A9 PREDICT

A9 Predict

Changelog

Changes associated with this build.
v0.2.0Build identifier supplied

Build date · Eastern: Oct 25, 2026, 1:41:30.352 PM EDT

Technical build details
Commit
ef695042b6cc7b8d0132bc9eb2a0a1890d123d57
Version source
Running build metadata

Build metadata does not confirm production deployment or service health.

v0.2.0 This build

Keep this round's issued prediction readable, with a separate fresh entry comparison when matching evidence is available.

Added

  • One-contract entry estimate from matching fresh forecast, quote, depth and timestamped series fees; no orders are placed.
  • Current Kalshi round, threshold, close countdown and timestamped public quotes, separate from immutable A9 forecasts.
  • Eastern-day count of unique rounds with acknowledged forecasts and saved settlement status; real trades remain unavailable without fills.

Fixed

  • Saved history stays historical, including records from an open round; it does not show a current pick or start quote refreshes.
  • An issued prediction stays visible until its round closes, with its original probability and an explicit stale-input label after freshness expires.
  • Invalid reference timestamps, conflicting records and different-round quotes cannot supply a selection or entry comparison.
  • Forecast-only pilots can use the approved budget while learning stays off.
  • Quotes retire when stale, closed, mismatched or incomplete; an old forecast is never relabeled as current.
  • Bounded response parsing and shared in-flight public reads prevent simultaneous viewers multiplying upstream work.

Changed

  • Target prices spell out At least versus Above from the contract comparison; original Kalshi wording and rules remain in Details.
  • Untracked no-pick counts and the absence of a trading connection are labeled explicitly instead of appearing as a temporary data outage.
  • Canonical ACE9 Living Glass theme, compact historical evidence, and Eastern Time throughout.
  • Accessible version dialog groups Added, Fixed and Changed notes for this exact build.
  • Page reads do not start forecast collection, learning or order execution.

A9 Predict · Track record

Does A9 actually beat the market?

What this means · Market is better

The market is better: across 1396 settled questions the market's odds were closer to what happened. A9 has no edge here yet.

The fair test is "Made at least 24h before close" . The real question is the "Researched questions" row : that is where paying for research has to prove itself against the market.

Edge scorecard by forecast group
GroupSettled A9 Jev raw Market A9 vs market Verdict
All settled forecasts 13960.2280.1140.226+0.002 (+0.000 to +0.003)Market is better
Made at least 24h before close 1130.0950.1170.085+0.010 (+0.004 to +0.016)Market is better
Current method (a9-predict-0.3.0) 890.1040.1330.103+0.002 (-0.000 to +0.004)No clear difference
Researched questions 13060.2370.1350.236+0.001 (-0.000 to +0.002)No clear difference
Market-efficient questions 900.0960.1050.084+0.012 (+0.006 to +0.017)Market is better

Scores are Brier scores : lower is better. In "A9 vs market", a negative number means A9 was closer to what actually happened.

Settled 1396
Leaned the right way 56%
Calibration Excellent

When A9 said...

~6%happened 3%n=30
~14%happened 3%n=32
~26%happened 17%n=29
~36%happened 32%n=104
~46%happened 45%n=453
~53%happened 52%n=590
~64%happened 64%n=101
~73%happened 91%n=32
~85%happened 87%n=15
~93%happened 100%n=10