vision-eval-harness

Live demo — every page below is a real artifact rendered by the harness in CI on the latest commit: two seeded evaluation runs, one healthy and one deliberately broken, gated statistically.

GATE PASSED

Passing run report

Candidate matches the promoted baseline — McNemar finds nothing, every check green. Metrics, check table, latency histogram.

GATE FAILED

Failing run report

Same model, images saturated with --brightness 3.0 — accuracy collapses, the gate blocks with exit 1. Worst samples name the damage.

TREND

Trend report

Accuracy and p95 latency across runs, with gate-colored markers — the regression is the red dot.

MACHINE-READABLE

Gate artifact JSON

The failing decision as downstream tooling sees it: per-check verdicts, values, thresholds, p-values.