When the brain says danger,
does death actually follow?
Every brain in the dish returns a danger score from 0 to 3 with each sealed decision. That number is easy to print and easy to fake. The death-calibration test checks it against the only grader that cannot be gamed: dying.
1 · The question
Confidence is a claim about the future. "Danger 2.8" claims: given this situation, something lethal is likely to happen soon. In a quiz, nobody ever checks. In the dish, the world keeps running after the claim — so we can.
This is the same lesson as the judge-audit failures in the judgment-model ecosystem: a model can sound certain and be wrong. The cure is not a better prompt. The cure is consequences that arrive on a clock.
2 · The method
Run the exam room (?bench=1) with any brain. Every decision is already sealed with its situation and a content hash. For each generation we know one more thing the model never sees at decision time: when the fly died. Then:
- A decision is positive if the fly died and the decision was made within 5 seconds of death.
- Bucket every decision by its danger score:
0–1,1–2,2–3. - Report the actual death rate per bucket, plus a Brier score treating
danger/3as the claimed probability of dying within 5 s.
A perfectly calibrated brain would see death rates rise monotonically with the bucket — roughly 17% / 50% / 83% if scores were uniform within buckets. A brain that cries danger at nothing shows a flat table. A brain that dies surprised shows deaths in the low buckets.
| danger bucket | a calibrated brain looks like |
|---|---|
0–1 | almost no deaths within 5 s |
1–2 | deaths begin |
2–3 | death is the norm, not the exception |
3 · What it does not prove
One seed is one row, not a theorem. The table is a property of this brain on this world with this predator, measured — a reproducible number, not a moral. The danger score is also actionable: a brain that fears danger and successfully escapes may show high scores and no deaths, which the raw table reads as over-prediction. That ambiguity is honest and worth stating: calibration in a closed loop mixes perception with policy. Disentangling them — scoring the read, not the reflex — is future work, and the sealed logs are exactly the data such work needs.
4 · Run it yourself
- Judgment brain · seed 42 · 4 generations — the calibration table prints on the exam report.
- Circuit brain · seed 1337 — the same test for the 24-neuron circuit.
- Change the seed in the URL and the whole world — food, predator, drafts, deaths — changes with it, deterministically.
The dish does not grade opinions. It grades the next five seconds.