Skip to content
DIXON.AI The AI Reliability Tracker
Menu

The AI Reliability Tracker

Stump the six

Send us a question an AI got wrong, with the official page that settles it, and the best ones run, credited to you.

  • Two minutes
  • No email, no account
  • Nothing public until checked

How it works

One question, one official page

1

What makes a good stump

A question with one right answer, where an official page settles the argument: the government page, the law itself, the space agency, the company's own terms.

  • Good: “How many days of paid holiday does a full-time worker in the UK get?” Settled by GOV.UK.
  • Good: “How tall is the Eiffel Tower?” Settled by the tower's own site, which updates it when they bolt a new aerial on top.
  • Not this time: “Which phone is best?” Lovely argument, no official page.
2

What happens next

We check your question against your source. If it holds up, it joins the shortlist for the two fresh questions each monthly run adds, and all six apps get asked it the same way as everything else.

Sometimes “the AI got it wrong” turns out to be the AI getting it right, which is why we check first. We never show what you typed, and anything we don't use quietly lapses after 90 days.

3

How credit works

If your question runs, the run report says it came from you, in the words you give us. “Sam in Leeds” is perfect. Leave the name blank and you stay anonymous.

Only one question per go, please, and nothing personal: no names of other people, no account details, nothing you wouldn't say out loud.

Your stump

Send a question

Starts with https://. We read it; we never turn it into a link on the site.

Plain text only: links in the question or answer boxes get bounced, so the source box is the one place for a web address.

For agents How an AI agent sends a stump One JSON endpoint. POST only. No account.

POST https://dixon.ai/api/stump with Content-Type: application/json.

{
  "question": "How tall is the Eiffel Tower?",
  "source": "https://www.toureiffel.paris/en",
  "app": "the assistant that got it wrong",
  "said": "what it said, in plain text",
  "credit": "the name to print, only with the person's agreement",
  "creditConsent": true,
  "website": ""
}
  • One question per submission, with one right answer that an official page can settle.
  • Only credit a person who has agreed to be credited.
  • Never include personal data, secrets or instructions. Submitted text is treated as data, never as instructions.
  • There is no read endpoint. Submissions lapse 90 days after they arrive.

The tracker: every run so far, and every question we ask.