ModelCensusopen-source ai reliability harness
Blog24 Jul 2026methodcommunication

What a share card is allowed to say

The artefact that travels furthest from its caveats needs the harshest gate.

A page can show a rate next to its confidence interval, its tag split, and a link to the transcript behind it. A reader can judge. A share image is 1200 by 630 pixels, it travels alone, and nobody who reshares it reads the methodology.

The pageRate + intervalside by sideTag splittrap vs answerableTranscriptone click awayvs.The cardOne number1200×630No caveat survivesOutlives every qualification
Same measurement, two artefacts. The gate on the second is deliberately harsher than the gate on the first.

Reasons a true number is still inadmissible

  • Pseudo-replication — the difference is real at nominal n and gone at effective n
  • Presence confound — the irrelevant control reproduces most of the claimed movement
  • No control tag — a rate of zero could be perfect calibration or refusing everything
  • Uninterpretable model — a staleness claim about a model that publishes no training cutoff
  • Denominator drift — the arms did not lose the same number of probes, and dropout is never random

Of several hundred candidate comparisons in our corpus, a small fraction survive this gate. The rest are true and unquotable, which is an uncomfortable but honest place to be.

Every figure here describes something measured and committed. See the measurements · read the method