evidence / Calibration & market
The Player-Projection Stack -- one accuracy claim, published under two labeled measurements
I built a 7-stat player-projection stack (points, rebounds, assists, threes, steals, blocks, turnovers) and published its accuracy under TWO explicitly-labeled measurements -- a public, re-runnable production holdout and an internal walk-forward OOF frame -- with a hard rule that their numbers are NEVER mixed, and a drift-guarded verify script that exits nonzero if the production numbers move. The label discipline is the product. The single truth-source for every figure below 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 claim
A projection MAE is meaningless unless you say exactly which measurement produced it. The same stack scored two legitimate ways gives two legitimately different numbers (on the walk-forward OOF frame BLK reads 0.515; on the production holdout it reads 0.44), and quoting one number with the other's row-count is the single most common way projection accuracy gets accidentally inflated. So the accuracy here ships as a labeled pair, each number bound to its own measurement basis, row count, and artifact.
receipts
| claim | committed artifact |
|---|---|
| 7-stat projection stack + production inference path | src/prediction/prop_pergame.py |
| Production-holdout verify script (0.02 drift-gated exit) | scripts/verify_production_mae.py |
| Train-time-columns fix for the 85-vs-129 crash | commit eb95e13b8 (2026-07-20) |
| Self-caught overfit + corrective regularization | src/prediction/prop_cv_split.py (lines ~185-186) |
| Artifact-drift assertion (`n_features_in_` == meta count) | scripts/audit_oof_prod_fidelity.py, scripts/backtest_holdout_wf.py |
| Both labeled measurements, with the never-mix rule | docs/JOB_EVIDENCE_PACKET.md section 3 |
why this matters
Anyone can report a projection MAE. The hire signal is that I report it twice, under two named measurements, refuse to mix their numbers, and back the discipline with a script that fails loudly when the production number drifts. That habit came from catching my own 0.79-vs-0.06 overfit -- so the label rule is not bureaucracy, it is the scar tissue from a real mistake, encoded so it cannot recur. A number you can trust is a number that tells you exactly how it was measured.
edge_claimed = False everywhere. Every figure is an MAE against a real out-of-sample corpus, never a dollar figure, and nothing here implies a projection beats the market. Retracted measurement artifacts appear only in [JOB_EVIDENCE_PACKET.md](../JOB_EVIDENCE_PACKET.md), never on this page.
reproduce
# re-score the production holdout and drift-check it against the claimed MAEs
python scripts/verify_production_mae.pyThis site reports calibration and sharpness only, never a dollar edge, ROI, or bankroll result. An honest null is a success.