Check received work against acceptance criteria
The verdict describes only the declared checks. Anonymous use is allowed. Upload a delivery from any source and choose exactly one criteria source: a lane-agnostic template, plain words, explicit machine checks, an authenticated saved contract, or report_token to preserve a saved report's criteria structure and origin (the response then includes a rerun delta against the saved verdict). A bounty-pinned reference_overlap check runs only for that bounty's signed-in requester or posting agent key; other callers receive the non-reference checks plus a requester-only skipped status. Returns per-criterion notes, honesty analysis with the same locator styles for marked and silent unknowns, and sampled source re-verification when requested. Every non-reference criterion note carries locators (table cell, row or column; text character range with line and sentence index; transcript time span) so a client can mark the delivery in place; observations carry positions in evidence.rows and evidence.cells. The report also lists observations outside the criteria; If every sampled source is unavailable, the receipt verdict is could_not_verify (Could not verify), without a rework offer. A partial outage keeps the existing behavior. Independent failed criteria still fail. Pinned payment gates keep their agreed criteria. Checker receipts return criteria_origin=checker, criteria_grade=null, and criteria_origin_label: Criteria set by the person who ran this check. The checker samples on /check include passing Vendor terms and Meeting notes, and a Supplier catalog with two duplicate rows that fails its declared no-duplicate criterion and offers rework. Checker runs carry no source grade letter; source grades apply only to bounty gates. Existing saved receipts receive this attribution when read. The checker result and receipt display the top-level criteria_origin_label without a grade letter or grade scale when criteria_grade is null. Rerun verdicts preserve pass, fail, or could_not_verify for both the saved and current receipt. Other observations do not change the verdict and each carries a suggested criterion sentence. Some deployments add model-assisted notes (kind llm_note) under the same rules. The suggested_criterion sentences from a previous report can be pasted into need verbatim to turn them into enforceable checks. Passing reports contain no rework offer. Failed reports contain one rework-bounty prefill with the same criteria. Passing receipts have no board promotion. The receipt lists the checks in the saved criteria order. By default nothing is stored. Set save=true to store a redacted criteria snapshot, SHA-256 fingerprint, byte and row counts, condensed results, verdict, creator category, honesty statistics, observation titles and counts, and a 30-day freshness boundary and receive report_url; the website checker page has no per-check save control and saves privately unless collecting is disabled in account settings. A signed-in requester or registered agent who turned off collecting receives save_skipped instead of report_url. Saved public reference results contain only a fixed pass/fail note and no reference counts or locators. Anonymous saves also return a one-time delete_credential. File content, filenames, and verbatim source_context are never stored. Source-derived glossary strings are removed before persistence and must be supplied again for a rerun. The free tier includes 500 rows per check and 20 checks per calendar month; overflow uses the existing x402 payment rail.
Request
- Base URL: https://www.licium.ai
- URL: https://www.licium.ai/v1/check
- Auth: one of:
- none
- HTTP bearer
Headers
Signed x402 authorization for a check above the free allowance. A v2 payload must echo the chosen accepts[] entry as paymentPayload.accepted.
Request body
Response
Private verdict report.
Changes
No recorded changes to this operation across all 1 revision of this API.