In one line

Grades the miner's facts against the answer key in 7 areas, using whole numbers only — so grading twice gives the identical result.

6 areas grade overlap: how much of what you found is in the key, and how much of the key you found (0–10,000).
The 7th area — the answers — is worth 40 of 100 points. Right + confident = high. Wrong + confident ≈ 0. Hedging = middle. (That rule is called Brier — "confidence honesty.")
Also graded: citing the right evidence, the right guidelines, and the right "day a doctor could act" — each day late −5%, each day early −2.5% (late = a patient harmed; early = a false alarm).
Mistakes on dangerous questions cost up to ×3.

area	worth	baseline	miner-a	miner-b
question answers	40	8400	9934	6014
where facts came from	15	7692	10000	5778
facts & values	12	10000	10000	10000
connections	12	5000	10000	0
dates	8	10000	10000	10000
contradictions / missing	8	6667	10000	0
ranking	5	5455	10000	0
total		7920	9974	5272

Two quiet fairness rules: each language counts equally (Spanish can't be drowned out by English having more questions), and spotting a conflict across two providers is reported but never scored — so it can't be gamed.

Limits

This measures agreement with our answer key — not medical truth. Getting truth right is a separate, production-level process (native review, adjudication, audit).

→ Go inside (button below, or press →) to see the seven graders as their own scene.