I built Owin because a buyer’s security questionnaire asks whether your AI providers keep customer data, and the true answer differs per endpoint.

Three guided paths, about a minute each

1. 1 Replay a real change Not true Answer 3 is not true for /api/crm-sync: Anthropic’s MCP connector is outside ZDR. Answer 5 became outdated when OpenAI changed its sub-processor list on Oct 14, 2026. Open the replay
2. 2 Change one input Verified Point /api/crm-sync at Anthropic’s plain Messages API. The rules run again on the server and answer 3 becomes true. Change it
3. 3 Try thin evidence Not stated Point /api/crm-sync at Anthropic’s rate-limits page, which says nothing about data use. Owin marks those answers Not stated. It never invents one. See it marked Not stated

Also: the buyer page shows every answer per endpoint with the vendor’s sentence, and you can see an answer that needs review, where three readings of OpenAI’s page split 2-to-1, so Owin hands the /api/embed cell to you.

Evidence

• Claims ledgerEvery number here, with the command behind it. A test fails if one drifts.
• READMEWhat failed, the labelling limit, the label error we found ourselves, and what is fictional.
• The buyer’s copyThe buyer report as a buyer receives it. No internal reasons, no fix drafts.
• CSVOne row per question and endpoint, with the vendor sentence and its link.
• What runs itTavily extract fetches each vendor page. NVIDIA Nemotron 3 Super, an open-weight model on Nebius Token Factory, reads each page three times; this page replays those recorded runs. Code, not the model, decides every answer.