserts is a refusal that can be deleted by accident. Measured 12 October 2026 against service 0.14.5; reproduce with python measure_quality.py --mutation. The run log is append-only and hash-chained, which makes alteration detectable — not impossible.