Fast verdict via a 3-model panel (sync, ~10-25s).

Escalation contract: a claims[].claim text POSTed back verbatim to /verify within 24 hours is verified as-is — in normal operation it does not return needs_input and its wording is not rewritten.

Two input forms, one response shape::

claim / text  ─► ONE text. Lenz reads the claims in it, and each becomes a
                 row, up to 20 a call; the rest go back, unchecked and
                 free, in ``more_claims``.
claims[]      ─► a LIST of up to 20 items, each one claim. One row per item, same order.
                 A compound item is assessed on its primary reading and
                 the rest is listed in that row's ``more_claims``.

The list form is the batch rung of the extract → assess → verify ladder: one /assess call answers up to 20 of /extract's claims (it finds up to 100) in about the time a single call takes, because every item is assessed in parallel.

A verdict can be served from memory: a completed /verify of the exact same claim text, or a prior /assess of it, within the last hour, whichever account ran it — the verdict is a function of the input. verification_url is set only when the caller could read that verification by id.

Billing: 1 credit per row that carries a verdict; error rows are free; a duplicate item is a second row and a second credit (the panel runs once).

Versions. 200: 2026-10-11 receives AssessOut, 2026-05-13 receives AssessOutLegacy; 422: 2026-10-11 receives ValidationErrorOut, 2026-05-13 receives ValidationErrorOutLegacy; default: 2026-10-11 receives ErrorOut, 2026-05-13 receives ErrorOutLegacy.