AI accuracy report
Real questions. Real AI answers. Judged blind by licensed humans.
Every signed letter on the registry adds four scores per AI answer (1–5): found the core issue · facts correct · raised the missing question · caught the risk. Customers add two: understood · helped me decide. This page is their monthly sum.
0
checks opened
0
letters signed
—
best AI answer still needed a correction
0
licensed reviewers who signed
No signed letters yet.
The first cells appear with the first signed checks. Model names are shown only when a model × kind × month cell has 100+ judged answers; until then the kind trend is published without names.
How to read it
- Reviewers score answers labelled A/B/C; the model names are revealed to them only after they sign.
- Four scores per answer, 1–5: found the core issue · facts correct · missing question raised · risk caught.
- Model names appear in the report only when a model × kind × month cell has at least 100 judged answers; thinner cells roll up into the kind trend.
- Questions come from people who paid for a check, so the sample is skewed toward hard, personal questions — the ones where AIs disagree. That is the point.
JSON: /api/report · Add your own data point: check an AI answer. Generated 2026-10-20 11:18 UTC.