Gary AI model audit

PUBLIC LEDGER

A recomputed view of the same results Gary publishes on the record: monthly performance, outcome calibration by confidence band, and the raw data behind both.

← Track recordDownload CSVDownload JSON

Monthly game-pick tape

Every decided game pick remains in its month. Units use flat one-unit stakes at the listed odds; pushes count as graded but not as decisions.

MonthRecordWin %NetGraded
August 2026196-20049%-26.9u396
July 2026203-19551%-19.9u398
June 2026268-17960%+64.1u447
May 2026145-11955%+6.4u264
April 2026217-18654%+5.2u403
March 2026279-249-653%+0.2u534
February 2026116-116-150%-10.4u233
January 2026124-119-151%-4.4u244
December 2025122-10654%+5.6u228
June 20255-550%-0.8u10
May 202543-2563%+12.4u68
April 20251-150%-0.1u2

Confidence calibration

Confidence is Gary's ranking signal, not a promise of win probability. This table shows what actually happened inside each displayed band.

60–69% confidence

92-82

53% won · 174 graded · avg label 66%

70–79% confidence

53-37

59% won · 90 graded · avg label 75%

80–89% confidence

13-18

42% won · 31 graded · avg label 83%

90–100% confidence

1-1

50% won · 2 graded · avg label 92%

Player props

2528-3369 · 43% · -487.3u · 5897 graded

Some historical winning props have no recorded odds and use the disclosed 0.9-unit fallback in net-unit calculations. The raw export retains the original missing values.

Methodology, known corrections, and source policy are documented in the editorial standards, corrections process, and data-source policy.