A worked example can show that a rule fired once. It cannot show that a rule always fires. The whitepaper's claims are universal — a headline is never published below the coverage threshold, GII-M always lies in 0–100, branch series always sum to the national measure — so they are tested the way universal claims have to be: against a corpus of generated inputs rather than a single record.
Run at build time from @gii/conformance · seed gii-conformance — the entire run reproduces from it.
What the harness found#
It earned its place on the first run, surfacing two defects the worked example could not reach. Both are fixed at the source, with regression tests in the packages that own them.
Both were defects in this implementation rather than in the methodology, and both are fixed at the source with regression tests in the packages that own them. One would have let a reliability statistic pass a production gate on data agreeing worse than chance; the other would have reported a sound national series as internally inconsistent. Neither was reachable from the worked example — which is the argument for generating inputs rather than choosing them.
The invariant catalog#
45 claims, 45 held when this page was built, across 6750 generated cases.
The claims themselves are not published. Each is a sentence specifying exactly what must hold everywhere in the pipeline, and together they are a complete specification of the index’s rules — stated more precisely, in fact, than prose ever states them. That the set holds against adversarial generated input is the claim worth making publicly; the set is the method. Operators run gii conformance for the full catalog.
Reproducibility#
No Math.random and no wall clock — every case is a pure function of its seed, drawn from SHA-256 the way the backoff jitter is. A failure report therefore carries the seed that reproduces it rather than a blob, and a shrinker walks the failing input down to a minimal form that still violates the claim.
npm run gii -- conformance # summary table npm run gii -- conformance --cases 1000 # deeper run npm run gii -- conformance --json # the full report GET /api/v1/conformance?cases=200&seed=my-seed