Every paper-bet rule we have ever recorded. LIVE = actively recording new bets. RETIRED = known broken, kept recording as negative control. SUPERSEDED = replaced by a later rule_version. Diagnostic + audit trail only — does not influence any model, prediction, or bet decision.
| Rule version | Status | Bets | Settled | Wins | WR | PnL (u) | First | Last | Reason |
|---|---|---|---|---|---|---|---|---|---|
| v1_multibook_5ev_raw Multi-book consensus; ≥5%-EV gate; raw uncalibrated probabilities. | SUPERSEDED | 211,006 | 211,002 | 85,489 | 40.5% | -8416.2u | 2021-04-01 | 2026-05-06 | Replaced by v2 isotonic + Pinnacle-anchored variants 2026-05-06. |
| v4_draftkings_7pp_isotonic DraftKings opening; ≥7pp edge gate; isotonic-calibrated probabilities. | SUPERSEDED | 7,628 | 7,628 | 2,835 | 37.2% | -152.7u | 2021-04-09 | 2026-05-06 | Replaced by v5 tighter-gate variant 2026-05-06. |
| v5_draftkings_15pp_isotonic DraftKings opening; ≥15pp edge gate; isotonic; dogs+favs. | SUPERSEDED | 1,792 | 1,792 | 532 | 29.7% | +89.4u | 2021-04-09 | 2026-05-06 | Replaced by v6 DOGS_ONLY variant 2026-05-06 after favorites bled -7.59% ROI. |
| v6_dk_situational DraftKings opening; ≥15pp edge; isotonic; DOGS_ONLY. | RETIRED | 1,440 | 1,433 | 479 | 33.4% | +76.8u | 2021-04-10 | 2026-05-25 | Live 0 wins / 43 same-day bets (Apr-May 2026). Recorder docstring admits gate=15pp + book=DK were tuned on the same data (in-sample). The 2024 backfill +73u headline doesn't generalize — DK has tightened openings since. Per Dave 2026-05-25: keep recording as negative control, but treat as RETIRED for any decision-making purpose. |
| v2_pinnacle_7pp_raw Pinnacle opening; ≥7pp edge gate; raw uncalibrated probabilities. | SUPERSEDED | 497 | 496 | 197 | 39.7% | -46.7u | 2026-04-17 | 2026-05-06 | Replaced by v3 isotonic variant 2026-05-06. |
| v3_pinnacle_7pp_isotonic Pinnacle opening; ≥7pp edge gate; isotonic-calibrated probabilities. | SUPERSEDED | 190 | 190 | 67 | 35.3% | -23.3u | 2026-04-17 | 2026-05-06 | Replaced by v4 DraftKings variant 2026-05-06. |
The smoking gun from the v27 audit. The backfilled cohort (bets inserted >7 days after the game) shows the +118u headline. The same-day live cohort shows the real strategy performance: zero wins out of 43 bets. Probability of 0 wins in 43 if true rate were 30% ≈ 4 in a million. Both cohorts use the same model + same walk-forward calibration, so the disconnect is not a calibration bug — it is the in-sample tuning of the 15pp gate + DK book (admitted in the recorder docstring).
| Cohort | Bets | Wins | WR | PnL (u) |
|---|---|---|---|---|
| backfilled (>7 days after) | 1,387 | 477 | 34.4% | +118.3u |
| same-day (live) | 43 | 0 | —% | -43.0u |
| within 7 days after | 3 | 2 | 66.7% | +1.6u |