RReasoningReceiptlive →

Calibration — the agent measures itself

Once a market on Polymarket or Kalshi resolves, the resolver back-fills the outcome onto every receipt the agent emitted for it. That gives us ground truth for every prediction. We then compute a Brier score (mean squared error between predicted probability and actual outcome — lower is better, perfect forecaster scores 0) plus a 10-bucket reliability curve.

A trivial “50% on everything” forecaster scores ~0.25. A good prediction-market analyst typically lands between 0.10 and 0.18.

Brier score
0.2471
across 964 resolved receipts
High-conf Brier
0.2475
confidence ≥ 0.7
Low-conf Brier
0.2462
confidence < 0.7
Resolved markets
21
out of 964 receipts
Reliability — predicted vs actual
dots near y=x = well calibrated · dot size = receipts in bucket
Brier over time — is the agent learning?
lower = sharper · dashed 0.25 = coin-flip baseline
Bucket breakdown
Bucketnmean predictedmean actualdrift
0.10-0.20118.0%0.0%-18.0 pp
0.20-0.301126.6%0.0%-26.6 pp
0.30-0.404535.4%4.4%-31.0 pp
0.40-0.5015446.2%17.5%-28.7 pp
0.50-0.6035554.7%26.5%-28.2 pp
0.60-0.7025764.0%59.5%-4.4 pp
0.70-0.806573.7%70.8%-2.9 pp
0.80-0.904383.6%60.5%-23.1 pp
0.90-1.002794.2%55.6%-38.7 pp

Note: a market is treated as resolved YES iff Polymarket Gamma reports closed=true and the YES close price is within 5% of 1.0; resolved NO iff the YES close price is within 5% of 0.0. Ambiguous markets (close price near 0.5) are not counted.