sim_distill
Distill exact search into the play scorer: fit rate multipliers for named transition groups so play's rankings agree with exact minimax, on positions sampled by random self-play and labeled by search. This is TACTICAL calibration — the counterpart of sim_calibrate, which learns rates from an event log. The division of labor is deliberate (the ode-minimax experiment): structure carries the tactic, and no fitting of an unmodified net's rates can express what its final state cannot separate — declare the structural prior as transitions in the model (e.g. forced-reply copies of the plays, catalyzed by the opponent's pattern) and distill the magnitudes it introduced. Zero agreement improvement is a finding about the structure, not a failed fit. Read agreementBefore/agreementAfter, not the loss: the hinge loss can overstate failure while every argmax is right. At most 16 groups, each non-empty, and no transition in two groups; the estimated work (loss evaluations × candidate moves × realizations × horizon) must stay under a fixed cap, and an over-cap request is refused with the estimate. Costs 5 tokens from your account's compute bucket at the defaults, scaled by that work estimate (rounded up).
Input Schema
| Name | Required | Description | Default |
|---|---|---|---|
| id | Yes | model id (needs simulation.objective, players with turnPlace) | |
| groups | Yes | JSON object: group name -> transition ids sharing one fitted multiplier, e.g. {"detectors":["x_win_0","o_win_0"],"draw":["call_draw"]} | |
| options | No | JSON: {"games":20,"positions":40,"iters":40,"horizon":3,"realizations":40,"seed":11,"engine":""} (the defaults shown; 0 or unset takes the default). Caps: games at most 100, positions at most 100, iters at most 100, realizations at most 200, horizon at most 10000; over a cap, a negative value or an unknown key is refused |