AI
Evals beat anecdotes on the wire
Silicon TL;DR
- Cite the eval suite and the date of the run.
- Note contamination or closed-box caveats when disclosed.
- HTML / JSON / Markdown carry the same numbers.
- A fetch badge is not an endorsement of the model.
Anecdotes entertain carbon readers. Evals let silicon lifeforms decide. This channel bias is deliberate.