Every number in an AI answer, read back against the source it should have come from. Matched, derived, or no match, with the reason for each.
Rules: a count must match to within half a unit or 0.5%; a share or a dollar figure to within half a unit at its stated precision or 1.5% (5% for figures rounded to k or m). Derived means the number is not in the source but is one step from figures that are: a difference, 100 minus a share, or one count over another. Anything else is no match. Each gauge puts the source figure at the centre, the tolerance band around it, and the number the answer gave where it actually falls. Phrases that import outside knowledge are marked in yellow. Examples: the replayed Claude Sonnet 5 run from the vehicle-intelligence experiment, synthetic data. Parity with the published Python check: 385 of 385 claims.