Start with the gaps.

The two-authority crosswalk — where every board row went

The payer-facing denominator is board rows — the Missouri Dental Board roster, not the NPPES pool (a federal source whose result count we do not control). Every board row goes to exactly one of three classes: joined (Tier A, identifier equality), not joined (each with a stated recorded reason), or a cross-field contradiction finding. The contradiction findings sit OUTSIDE the join/not-join partition — a contradiction row is a finding, never also a not-joined row — and INSIDE the board-rows denominator.

Why rows do not join — all six reason classes, with denominators

The non-join reasons partition the not-joined rows — one reason per row, so the non-join reason counts sum to the not-joined row count. The denominator for the five non-join classes is the not-joined row count in the disposition above. The contradiction class counts finding rows — a finding class, never also a not-joined row.

Method supersession — method change on fixed evidence, never an improvement claim

Ruleset changes are method effects — the same snapshot evidence, processed differently. A changed ruleset joins the SAME board roster and NPPES pool the prior one joined. The observable diff between two epochs is the method effect, not an accuracy improvement. No epoch can claim to be “better” — only different by a stated method change.

Our own error rate

We re-draw a random sample of 50 recorded verification rows from the verification ledger — our own recorded verification events, not the propositions published in signed records — and re-derive each drawn verdict independently from its cited observations, not from the code that produced the verdict. Read what the number measures before the number: the result is a bound on our own reproduction of our own verdicts. It is not a provider score, not a measure of source correctness, adjudication quality, or any plan's directory accuracy. An auditor who also sells you the answer is not an auditor.