The Retraction Story -- I built the instruments that refuted my own headline numbers
The most transparent sports forecaster you can audit -- every prediction pre-registered, every number gated, including the ones that refuted me. The single truth-source for any figure below is docs/JOB_EVIDENCE_PACKET.md; the retracted numbers appear here only inside explicit retraction framing, exactly as the packet handles them.
The same person who built this system also built the instruments that refuted his own
flagship numbers -- and documented the negative results in writing rather than quietly
deleting them. The product is a calibrated predictor, not a betting-edge product. The
strongest signal in the repo is not any metric; it is the self-refutation trail.
---
The single truth-source for every figure below is docs/JOB_EVIDENCE_PACKET.md. The retracted numbers appear here, and only here, inside this retraction framing.
six headlines my own harnesses took apart
retracted
+18.38% pregame ROI on 1,535 walk-forward bets vs real closing lines
what was wrong
Market-follow artifact, confirmed at the source-code level. The grader picked bet direction from the market's own devigged lean and never read the model (the eval CSV had no prediction column), priced at a flat -110 that real books do not offer, and tuned its filters in-sample on the same file.
honest replacement
Roughly break-even-minus-vig vs real closing lines. Every candidate edge, including assists, was ultimately rejected or retracted by my own gates.
proofJOB_EVIDENCE_PACKET s4 (model's own unfiltered number: -2.00%)
Leak-inflated and mis-sourced. Two features were computed from fourth-quarter data, so the model predicting Q4 was peeking at Q4; the cited file actually reported 0.1354, a different number.
honest replacement
Leak-free walk-forward end-of-Q3 Brier ~0.141, after I removed a Q4 feature leak I found in my own pipeline. Framed as a leak I caught, not a competitive number.
Graded against an L5 line proxy, not real closing lines. A model-quality ceiling on a soft proxy, never a tradeable result.
honest replacement
On a soft L5 proxy the in-play backtest reaches that ceiling. I treat it strictly as a model-quality ceiling, never as realized edge.
proofJOB_EVIDENCE_PACKET s4
retracted
Aggregate CLV +8.94pp
what was wrong
Circular -- computed on the same model-unused, devig-direction corpus. No real Pinnacle-close CLV exists yet; a full-season backtest shows CLV about zero vs real closes.
honest replacement
Real closing-line CLV cannot be measured yet. I built the methodology that will measure it; I do not quote a CLV figure until it can be.
proofJOB_EVIDENCE_PACKET s4 (full-season backtest: CLV ~= 0 vs real closes)
retracted
Steals/blocks prop grid search: training R^2 ~0.79
what was wrong
Textbook leakage: the ~0.79 training R^2 collapsed to ~0.06 on a leak-free holdout.
honest replacement
Caught and hard-corrected a leakage-driven overfit. The corrective regularization takes precedence over the stale tuned parameters, so the mistake cannot silently reappear.
proofsrc/prediction/prop_cv_split.py (documents the gap; hard-codes corrective regularization)
retracted
The assists ROI edge (the strongest surviving candidate)
what was wrong
Regime-dependent -- it broke in the playoffs -- and retracted 2026-07-21. Under the no-edge rail, no dollar/ROI edge is claimed anywhere.
honest replacement
No dollar or ROI edge is claimed anywhere. The historical measurement remains only as a record of the stress-testing methodology.
proofJOB_EVIDENCE_PACKET s3 (historical record only, in the gate artifacts)
The through-line: against real closing lines the market is efficient, the model is break-even-minus-vig, and every candidate edge, including my strongest, was rejected or retracted by my own gates. That is the honest, correct result for an efficient market, and I have the harnesses that prove it.
