How it worksVerified at 4e59cca

Known limits and design decisions

What this site has not proved, and what it gave up on purpose, knowing the cost.

Things this site has not proved

  • The threat model is my own review of my own design. Nobody else has attacked it. (ADR-0055)
  • The reasoning that no cookie banner is needed, and the privacy notice itself, were written from the regulations and the regulators' guidance. No lawyer has read either. (ADR-0053, ADR-0083)
  • The injection results are one run per model, and three of their conditions are judged by models rather than checked by code. (ADR-0055, ADR-0062)
  • The scope gate refuses some questions whose writer introduces themselves first, such as a developer asking which library he uses: 4 of 25 in a set written to test this, none in its holdout. (ADR-0081)
  • The page counts are Cloudflare's estimate and include crawlers. (ADR-0086)
  • Three of the metrics I set out to track, such as the time it takes you to find something, are not measured at all, because the only way to measure them would be to watch you. (ADR-0057)

Things I chose, knowing the cost

  • The assistant forgets you when you close the tab. (ADR-0036)
  • A question it cannot place is recorded as a topic it looked for, never as your sentence, so the most useful editorial signal I could have is the one I gave up. What is left is one of nine words for what the question was about. (ADR-0087, ADR-0054)
  • Asking the assistant sends your words to two companies in the United States. I would rather tell you exactly which, and what they keep, than promise something I could not check. (ADR-0073)

Each names the decision record behind it. Those records are not published.