Demo -- read-only snapshot of the live system as of 2026-07-16T01:45:31Z; no live data, paper units only.
Skip to main content

evidence / Calibration & market

In-Game Conditioning -- the one measured calibration win

Pregame, the market is efficient: across 4 sports and 6 independent corpora my leak-free forecaster MATCHES the Shin-devigged closing line within noise and beats nothing. The one place the machinery measurably sharpens is IN-GAME -- fusing the pregame rating prior with the realized mid-game state improves win-probability calibration. A live book sees the score too, so this is a CALIBRATION result, not a claim of beating anyone. Every number below is Brier / MAE / row-count with its source artifact. The single truth-source for any figure is docs/JOB_EVIDENCE_PACKET.md; edge_claimed = False throughout, and no dollar, ROI, or edge figure appears anywhere on this page.

strongest single receipt

The honest question is not "can I beat the close" -- I proved against myself that I cannot, pregame -- but "does the system's mid-game conditioning actually sharpen the forecast." It does, and the win is measured, not asserted: conditioning the win-probability forecaster on the realized in-game state improves calibration (Brier) on real out-of-sample corpora in two sports

the claim

The honest question is not "can I beat the close" -- I proved against myself that I cannot, pregame -- but "does the system's mid-game conditioning actually sharpen the forecast." It does, and the win is measured, not asserted: conditioning the win-probability forecaster on the realized in-game state improves calibration (Brier) on real out-of-sample corpora in two sports. This is forecaster quality. A live book also sees the score, so no beat-the-market test applies to an in-game number and none is claimed.


receipts

claimcommitted artifact
NBA static->conditional Brier 0.209 -> 0.159scripts/platformkit/proof_nba/ingame_accuracy.py
MLB static->conditional Brier 0.241 -> 0.126scripts/platformkit/proof_mlb/ingame_accuracy.py
Cross-sport roll-upscripts/platformkit/ingame_scoreboard.py
end_q1 MARKET_SHARPER delta -0.0084, halftime/end_q3/q4 UNDERPOWERED, n~1593scripts/platformkit/benchmarks/crps_market/last_run_ingame_nba_winprob_ALLGAMES_v3.json
Same, pre-registered ledger row[RECEIPTS.md](../../RECEIPTS.md) (batch 6eba592d3a77, nba end_q1 row)
~46% pooled / ~26% learned-head MAE liftleak-clean 14-feature schema, two walk-forward harnesses; see [JOB_EVIDENCE_PACKET s3](../JOB_EVIDENCE_PACKET.md)
Full receipts write-up[docs/INGAME_PROOF.md](../INGAME_PROOF.md)

why this matters

The in-game conditioning result is small, honest, and load-bearing for exactly that reason. The static -> conditional Brier improvement is a real calibration gain, most of it mechanical, with the model's own prior contributing a measured ~0.014 in NBA. Benchmarked against the live market, the conditional forecaster is sharper nowhere it is powered: the market wins the one resolvable checkpoint and ties everywhere else, the loss stated as plainly as any win. The hire signal is not the metric. It is a forecaster whose author scores it against the market, publishes the checkpoint where the market wins, and refuses to convert a real calibration improvement into an edge claim it did not earn.


edge_claimed = False everywhere. Every number is Brier / MAE / row-count vs a real out-of-sample corpus, never a dollar figure. Retracted measurement artifacts appear only in [JOB_EVIDENCE_PACKET.md](../JOB_EVIDENCE_PACKET.md), never on this page.


reproduce

reproduce on a fresh clone
# NBA + MLB in-game calibration scoreboard (real corpus if present)
python -m scripts.platformkit.ingame_scoreboard

# the per-sport proof harnesses directly
python -m scripts.platformkit.proof_nba.ingame_accuracy
python -m scripts.platformkit.proof_mlb.ingame_accuracy

# the leak property test (truncation invariance) -- runs on any clone
python -m pytest tests/test_ingame_leak_free.py -q
no edge claimed

This site reports calibration and sharpness only, never a dollar edge, ROI, or bankroll result. An honest null is a success.