← Resources

The AI confidence score, explained properly

2026-08-08 · 5 min read

A confidence number is worthless unless the scale is defined and the basis is stated. Here's the definition, in full.

The bands

  • 100 — directly verified in a primary source that was opened during the answer.
  • 80–99 — corroborated by strong sources.
  • 50–79 — partial support, or some inference on top of evidence.
  • 1–49 — largely reasoning, not verifying. The answer must open with the admission.
  • 0 — no basis at all.
  • Caps and floors

    Any answer under 50 sets the knowledge_gap flag. Any contested figure with no rankable source caps the score at 70, however confident the prose sounds — the number is not allowed to outrun the evidence.

    Every score ships with a one-line basis and a "what would change this" statement. If those two lines are weak, distrust the number, including ours.

    FAQ

    Is the score a probability?
    No. It's an evidence grade: how much of the answer was verified in opened sources versus inferred. Treat it as an audit label, not a calibrated likelihood.
    Share the receiptPost on XRedditLinkedIn

    Ask it something it could get wrong

    Three questions, no card, no account. Every answer comes back with its score, its sources and an exportable trace.

    Ask 3 questions free

    Founding member · $9.99/mo billed yearly · first 100 members