How reliable the published decisions have been, measured on their outcomes — with the evidence behind every number, every intelligence release and its measured change, and the price ladder. Every number publishes with its n.
Decision intelligence
One composite of the measured quality of the published decisions, on a scale of 0 to 100. It publishes only past the evidence floor.
15.8
out of 100
n = 3,413 realized outcomes · window 2026-09-18 – 2026-10-17
index 1.0.0 · snapshot taken 06:57:22 UTC
The counts
3,935
Decisions in the window
3,413
Realized outcomes
11
Published calibration bands
0
Releases published
0
Rejected candidates
Window 2026-09-18 – 2026-10-17 · every number publishes with its n.
The learning curve
Error falls with evidence on a published power law.
Each dot is a measured error at that many realized outcomes. The chart shows 400 of the 8,314 measured points (evenly spaced, first and last kept) — the full series is served at /v1/intelligence/curve and the published fit is computed over all 8,314. Dashed vertical lines mark the intelligence releases. The solid line is the published fit — nothing else is drawn through the dots, and nothing is interpolated across a gap.
Published fit: BS(n) = a·n^(−b) + c with a = 0.01239029332009478, b = -0.37226348545630417, c = 0 — the exponent's 95% interval -0.37283053133376265 to -0.37175864055689. Measured over 8,314 points, n = 3,413 realized outcomes.
No qualified crowd benchmark yet. At least 10 graded player calls per question are required before the crowd line is published.
The reference line is a forecaster that always predicts the base rate — every skill score on this page is measured against it. The crowd line appears only where at least 10 scored calls exist for the question. It is a comparison only, never an input.
Intelligence releases
Every intelligence release publishes with its measured change. A rejected candidate stays on the record with the failing gate’s reason — the failures are part of the record.
No release has yet proved improvement over the current intelligence. Rejected candidates remain on the public record.
The ladder
Price rises exponentially with evidence — the ladder, +25% a rung.
Rung 1 · prices are ×1.25 of the opening price · in effect since 2026-10-04
Evidence gate
realized outcomes · 3,413 of 10,000 · not met
Intelligence gate
release holdout skill · — of holdout Brier skill > 0 (95% CI) · not met
What remains: 6,587 outcomes and the release holdout skill gate.
A rung, once crossed, stays crossed: the price does not fall back when a measured number dips, so between rungs the price can sit above the measured skill of the moment. It follows the published schedule, never a live estimate. The price on a decision already in progress holds.
Every number publishes with its n and the window it was measured over. Nothing on this page is estimated in place — every figure is computed from the recorded outcomes and then published.
The decomposition
Resolved pairs are scored with Murphy’s decomposition of the Brier score:
BS = REL − RES + UNC
Reliability (REL) is the cost of miscalibration — small means that when Tollbridge says 70%, it happens about 70% of the time. Resolution (RES) separates events from non-events — larger is better. Uncertainty (UNC) is the base rate’s own noise. Against a reference forecaster that always predicts the base rate:
BSS = 1 − BS/BS_ref
Skill is exactly the resolution earned minus the reliability paid, normalized by the base rate’s noise. A negative skill score publishes as negative.
Skill carries 0.60 of the weight, calibration error (ECE) 0.25, and the resolution rate — how often a paid question actually settles — 0.15. Weights change only with a new index version, and curves never mix index versions.
The floors
30 realized outcomes publish a question’s calibration. 10 outcomes in a band publish that band’s frequency. 100 realized outcomes publish the index. Below a floor the counts are shown and no claim is published.
What the record claims
Error falls with evidence on a published power law.
Price rises exponentially with evidence — the ladder, +25% a rung.
Each new question family launches at the accuracy of the whole record.
What it does not claim
Accuracy does not grow exponentially. Error falls on a power law in evidence — steep early, asymptotic later.
One question teaches nothing below the floors.
Unresolvable outcomes are never scored as errors, and a declined question carries no probability claim.
A rejected candidate publishes with its deltas. The failures are part of the record.
Predicted against actual, by range and by question, is published on the results page.