Kalshi Chess Market-Making

The model explainer: inference pipeline (Stockfish → features → logistic → encounter prob), why eval alone fails, the training data + method, and blind validation (skill +0.34, calibration ECE 3–7%).

Books are 16k-deep → incentive farming is dead; pivoted to directional (take mispricings). Better model: clock+elo nearly double skill (+34%). Directional backtest positive but ~3-game sample. Tomorrow: $250 live test.

7-agent red-team of the spec; built engine+PGN pipeline; two backtests on captured data — raw Stockfish WDL is miscalibrated for humans (skill −1.12), a human-calibrated matrix beats it (+8.9%); market tracks the engine with 10–31¢ gaps. N=24 games (small).

Opportunity thesis, market universe, the read-only logger, and live data we started capturing.