It found where it was
systematically wrong.

I · The rewrite

Watch the rules edit themselves.

Four rules. The model began with naive priors — then, generation by generation, it struck out what the evidence contradicted and wrote in the base rate it had measured. Each margin note is the diagnostician's own recorded justification for the change.

II · The diagnostician

Where it caught itself.

After each generation an isolated grader handed back only a Brier score and the miss distribution. The model had to find its own bias and decide what to do — including, twice, deciding to do nothing.

III · The evidence

It actually got better.

With no information, optimal forecasting is learning the conditional base rates — the model discovered which structure to condition on (deadline vs threshold) and when to stop. That's calibration discipline, not curve-fitting.

IV · The guarantee

Earned, not looked up.

The whole point is that improvement came from correcting its own calibration errors — never from peeking at the answer or searching the web.

V · Appendix

VI · The transfer

The same engine, a domain it had never seen.

The real product is the loop, not the specific rules it learned. To prove it, the byte-identical workflow was pointed at a different world — "will this GitHub issue be closed within 30 days?" — with only a new dataset and a fresh, naive gen-1 playbook. No code changed. If the curve falls here too, the engine generalizes.