# DN Research Export Acceptance Lab v0.1
+
+Published 3 November 2026. A fixed synthetic task rubric, custom checker and 26 invented cases. This is not a vendor benchmark, fraud detector, factual verifier, payment gate or legal acceptance standard.
+
+The rubric is compiled into checker.mjs outside the candidate export. A candidate cannot change it by adding a policy field. The task requires asset-a, asset-b and asset-c exactly once each, in any order; USD values in integer cents from zero to 1,000,000,000,000; strict UTC timestamps at whole-second resolution; and source references using credential-free HTTPS on research.example with no nondefault port. Generated-at and observation-time claims must be within the inclusive 24 hours ending 2026-11-03T18:00:00Z. An observation claim cannot be later than generation. The fixed clock is synthetic, not a freshness recommendation for every task.
+
+The .example host is reserved for documentation and no source is fetched. Passing checks timestamp and source claims, not their authenticity. The invented-source-path and invented-value controls deliberately pass: a well-shaped record is not evidence of factual truth.
+
+26 locally executed cases match their declared expectations: six meet this rubric and twenty do not. No model calls, paid requests, provider calls or source fetches occurred. Results are not an observed agent success rate. The custom checker does not use a generic JSON Schema engine; export.schema.json is a structural guide. Its format annotations do not replace the custom clock, coverage and exact-host checks. Configure any separate validator explicitly before relying on format assertion.
+
+Download checker.mjs, fixtures.json and reproduce.mjs to one folder; run `node reproduce.mjs`. No dependencies or network are required. results.json includes the SHA-256 of the executed checker. A hash identifies source bytes; it is not publisher authentication or a trusted timestamp.
+
+The browser checker accepts up to 50,000 characters and the function rejects record arrays outside 1–100 items. The fixed rubric still requires exactly three named records. Use only synthetic or redacted examples. Never paste credentials, personal data or live commercial exports. The checker itself performs no request or persistent storage; ordinary website analytics are separate.
+
+License: CC BY 4.0. Attribute Decentralised News and link to https://decentralised.news/ai-agent-acceptance-tests. Retain the limitations. Third-party standards and documents retain their own terms. Corrections can be submitted through DN Contact with the case identifier and reproducible evidence.
+