Back to home
Public learning proof

Learning Proof

A first-class proof chain showing whether the V2 engine is actually improving over time, whether the aggregate still matches ledger truth, and whether calibration and challenger workflows are healthy enough to support public claims.

Generated 23/09/2026, 05:04:03

HEALTHY
Predictions Analysed
115K+
114,771 exact surfaced forecasts
Settled Forecasts
110,688
Exact settled surfaced V2 ledger
Model Brier
0.2089
Lower is better
Alpha
-0.0031
Market Brier minus model Brier
ROI
-1.43%
Flat-stake surfaced ROI
Calibration Freshness
0.8h
SLA: 24h
Aggregate Reconciliation
Healthy
Last reconciled 23/09/2026, 05:05:38
Max drift 0.000000
Calibration SLA
Within SLA
Last updated 23/09/2026, 04:14:24
Challenger Activity
Active
Last activity 21/09/2026, 05:04:50
Pending promotions: 0 · recent decisions: 25

Rolling Walk-Forward Evidence

Shareable audit of persisted pre-match V2 forecasts. Rows are grouped by forecast-created week, and a cohort is only evaluated once that market has at least 100 prior surfaced settled forecasts. Forecast rows are not the same as executed bets: many markets can come from one fixture. This is deliberately separate from the tiny scheduled weekly comparison jobs.

Generated 23/09/2026, 05:06:07

Scope: 365d, surfaced-only, engine opportunities-engine-v2, cohort field createdAt, settled-outcome field settledAt. This validates stored forecasts against the market Brier baseline; it is not a claim that every row was executable at unlimited stake size.

Evaluated Holdout
110,190
9,694 fixtures, 110,688 eligible rows
Cohorts
25
1 warm-up cohorts skipped
Alpha vs Market
-0.0031
Market Brier minus model Brier
Flat ROI
-1.45%
-1597.95u over holdout rows
CLV Check
-1.26%
25.3% hit, 60,260 samples
MarketEligibleHoldoutFixturesCohortsAlphaROIModel BrierMarket BrierAvg CLVRange
All Markets110,688110,1909,69425-0.0031-1.45%0.20890.2058-1.26%27/03/2026 to 22/09/2026
Asian Handicap30,16427,5086,73024-0.0081-7.36%0.22310.2150-1.21%27/03/2026 to 22/09/2026
Goals Over/Under19,99319,8705,80325-0.0037-1.28%0.19380.1901-1.60%27/03/2026 to 22/09/2026
Half Goals15,75113,6855,60624+0.0221+2.81%0.20770.2298+0.37%27/03/2026 to 22/09/2026
Match Winner14,50213,3055,86324-0.0017+2.18%0.20000.1983-1.37%27/03/2026 to 22/09/2026
Double Chance8,0036,6963,32724-0.0063-2.45%0.23130.2250-0.76%27/03/2026 to 22/09/2026
Corners6,3984,4061,64618-0.0105-1.91%0.21780.2073n/a28/03/2026 to 22/09/2026
BTTS4,1923,4222,13524-0.0045-7.07%0.24410.2396-0.70%28/03/2026 to 22/09/2026
Team Goals3,0832,85784616+0.0195+64.12%0.21940.2389-1.12%06/05/2026 to 21/09/2026
Correct Score3,2412,4321,1874-0.0019-25.15%0.05250.0506n/a29/03/2026 to 08/05/2026
Clean Sheet2,1672,0001,92624-0.0063-7.65%0.21520.2089n/a28/03/2026 to 22/09/2026
Even/Odd Goals2,2471,03796021-0.0026-5.58%0.25240.2498+0.24%27/03/2026 to 21/09/2026
HT/FT6093092613-0.0031-16.89%0.11720.1141n/a03/04/2026 to 05/05/2026
Win to Nil evidence quarantined

Old Win to Nil rows were recomputed with strict settlement rules, then excluded from headline proof because the original odds feed mapped stale market IDs into Win to Nil. These rows remain visible here as an audit trail, but they do not count toward public ROI, Brier, alpha, CLV, or market-family proof.

Quarantined Rows
6,459
6,801 historical Win to Nil rows
Active Proof Rows
0
0 settled rows currently counting
Removed ROI
+846.47%
Audit-only; +57568.60u is excluded
Strict Audit ROI
+1007.87%
5,957 recomputable rows, still excluded
Dominant Bad Pattern
31.27
Win to Nil - Home / No / key=no / side=home
Reason: win_to_nil_market_id_contamination_2026_04_14. Last quarantine write: 04/07/2026, 13:58:27. Latest captured row: 11/04/2026, 17:43:06. Latest active row: n/a. Generated 23/09/2026, 06:09:29.
MatchStored BetBookCapturedScoreStoredStrictUnit
Rayo Vallecano vs Elche
La Liga
Win to Nil - Home / No
@ 41.00 - Stored outcome differed from strict Win to Nil settlement
bet36503/04/2026, 14:21:341-0wonlost-1.00u
Grenoble Foot 38 vs Clermont
Ligue 2
Win to Nil - Home / No
@ 29.00 - Stored outcome differed from strict Win to Nil settlement
bet36503/04/2026, 14:21:242-2lostwon+28.00u
Grenoble Foot 38 vs Clermont
Ligue 2
Win to Nil - Home / No
@ 29.00 - Stored outcome differed from strict Win to Nil settlement
bet36503/04/2026, 14:18:322-2lostwon+28.00u
Nottingham Forest vs Burnley
Premier League
Win to Nil - Home / No
@ 41.00 - Highest contaminated stored odds
bet36511/04/2026, 14:55:364-1wonwon+40.00u
Rayo Vallecano vs Espanyol
La Liga
Win to Nil - Home / No
@ 41.00 - Highest contaminated stored odds
bet36511/04/2026, 14:05:011-0lostlost-1.00u
Cambridge United vs Notts County
League Two
Win to Nil - Home / No
@ 41.00 - Highest contaminated stored odds
bet36511/04/2026, 13:46:514-0lostlost-1.00u
Proof Chain Windows
Exact surfaced settled windows comparing older periods against recent behavior.
WindowSampleBrierMarket BrierAlphaROIAvg CLVCLV Hit
Older 7d7,0430.21340.2130-0.0003-0.16%-0.9323.4%
Recent 7d5,5350.20870.2072-0.0015-7.55%-0.7324.2%
Older 30d11,1210.21840.2214+0.00296.21%-0.9928.0%
Recent 30d24,1830.21380.2139+0.00010.01%-0.9022.5%
7-Day Improvement
Older 7d vs recent 7d, with 95% confidence intervals on the change.
Brier Improvement
+0.0047
95% CI -0.0007 to 0.0102
Directional only · improved
Alpha Improvement
0.0000
95% CI 0.0000 to 0.0000
Directional only · worsened
ROI Improvement
-7.40pp
95% CI -12.23 to -2.57pp
Statistically clear · worsened
Avg CLV Improvement
+0.19%
95% CI -0.03 to 0.42%
Directional only · improved
CLV Hit Improvement
+0.83pp
95% CI -0.71 to 2.36pp
Directional only · improved
30-Day Improvement
Older 30d vs recent 30d for a broader view of whether the learning story survives over longer horizons.
Brier Improvement
+0.0046
95% CI 0.0012 to 0.0081
Statistically clear · improved
Alpha Improvement
0.0000
95% CI 0.0000 to 0.0000
Statistically clear · worsened
ROI Improvement
-6.20pp
95% CI -9.50 to -2.89pp
Statistically clear · worsened
Avg CLV Improvement
+0.09%
95% CI -0.10 to 0.28%
Directional only · improved
CLV Hit Improvement
-5.41pp
95% CI -6.44 to -4.38pp
Statistically clear · worsened