Calibration
Graded calls with a market price recorded at open, scored against that price on the same question. The Brier score measures squared error — lower is better — and it punishes confident misses hardest. The market price on each row is the one we recorded at call time, not an independently sourced venue history, so this comparison is not yet a verified measure of skill against the market.
32 verdict-bearing calls with both stamps, of 65 resolved. Row-by-row: EHIQ closer on 19, market closer on 12, ties 1.
By venue
| venue | n | EHIQ | market | delta |
|---|---|---|---|---|
| polymarket | 24 | 0.2140 | 0.2180 | -0.0040 |
| kalshi | 5 | 0.1431 | 0.1228 | +0.0203 |
| manual | 1 | 0.2025 | 0.4225 | -0.2200 |
| instrument_implied | 1 | 0.3025 | 0.5439 | -0.2414 |
| sell_side_consensus | 1 | 0.1936 | 0.4225 | -0.2289 |
day-0 baseline 2026-09-13: the machine-proposal batch (#371-376) resolves inside the next 60 days; this curve is where "is the machine beating the crowd" gets answered.
How the outcome is recovered: hit=True means the call was right, not that the event happened — a 2% call that “hits” is the event not happening. Outcome = (stamp ≥ 0.5) == hit, exact at the 0.5 threshold. The first readout of this page scored 0.42 by forgetting that; the rule is printed here so it never happens again.