# Luna deep analysis — slate 20260901 > Coverage receipt: **COMPLETE** · 9/8 sections; 15/15 games; 8/8 causes; 6/6 sources # 1. Evidence receipt and gaps ## Sources received All seven manifest entries were checked against the supplied evidence: | Source | Path | Size | Status | Use | |---|---|---:|---|---| | Prediction review | `/app/output/reports/prediction_review_20260901.json` | 373,178 bytes | Loaded | Official outcomes and scoring errors | | Forensic deep audit | `/app/output/reports/PREDICTION_SYSTEM_DEEP_AUDIT_20260901.json` | 240,806 bytes | Loaded | Probability, provenance, starter, and decision contracts | | Continuous learning | `/app/output/learning/continuous_learning_daily_20260901.json` | 1,939,278 bytes | Loaded | Stage quality, calibration, lineage, score distribution, starter comparisons | | Detailed report | `/app/output/reports/detailed_20260901.html` | 355,722 bytes | Loaded | Full game and simulation rendering | | Ultra-detailed player audit | `/app/output/reports/20260901_ultra_detailed_player_audit.html` | 6,288,497 bytes | Loaded | Player-level prediction trace | | Structural audit | `/app/output/reports/audit_report_20260901.html` | 383,197 bytes | Loaded | Invariant findings | | Integrity revision audit | `/app/output/reports/audit_report_20260901_integrity_revision.html` | Not supplied | **Missing** | Post-remediation proof for immutable revision artifacts | The missing integrity-revision audit is the principal evidence gap. It does not, by itself, prove a prediction error, but it prevents a fully closed artifact-revision audit. ## Slate coverage The evidence reports: - Expected games: 15. - Games reviewed: 15. - Continuous-learning games: 15. - Forensic deep-audit games: 15. - Structural audits completed: 15. - Starter comparisons: 15. - Canonical artifacts loaded: 15. - Settled games: 15. - Hits: 7. - Misses: 8. - Learning eligibility: eligible because all game artifacts passed critical/error gates. The 15 reviewed games were: `ATL@WAS`, `BAL@COL`, `CHW@HOU`, `DET@MIN`, `MIA@KC`, `MIL@CHC`, `NYM@TB`, `NYY@LAA`, `OAK@TEX`, `PHI@ARI`, `SD@CIN`, `SEA@BOS`, `SF@PIT`, `STL@LAD`, and `TOR@CLE`. ## Internal contradictions and gaps 1. **Missing integrity-revision artifact.** The manifest explicitly marks one source as missing. 2. **One material probability-lineage defect.** `TOR@CLE` states that the publication delta has “no recorded transformation.” The delta is published probability 0.555236 versus governed/calibrated probability 0.545, a +0.010236 movement. This is a lineage/contract defect even though it did not change the winner. 3. **One structural warning.** `STL@LAD` has a pitch-count mismatch for Michael McGreevy: summary BF 24 versus box-score BF 23. 4. **Aggregate audit contradiction requiring care.** The slate-level structural summary reports zero critical and zero error findings, one warning, and 108 informational findings. That is consistent with the per-game records supplied, but “all artifacts passed critical/error gates” does not mean the slate is artifact-perfect: the warning, unexplained TOR probability delta, and missing integrity-revision source remain. 5. **Calibration-label inconsistency.** The calibration object defines error as average published confidence minus observed hit rate, with positive meaning overconfidence. The 0.50–0.55 and 0.55–0.60 buckets are labeled “underconfident” despite negative errors, which is directionally consistent with the definition. The high buckets have zero hits and are clearly overconfident. However, the displayed bucket errors for the high buckets equal average confidence rather than a separately rounded confidence-minus-hit-rate field; this is mathematically expected when observed hit rate is zero, not necessarily a defect. 6. **Starter and player data are retrospective comparisons.** They support diagnosis but do not establish that a pregame input was wrong unless the relevant pregame signal and transformation are shown. --- # 2. Executive diagnosis ## Overall slate result The system went **7–8, or 46.67% winner accuracy**. The continuous-learning summary reports: - Winner accuracy: 0.4667, 7 hits. - Winner Brier score: 0.3364 is not supplied slate-wide, but per-game Brier values are supplied. - Historical walk-forward accuracy: 0.566181. - Historical walk-forward Brier: 0.243016. - Model-validation winner accuracy: 0.592233. - Model-validation winner Brier: 0.238561. - Weighted absolute calibration error on this slate: **0.2476**. - Mean team-run error: materially poor, with the `enhanced_prediction` stage the worst measured stage. - Exact run-model holdout fit: weak, with home/away/total R² of **0.0361 / 0.0227 / 0.0280** over 2,062 rows. A correct winner is not equivalent to a good prediction. The slate contains several correct winners with large score errors, including `SF@PIT` and `CHW@HOU`. ## Winner accuracy Winner accuracy was poor relative to the model’s historical validation: - 7 of 15 winners correct. - 8 misses. - Four high-raw-probability selections lost: `BAL@COL`, `MIL@CHC`, `SD@CIN`, `STL@LAD`. - The high-raw cluster was 0–4 despite raw probabilities of 0.6793–0.7934 and published pick probabilities of 0.58–0.7461. The main winner-direction problem was not uniformly one thing: - Starter projection was listed as the primary driver in four losses: `BAL@COL`, `MIA@KC`, `SD@CIN`, and `SEA@BOS`. - Run-environment miss was primary in two: `DET@MIN` and `STL@LAD`. - Small-edge losing-side miss was primary in `ATL@WAS`. - Wrong conviction on the selected side was explicitly secondary or primary in `MIA@KC`, `MIL@CHC`, `SD@CIN`, and `STL@LAD`. ## Calibration quality Calibration was the clearest probability-quality failure. - Weighted absolute calibration error: **0.2476 across 15 games**. - 0.50–0.55 bucket: average confidence 0.5352, observed hit rate 0.625, error −0.0898. - 0.55–0.60 bucket: average confidence 0.5651, observed hit rate 0.6667, error −0.1015. - 0.60–0.65 bucket: average confidence 0.6232, 0 hits in 2 games. - 0.65–0.70 bucket: average confidence 0.6986, 0 hits in 1 game. - 0.70–0.80 bucket: average confidence 0.7461, 0 hits in 1 game. The low-confidence buckets were underconfident on this sample, while the higher-confidence buckets were severely overconfident. The evidence supports a bounded shadow adjustment, not an immediate permanent recalibration, because the high-bucket sample is only four games and the causal register correctly warns that a single slate is too small to establish a stable calibration curve. ## Score accuracy Score accuracy was materially poor and structurally biased toward narrow margins: - 13 of 15 predicted margins were 2 runs or fewer. - 9 of 15 actual margins were 4 or more. - Mean predicted margin: 1.67. - Mean actual margin: 4.20. - Actual margins ranged from 1 to 13. - The largest miss was `DET@MIN`: predicted margin 1, actual margin 13. - The `enhanced_prediction` stage had the largest team-run regression versus historical ML: - Team-run MAE: 3.2667 versus 3.0977 historical ML. - Team-run MAE delta: +0.169. - Margin MAE: 4.9333 versus 4.1769. - Winner accuracy: 6/15, or 40%. The `score_margin_compression` causal issue is therefore materially supported and has severity **high**, not critical. The do-not-infer guardrail remains important: one blowout is not enough to widen all scores. Here, however, the distributional evidence spans nine actual margins of at least four and two consecutive slates in the supplied trend record. ## Starter accuracy Starter projections were a significant source of score and winner error: - Starter errors were cited in four losses as the primary driver. - The starter comparison fields are present for all 15 games. - Some large errors were likely variance: - `SEA@BOS`: Jedixson Paez +6 ER over projection. - `SF@PIT`: Logan Webb +7 ER and Paul Skenes +4 ER. - `MIA@KC`: Randy Dobnak +5 ER and Tyler Phillips −4 ER. - Some evidence supports underweighted context rather than pure variance: - `MIA@KC`: Dobnak’s home ERA was 2.945 versus 1.141 on the other side, with a credible venue split. - `MIL@CHC`: Robert Gasser’s away ERA was 5.595 versus 4.05 on the other side, with a credible venue split. The lever-tension record says the pitcher-weight lever “helped” 10 games and hurt none by the supplied attribution, but that is a descriptive slate count, not a causal estimate. ## Artifact integrity Artifact integrity was mostly sound but not clean: - Structural summary: 0 critical, 0 error, 1 warning, 108 info. - Fourteen games had structural-audit status PASS. - `STL@LAD` had status WARNING because of a pitch-count discrepancy. - `TOR@CLE` had an unexplained +0.010236 publication delta. - The integrity-revision audit was missing. - All games were nevertheless marked learning-eligible because no critical/error structural gate failed. The high-severity `artifact_integrity_warning` issue is appropriately preserved: the warning is a presentation or contract risk, not proof that the model chose the wrong side. ## Causal issue register reconciliation | Issue | Severity | Evidence and affected games | Action | Guardrail | |---|---|---|---|---| | Canonical artifact warning requires reconciliation | High | 1 warning, 108 info; no critical/error. No specific game list. | Fix the flagged stat/presentation contract before considering the slate clean. | Does not prove a wrong-side pick. | | Published margins compressed | High | 13/15 predicted margins ≤2; 9/15 actual margins ≥4; mean margin 1.67 versus 4.20. Affected games listed as `CHW@HOU`, `DET@MIN`, `MIL@CHC`, `NYM@TB`, `PHI@ARI`, `STL@LAD`, `TOR@CLE`. | Check rolling-slate tail coverage and margin MAE before widening score spread. | Do not widen every score because of one blowout. | | Single-slate calibration instability | High | WAC error 0.2476 across all 15 games. | Monitor the same buckets over the rolling window before changing calibration weights. | Fifteen games do not establish a stable calibration curve. | | Weak run-model signal | High | Home/away/total R² 0.0361/0.0227/0.0280 over 2,062 rows. | Investigate features and target via isolated training backtest. | Positive R² does not make exact scores reliable. | | High raw-probability cluster | Medium | 0/4 high-raw selections won. | Track across multiple slates; preserve raw→calibrated→published lineage. | Do not create a new bucket-shrinkage rule from one slate. | | Complete slate audit coverage | Info | 15/15 structurally audited. | Keep per-game audit receipts. | Coverage does not cancel internal findings. | | Publication delta explained | Info | 9 material movements had named transformations. | Continue displaying every layer and transformation. | Published > calibrated is not a bypass when explained. | | Simulation inert on slate | Info for current slate; escalated across slates | 0 flip candidates and 0 published flips. | Measure simulation contribution over rolling slates. | No winner flips does not prove no score reconciliation occurred. | --- # 3. Every game Probabilities below are **published probability for the selected/predicted side**. The signed confidence edge is `home probability − away probability`; therefore negative values indicate an away selection. ## ATL@WAS — Miss; wrong winner and wrong score - **Prediction:** ATL 7, WAS 5; winner ATL. - **Actual:** ATL 5, WAS 9; winner WAS. - **Published selected-side probability:** 0.5317 for ATL. Home-minus-away edge: −0.0634. - **Bucket:** Lean. - **Errors:** Brier 0.282705; team-run absolute error 6; total bias −2; predicted margin 2 versus actual margin 4. - **Starter comparison:** - AJ Smith-Shawver: projected 3 ER/13 outs; actual 5 ER/12 outs; +2 ER, −1 out. - Jake Irvin: projected 2 ER/15 outs; actual 2 ER/12 outs; 0 ER, −3 outs. - **Edges:** Season −0.6015; recent −0.318; pitcher −0.043; composite −0.865 home-minus-away. - **Controls:** Governor applied; raw pick probability 0.4969; governed 0.5082; recency guardrail applied; score coherence plus matchup-recency guardrail; volatility action no change. - **Score path:** Historical ML 4.77–4.83; variance 5–5; enhanced 6–5; Savant 7–5; later stages unchanged. - **Player adjustments:** Material changes 24. Positive adjustments included Drake Baldwin RBI +2.24, Ronald Acuña Jr. hits +1.82, Matt Olson RBI +1.42, and Brady House RBI +1.34. Washington also had deboosts to CJ Abrams, Abimelec Ortiz, Yohandy Morales, and Harry Ford. - **Audit:** Structural PASS, risk 0, no defects, seven informational notes. Forecast-file integrity clean. - **Primary cause:** Lean-side miss. - **Secondary cause:** None formally assigned. - **Forensics:** WAS offense was under-projected by 4. Smith-Shawver’s +2 ER was classified as an outlier start/variance; no credible venue or opponent-strength explanation was found. ## BAL@COL — Miss; wrong winner and wrong score - **Prediction:** BAL 7, COL 4; winner BAL. - **Actual:** BAL 2, COL 4; winner COL. - **Published selected-side probability:** 0.746135 for BAL. Home-minus-away edge: −0.4923. - **Bucket:** Strong. - **Errors:** Brier 0.556717; team-run error 5; total bias +5; predicted margin 3 versus actual margin 2. - **Starters:** - Kyle Bradish: projected 3 ER/16 outs; actual 2 ER/15 outs; −1 ER, −1 out. - Gabriel Hughes: projected 5 ER/13 outs; actual 1 ER/14 outs; −4 ER, +1 out. - **Edges:** Season −1.0874; recent −1.418; pitcher −0.0404; composite −1.5815. - **Controls:** Raw 0.736061; governed 0.7373; no recency guardrail; score-coherence movement to published 0.746135; supported blowout-margin expansion. - **Score path:** Historical 5.44–4.39; variance 5–4; enhanced 7–4; Savant and later stages 7–4. - **Player adjustments:** Material changes 24. Gunnar Henderson hits +1.95, Brett Sullivan hits +1.28, Jackson Holliday hits +1.27, Pete Alonso AB +1.244, and Colton Cowser AB +1.24 were among the largest. - **Audit:** Structural PASS, risk 0, forecast artifact clean. - **Primary cause:** Starter projection. - **Secondary cause:** Run environment over-projected. - **Forensics:** Hughes allowed four fewer ER than projected. The evidence suggests possible opponent-strength weighting relevance, but the supplied synthesis recommends validation rather than a permanent change. ## CHW@HOU — Hit; correct winner, wrong score - **Prediction:** CHW 6, HOU 5; winner CHW. - **Actual:** CHW 5, HOU 1; winner CHW. - **Published selected-side probability:** 0.53 for CHW. Home-minus-away edge: −0.0600. - **Bucket:** Lean. - **Errors:** Brier 0.2209; team-run error 5; total bias +5; predicted margin 1 versus actual margin 4. - **Starters:** - Sean Burke: projected 4 ER/15 outs; actual 0 ER/17 outs; −4 ER, +2 outs. - Ronel Blanco: projected 4 ER/13 outs; actual 0 ER/12 outs; −4 ER, −1 out. - **Edges:** Season −0.363; recent +0.283; pitcher −0.0483; composite −0.4152. - **Controls:** Raw 0.53; governed 0.5082; recency guardrail applied; score coherence/recency path; volatility no change. - **Score path:** Historical 4.84–5.00; variance 5–5; enhanced 3–5; Savant and later stages 3–5, although the supplied `predicted` score is 6–5, indicating a stored score-path versus published-score inconsistency that should be reconciled. - **Player adjustments:** No player-level forensic record was available. - **Audit:** Structural PASS, risk 0. - **Finding:** Correct winner/wrong score. - **Cause:** No formal primary driver; both starters materially outperformed ER projections, and HOU’s offense was over-projected. ## DET@MIN — Miss; wrong winner and largest score miss - **Prediction:** DET 4, MIN 3; winner DET. - **Actual:** DET 2, MIN 15; winner MIN. - **Published selected-side probability:** 0.53 for DET. Home-minus-away edge: −0.0600. - **Bucket:** Lean. - **Errors:** Brier 0.2809; team-run error 14; total bias −10; predicted margin 1 versus actual margin 13. - **Starters:** - Troy Melton: projected 2 ER/18 outs; actual 5 ER/16 outs; +3 ER, −2 outs. - Andrew Morris: projected 1 ER/14 outs; actual 1 ER/3 outs; 0 ER, −11 outs. - **Edges:** Season −0.2973; recent +0.534; pitcher −0.0249; composite −0.2648. - **Controls:** Raw 0.53; governed 0.505; recency guardrail; enhanced stage explicitly shrank confidence because of a one-run margin and recency controls; no volatility adjustment. - **Score path:** Historical 5.07–3.72; variance 5–4; enhanced 4–3; Savant and final stages 4–3. - **Player adjustments:** Material changes 25. Large adjustments included Kody Clemens strikeout +2.075, John Peck RBI deboost −1.86, Walker Jenkins RBI −1.47, and Troy Melton projected ER adjustment −1.0. - **Audit:** Structural PASS, risk 0. - **Primary cause:** Run environment under-projected. - **Secondary cause:** None formally assigned. - **Forensics:** MIN scored 12 more than projected as a team. Melton allowed +3 ER, while the five-run ninth-inning-type bullpen contribution was also important: Kyle Finnegan allowed 5 ER versus 0 projected. The model treated the realized explosion as variance rather than a demonstrated persistent input failure. ## MIA@KC — Miss; wrong winner and wrong score - **Prediction:** MIA 3, KC 5; winner KC. - **Actual:** MIA 6, KC 3; winner MIA. - **Published selected-side probability:** 0.646375 for KC. Home-minus-away edge: +0.2928. - **Bucket:** Strong. - **Errors:** Brier 0.417801; team-run error 5; total bias −1; predicted margin 2 versus actual margin 3. - **Starters:** - Tyler Phillips: projected 4 ER/15 outs; actual 0 ER/15 outs; −4 ER. - Randy Dobnak: projected 1 ER/17 outs; actual 6 ER/18 outs; +5 ER, +1 out. - **Edges:** Season −0.2701; recent +0.648; pitcher +0.052; composite +0.4878. - **Controls:** Raw 0.642335; governed 0.6227; no recency guardrail; score coherence raised published probability to 0.646375; no volatility action. - **Score path:** Historical 4.84–5.00; variance 5–5; enhanced 3–5; Savant and final stages 3–5. - **Player adjustments:** Material changes 21. Graham Pauley strikeout +2.341; Salvador Perez AB −1.89; Carter Jensen AB −1.73; Bobby Witt Jr. AB −1.65; Heriberto Hernández RBI +1.43. - **Audit:** Structural PASS, risk 0. - **Primary cause:** Starter projection. - **Secondary cause:** Wrong conviction side. - **Forensics:** Dobnak’s credible venue split showed materially worse home performance: 2.945 ERA at home versus 1.141 on the other side. The supplied recommendation is to enable or raise `context.venue_adjustment_weight` in shadow testing. ## MIL@CHC — Miss; wrong winner and wrong score - **Prediction:** MIL 6, CHC 8; winner CHC. - **Actual:** MIL 9, CHC 4; winner MIL. - **Published selected-side probability:** 0.698612 for CHC. Home-minus-away edge: +0.3972. - **Bucket:** Strong. - **Errors:** Brier 0.488059; team-run error 7; total bias +1; predicted margin 2 versus actual margin 5. - **Starters:** - Robert Gasser: projected 5 ER/15 outs; actual 2 ER/13 outs; −3 ER, −2 outs. - Matthew Boyd: projected 4 ER/18 outs; actual 3 ER/16 outs; −1 ER, −2 outs. - **Edges:** Season −0.0089; recent +0.482; pitcher −0.023; composite +0.425. - **Controls:** Raw 0.914; calibrated/governed 0.6442 due to sample size 10; no recency guardrail; score coherence yielded 0.698612; supported blowout-margin expansion. - **Score path:** Historical 4.47–4.63; variance 4–5; enhanced 5–8; Savant 6–8; final unchanged. - **Player adjustments:** Material changes 21. Tyrone Taylor RBI +3.37 and Jackson Chourio RBI +2.28 were large positive adjustments; Seiya Suzuki AB −2.24 and Christian Yelich strikeout +1.86 were also substantial. - **Audit:** Structural PASS, sparse historical bucket diagnostic flag. - **Primary cause:** Wrong conviction side. - **Secondary cause:** Small-edge/side error despite a strong published selection. - **Forensics:** Both starters outperformed ER projections, especially Gasser. His credible away split was materially worse than his other-side ERA, suggesting underweighted venue-role information. The game was also a high-raw-probability cluster loss. ## NYM@TB — Hit; correct winner, wrong score - **Prediction:** NYM 5, TB 6; winner TB. - **Actual:** NYM 2, TB 6; winner TB. - **Published selected-side probability:** 0.526945 for TB. Home-minus-away edge: +0.0539. - **Bucket:** Lean. - **Errors:** Brier 0.223781; team-run error 3; total bias +3; predicted margin 1 versus actual margin 4. - **Starters:** - Sean Manaea: projected 2 ER/18 outs; actual 5 ER/17 outs; +3 ER, −1 out. - Freddy Peralta: projected 4 ER/16 outs; actual 1 ER/18 outs; −3 ER, +2 outs. - **Edges:** Season +0.7945; recent +0.086; pitcher +0.1335; composite +0.9322. - **Controls:** Raw 0.53; governed 0.5082; recency guardrail; score coherence and recency path; no volatility action. - **Score path:** Historical 4.83–5.00; variance 5–5; enhanced 5–5; Savant/final 5–6. - **Player adjustments:** No player-level forensic record was available. - **Audit:** Structural PASS. - **Finding:** Correct winner/wrong score. - **Cause:** No formal primary driver; opposing starter errors offset in winner direction while NYM offense was over-projected. ## NYY@LAA — Hit; correct winner, wrong score - **Prediction:** NYY 6, LAA 4; winner NYY. - **Actual:** NYY 7, LAA 3; winner NYY. - **Published selected-side probability:** 0.54 for NYY. Home-minus-away edge: −0.0800. - **Bucket:** Lean. - **Errors:** Brier 0.2116; team-run error 2; total bias 0; predicted margin 2 versus actual margin 4. - **Starters:** - Gerrit Cole: projected 1 ER/18 outs; actual 3 ER/16 outs; +2 ER, −2 outs. - Grayson Rodriguez: projected 3 ER/15 outs; actual 4 ER/18 outs; +1 ER, +3 outs. - **Edges:** Season −1.2252; recent −0.490; pitcher −0.0343; composite −1.2318. - **Controls:** Raw 0.54; governed 0.5332; recency guardrail; score coherence and recency path; no volatility action. - **Score path:** Historical stage not numerically displayed beyond final trace; published score 6–4. - **Player adjustments:** No player-level forensic record was available. - **Audit:** Structural PASS. - **Finding:** Correct winner/wrong score, but relatively accurate runs. - **Cause:** No formal primary or secondary driver. ## OAK@TEX — Hit; correct winner, wrong score - **Prediction:** OAK 4, TEX 5; winner TEX. - **Actual:** OAK 5, TEX 8; winner TEX. - **Published selected-side probability:** 0.548725 for TEX. Home-minus-away edge: +0.0974. - **Bucket:** Lean. - **Errors:** Brier 0.203649; team-run error 4; total bias −4; predicted margin 1 versus actual margin 3. - **Starters:** - Brady Basso: projected 4 ER/13 outs; actual 2 ER/7 outs; −2 ER, −6 outs. - MacKenzie Gore: projected 3 ER/16 outs; actual 1 ER/16 outs; −2 ER. - **Edges:** Season +0.9405; recent +0.242; pitcher +0.0026; composite +0.7635. - **Controls:** Raw 0.5517; governed 0.5499; recency guardrail; published probability essentially matched governed value; no volatility action. - **Score path:** Published 4–5; detailed intermediate numeric path is not separately supplied. - **Player adjustments:** No player-level forensic record was available. - **Audit:** Structural PASS. - **Finding:** Correct winner/wrong score. - **Cause:** TEX offense under-projected by 3; total under-projected by 4. ## PHI@ARI — Hit; correct winner, wrong score - **Prediction:** PHI 4, ARI 3; winner PHI. - **Actual:** PHI 7, ARI 1; winner PHI. - **Published selected-side probability:** 0.560194 for PHI. Home-minus-away edge: −0.1204. - **Bucket:** Moderate. - **Errors:** Brier 0.193429; team-run error 5; total bias −1; predicted margin 1 versus actual margin 6. - **Starters:** - Jesús Luzardo: projected 1 ER/20 outs; actual 1 ER/20 outs; exact ER and outs. - Eduardo Rodríguez: projected 1 ER/18 outs; actual 3 ER/15 outs; +2 ER, −3 outs. - **Edges:** Season −0.3931; recent −0.648; pitcher −0.0855; composite −0.7463. - **Controls:** Raw 0.5884; governed 0.5816; recency guardrail; score coherence/recency path; no volatility action. - **Score path:** Published 4–3; intermediate numeric stages not separately supplied. - **Player adjustments:** No player-level forensic record was available. - **Audit:** Structural PASS. - **Finding:** Correct winner/wrong score. - **Cause:** PHI offense under-projected by 3; Rodríguez also allowed two more ER than projected. ## SD@CIN — Miss; wrong winner and wrong score - **Prediction:** SD 7, CIN 4; winner SD. - **Actual:** SD 3, CIN 4; winner CIN. - **Published selected-side probability:** 0.60 for SD. Home-minus-away edge: −0.2000. - **Bucket:** Strong. - **Errors:** Brier 0.36; team-run error 4; total bias +4; predicted margin 3 versus actual margin 1. - **Starters:** - Randy Vásquez: projected 3 ER/15 outs; actual 2 ER/14 outs; −1 ER, −1 out. - Nick Lodolo: projected 6 ER/12 outs; actual 2 ER/18 outs; −4 ER, +6 outs. - **Edges:** Season −0.8428; recent −0.699; pitcher +0.0764; composite −1.0102. - **Controls:** Raw 0.60; governed 0.6778, based on a raw/governor path with recency guardrail; published 0.60 after score coherence and matchup-recency guardrail; supported blowout-margin expansion. - **Score path:** Historical 4.46–3.81; variance 4–4; enhanced 6–4; Savant/final 7–4. - **Player adjustments:** Material changes 23. Xander Bogaerts RBI +3.53, Héctor Rodríguez RBI +2.55, Ty France RBI +2.26, and several SD lineup adjustments raised the selected-side projection. - **Audit:** Structural PASS. - **Primary cause:** Starter projection. - **Secondary cause:** Wrong conviction side. - **Forensics:** Lodolo allowed four fewer ER and six more outs than projected. The retrospective synthesis calls this variance and does not recommend a model change. ## SEA@BOS — Miss; wrong winner and wrong score - **Prediction:** SEA 3, BOS 5; winner BOS. - **Actual:** SEA 9, BOS 6; winner SEA. - **Published selected-side probability:** 0.53 for BOS. Home-minus-away edge: +0.0600. - **Bucket:** Lean. - **Errors:** Brier 0.2809; team-run error 7; total bias −7; predicted margin 2 versus actual margin 3. - **Starters:** - Bryan Woo: projected 1 ER/17 outs; actual 1 ER/18 outs; exact ER, +1 out. - Jedixson Paez: projected 2 ER/9 outs; actual 8 ER/11 outs; +6 ER, +2 outs. - **Edges:** Season +0.9535; recent +0.551; pitcher −0.0054; composite +1.0103. - **Controls:** Raw 0.53; governed 0.6273; recency guardrail and multiple confidence-shrink controls; published 0.53 after score coherence/recency; supported blowout-margin expansion. - **Score path:** Historical 6.19–0.95; variance 6–1; enhanced 3–5; Savant/final 3–5. - **Player adjustments:** Material changes 24. Lazaro Montes RBI +2.554, Caleb Durbin RBI +1.63, Mickey Gasper RBI +1.556, and multiple SEA lineup deboosts. - **Audit:** Structural PASS. - **Primary cause:** Starter projection. - **Secondary cause:** Run environment under-projected. - **Forensics:** Paez allowed six more ER than projected. The synthesis classifies this as an outlier start/variance; no credible venue or opponent-strength signal explains it. ## SF@PIT — Hit; correct winner, very wrong score - **Prediction:** SF 5, PIT 6; winner PIT. - **Actual:** SF 12, PIT 13; winner PIT. - **Published selected-side probability:** 0.544522 for PIT. Home-minus-away edge: +0.0890. - **Bucket:** Lean. - **Errors:** Brier 0.20746; team-run error 14; total bias −14; predicted margin 1 versus actual margin 1. - **Starters:** - Logan Webb: projected 2 ER/18 outs; actual 9 ER/6 outs; +7 ER, −12 outs. - Paul Skenes: projected 1 ER/16 outs; actual 5 ER/14 outs; +4 ER, −2 outs. - **Edges:** Season +0.4402; recent −0.3527; pitcher −0.1715; composite −0.0434. - **Controls:** Raw 0.559024; governed 0.56; recency guardrail; score coherence/recency path; no volatility action. - **Score path:** Published 5–6; intermediate numeric path not separately supplied. - **Player adjustments:** No player-level forensic record was available. - **Audit:** Structural PASS. - **Finding:** Correct winner/wrong score. - **Cause:** Both offenses and both starters were far outside the score projection. The winner survived a severe run-environment miss. ## STL@LAD — Miss; wrong winner and wrong score - **Prediction:** STL 5, LAD 7; winner LAD. - **Actual:** STL 13, LAD 8; winner STL. - **Published selected-side probability:** 0.58 for LAD. Home-minus-away edge: +0.1600. - **Bucket:** Moderate. - **Errors:** Brier 0.3364; team-run error 9; total bias −9; predicted margin 2 versus actual margin 5. - **Starters:** - Michael McGreevy: projected 3 ER/17 outs; actual 4 ER/15 outs; +1 ER, −2 outs. - Eric Lauer: projected 5 ER/16 outs; actual 7 ER/9 outs; +2 ER, −7 outs. - **Edges:** Season +1.1307; recent +0.5433; pitcher +0.0064; composite +1.1318. - **Controls:** Raw 0.58; governed 0.6696; recency guardrail; score coherence/recency guardrail yielded published 0.58; supported blowout-margin expansion. - **Score path:** Historical 4.21–4.42; variance 4–4; enhanced 4–6; Savant 5–7; final 5–7. - **Player adjustments:** Material changes 27. Tommy Edman RBI +4.554, José Fermín RBI −2.638, Alex Freeland AB +2.214, Kyle Tucker hits +2.18, and several LAD adjustments. - **Audit:** **WARNING**, risk 10. Michael McGreevy pitch-count row had summary BF 24 versus box BF 23. One structural defect/warning was reported in the game-level diagnostic. - **Primary cause:** Run environment under-projected. - **Secondary cause:** Wrong conviction side. - **Forensics:** STL offense was under-projected by 8. Lauer allowed +2 ER, and the player-level results included multiple unexpected STL home runs. The synthesis labels the starter miss variance. ## TOR@CLE — Hit; correct winner, wrong score; lineage defect - **Prediction:** TOR 4, CLE 5; winner CLE. - **Actual:** TOR 1, CLE 6; winner CLE. - **Published selected-side probability:** 0.555236 for CLE. Home-minus-away edge: +0.1105. - **Bucket:** Moderate. - **Errors:** Brier 0.197815; team-run error 4; total bias +2; predicted margin 1 versus actual margin 5. - **Starters:** - Spencer Miles: projected 2 ER/15 outs; actual 0 ER/3 outs; −2 ER, −12 outs. - Gavin Williams: projected 3 ER/16 outs; actual 0 ER/21 outs; −3 ER, +5 outs. - **Edges:** Season +0.4158; recent +0.3953; pitcher −0.1167; composite +0.0708. - **Controls:** Raw 0.548941; governed 0.545; no recency guardrail; score coherence probability 0.564983; published 0.555236. - **Score path:** Enhanced-to-fundamental transition changed the side from tie to CLE and improved winner accuracy. - **Player adjustments:** No player-level forensic record was available. - **Audit:** Structural PASS at the game level, but probability lineage contains a defect. - **Finding:** Correct winner/wrong score. - **Cause:** TOR offense over-projected by 3 and Williams allowed three fewer ER than projected. - **Lineage issue:** The +0.010236 published-minus-calibrated movement has no recorded transformation. This is an artifact/contract defect, not evidence that the winner was wrong. --- # 4. Miss reconstruction ## Earliest wrong-direction stage for each loss A later recovery cannot erase an intermediate regression. The stage audit must therefore distinguish final outcome from path quality. ### ATL@WAS - Historical ML selected WAS by the underlying home-minus-away direction in the transition record (`WAS → TIE` after variance). - `variance_applied` moved the direction from WAS toward a tie and cost one winner hit overall; `ATL@WAS` is explicitly listed as regressed. - The published pick ultimately favored ATL only weakly. - **Classification:** Small-edge side miss; likely winner-direction instability plus realized WAS offense. Starter variance contributed but was not the formal primary cause. ### BAL@COL - The earliest harmful information was already present in the historical/quality edge: strong BAL superiority. - The enhanced stage expanded BAL from roughly 5–4 to 7–4. - **Classification:** Bad weighting or incomplete context, especially Hughes/opponent-strength and venue-role response; then starter error and run-environment overprojection. The loss was not caused by a late winner flip. ### DET@MIN - Historical ML was near-neutral and variance moved the game toward DET/TIE dynamics. - Enhanced prediction reduced the projected score to DET 4, MIN 3 and retained the wrong DET direction. - **Classification:** Run-environment miss began in the narrow score construction and was amplified by an extreme MIN offensive/bullpen outcome. Melton’s +3 ER and Finnegan’s +5 ER were important, but the 12-run MIN underprojection is primarily a run-tail failure. ### MIA@KC - Historical ML was near 5–5; variance stayed neutral. - Enhanced prediction changed to MIA 3, KC 5 and produced the KC direction. - **Classification:** Bad weighting of pitcher/context inputs. Dobnak’s credible home split and Phillips’ expected weakness were not sufficiently represented. This was a starter projection and conviction failure, not merely a random one-run miss. ### MIL@CHC - The historical score was nearly even; variance moved to MIL 4, CHC 5. - Enhanced prediction sharply moved CHC to 8 runs and generated a strong CHC direction. - **Classification:** Winner-direction regression at the enhanced/recent-form stage, compounded by high-raw overconfidence and starter/venue mismatch. The model overweighted recent CHC offensive evidence and underweighted Gasser’s credible away weakness. ### SD@CIN - Historical ML favored SD slightly. - Variance applied moved the direction from SD to tie, an intermediate winner-direction regression. - Enhanced prediction then recovered SD and increased its score to 6–4, and Savant raised it to 7–4. - **Classification:** The final side was still wrong, but the pipeline had an explicit intermediate regression at `historical_ml → variance_applied`; later recovery does not erase it. The final loss was principally a starter projection error: Lodolo allowed four fewer ER and six more outs than projected, with conviction amplified by a 0.60 published pick probability. ### SEA@BOS - Historical ML produced SEA 6, BOS 1, favoring SEA by score. - Variance retained SEA direction. - Enhanced prediction reversed the score to SEA 3, BOS 5 based on recent form, starter-quality, recency, and season-quality controls. - This reversal was not listed as a winner regression in the transition table, but it is the earliest substantive movement toward the wrong final side in the game trace. - **Classification:** Bad weighting of recent/season and starter-quality controls, followed by an extreme Paez starter miss. The home confidence was weak in publication, so this was not a high-confidence calibration failure. ### STL@LAD - Historical ML was nearly neutral. - Variance stayed tied. - Enhanced prediction moved to STL 4, LAD 6; Savant increased LAD to 7 and supported a blowout expansion. - Fundamental edges strongly favored LAD, and the published side remained LAD. - **Classification:** Run-environment underdispersion and over-reliance on LAD quality; secondary wrong conviction. McGreevy’s pitch-count warning is an artifact/data-quality issue, but it does not establish that the pick was caused by the warning. ## Error-type classification across losses - **Bad input:** No critical bad-input defect was established. The STL pitch-count inconsistency is a warning; the missing integrity-revision file is an artifact gap. - **Bad weighting:** Best supported in `MIA@KC` and `MIL@CHC` through credible venue-role splits that were underweighted; also plausible in `SEA@BOS` due to aggressive form/quality controls. - **Calibration:** Strongly implicated in the confidence level of `BAL@COL`, `MIL@CHC`, `SD@CIN`, and `STL@LAD`; not necessarily the cause of the underlying side selection. - **Starter error:** Primary in four losses; substantial in nearly every loss. - **Offense/bullpen miss:** Strong in `DET@MIN`, `STL@LAD`, `SEA@BOS`, and `ATL@WAS`. - **Run-environment miss:** Primary in `DET@MIN` and `STL@LAD`, and broadly visible in the slate margin compression. - **Irreducible variance:** Supported for the large outlier starts in `ATL@WAS`, `DET@MIN`, `SD@CIN`, `SEA@BOS`, and `STL@LAD` where no credible contextual signal explains the deviation. ## Pipeline-stage quality across all stages | Stage | Winner accuracy | Winner hits | Margin MAE | Team-run MAE | Total-run MAE | |---|---:|---:|---:|---:|---:| | Historical ML | 46.67% | 7 | 4.1769 | 3.0977 | 4.6893 | | Variance applied | 40.00% | 6 | 4.1333 | 3.0667 | 4.6667 | | Enhanced prediction | 40.00% | 6 | 4.9333 | 3.2667 | 4.6667 | | Embedded Savant/run environment | 40.00% | 6 | 4.8667 | 3.1667 | 4.4667 | | Fundamental edge target review | 46.67% | 7 | 4.8000 | 3.1333 | 4.5333 | | Capped target | 46.67% | 7 | 4.8000 | 3.1333 | 4.5333 | | Simulation result | 46.67% | 7 | 4.8000 | 3.1333 | 4.5333 | | Published final | 46.67% | 7 | 4.8000 | 3.1333 | 4.5333 | ### Transition audit - `historical_ml → variance_applied` - Margin MAE improved by 0.0436. - Team-run MAE improved by 0.031. - Winner accuracy fell by 6.67 percentage points, from 7 to 6. - Changed directions: `ATL@WAS WAS→TIE`, `MIA@KC KC→TIE`, `SD@CIN SD→TIE`, `STL@LAD LAD→TIE`. - `ATL@WAS` was explicitly regressed. - `variance_applied → enhanced_prediction` - Margin MAE worsened by 0.8. - Team-run MAE worsened by 0.2. - Winner accuracy was flat. - `NYM@TB` improved; `SEA@BOS` regressed. - Direction changed in six games, including `SEA@BOS BOS→SEA` in the stored direction notation. - `enhanced_prediction → embedded_savant_run_environment` - Small margin improvement of 0.0666 and team-run improvement of 0.1. - `TOR@CLE` moved from TOR to tie; no winner accuracy change. - `embedded_savant_run_environment → fundamental_edge_target_review` - `TOR@CLE` moved tie to CLE, improving one winner. - Margin and team-run MAE improved slightly. - All later transitions were inert: - No changed directions. - No metric changes. - No simulation winner intervention. The key conclusion is that the enhanced stage was the main score-quality regression versus historical ML, while fundamental-edge review recovered one winner. The recovery does not erase the earlier regressions. --- # 5. Cross-game counterfactuals ## Lever tension 1: Widen score distributions **Lever:** `score.variance_multiplier / score.variance_floor` - **Helps:** `BAL@COL`, `MIA@KC`, `MIL@CHC`, `SD@CIN`, and `STL@LAD`. - **Hurts:** None listed. - **Net evidence:** 5 helps versus 0 hurts. - **Interpretation:** This is the strongest direct slate-level case for a bounded score-tail shadow experiment. It is consistent with: - 13/15 predicted margins ≤2. - 9/15 actual margins ≥4. - Persistent negative margin bias. - **Caution:** The lever should not be globally widened solely because `DET@MIN` was a 13-run margin. The broader distributional evidence, not that single blowout, justifies testing. ## Lever tension 2: Pitcher form weights and prior games **Lever:** `player.pitcher.v4 form weights / player.pitcher_prior_games` - **Helps:** Ten games by the supplied attribution: - `BAL@COL`, `CHW@HOU`, `DET@MIN`, `MIA@KC`, `MIL@CHC`, `NYM@TB`, `SD@CIN`, `SEA@BOS`, `SF@PIT`, `TOR@CLE`. - **Hurts:** None listed. - **Net evidence:** 10 helps versus 0 hurts. - **Interpretation:** Strong descriptive support for a shadow test, but the “help” labels are based on starter error and should not be treated as proof that more recent form always improves forecasts. Extreme starter outcomes create noisy retrospective attribution. ## Lever tension 3: Recent/season run-mean weighting **Lever:** `recent.* / season.* run-mean weighting` - **Helps:** 13 games: `ATL@WAS`, `BAL@COL`, `CHW@HOU`, `DET@MIN`, `MIA@KC`, `MIL@CHC`, `NYM@TB`, `OAK@TEX`, `PHI@ARI`, `SD@CIN`, `SEA@BOS`, `SF@PIT`, `STL@LAD`, `TOR@CLE` — the JSON’s help list contains 13 named games and omits two listed slate games from the help count. - **Hurts:** `NYY@LAA`, where team runs were already close, with an explicit warning that reweighting could disturb it. - **Net evidence:** 13 helps versus 1 hurt. - **Interpretation:** The net count supports investigation, but this lever is risky because recent and season signals already drove the wrong side in `SEA@BOS` and contributed to high-confidence errors in `MIL@CHC`/`STL@LAD`. ## Tempting one-game fixes that must not be made 1. **Do not widen every score because of `DET@MIN`.** The causal register explicitly prohibits this. 2. **Do not permanently shrink all high probabilities because `BAL@COL`, `MIL@CHC`, `SD@CIN`, and `STL@LAD` lost.** The four-game cluster is a warning, not a stable calibration curve. 3. **Do not increase venue weighting solely from `MIA@KC`.** `MIL@CHC` provides supporting evidence, but the venue split must be validated across the rolling window. 4. **Do not remove recency controls because of `SEA@BOS`.** `NYM@TB`, `NYY@LAA`, and `PHI@ARI` show that recency/quality information can coexist with correct winners. 5. **Do not treat the simulation as a failed winner selector solely because it produced zero flips.** Its persistent inertness requires a code/diagnostic check, not an arbitrary parameter move. 6. **Do not treat the STL pitch-count warning as the cause of the STL loss.** It is a data-contract issue, not evidence of wrong-side causation. --- # 6. Data and contract defects ## Critical findings - **Critical count: 0.** - No critical data or contract finding was supplied. ## Error findings - **Error count: 0.** - No error-severity structural finding was supplied. ## Warning findings 1. **`STL@LAD`: pitch-count mismatch** - Player: Michael McGreevy. - Rule: `PITCH_COUNT_PITCHER_MATCH`. - Summary BF: 24. - Box-score BF: 23. - Source: simulation. - Severity: warning. - Game risk score: 10. - This is an artifact/data-contract warning, not established prediction causation. 2. **Canonical artifact warning** - Slate structural totals: 0 critical, 0 error, 1 warning, 108 info. - The exact aggregate warning above is the only supplied warning-level finding. ## Informational findings The evidence supplies an aggregate count of **108 informational findings** but does not enumerate their individual messages. The major named informational categories are: - Complete 15/15 audit coverage. - Nine material calibrated-to-published movements with named transformations. - Zero simulation flip candidates and zero published flips. - Per-game forecast-file informational notes, usually seven or eight per game. - No material integrity defects in most game artifacts. Because individual info messages are not present in the supplied JSON, they cannot be exhaustively itemized without inventing data. ## Prediction-model defects versus presentation/artifact defects ### Prediction-model/process risks - Compressed predicted margins and weak score tails. - Weak holdout run-model signal. - High raw-probability cluster losing 0–4. - Poor current-slate calibration. - Persistent simulation inertness. - Persistent run-model weakness. - Potentially underweighted pitcher venue-role effects in `MIA@KC` and `MIL@CHC`. - Potentially harmful enhancement-stage score regression. ### Presentation/artifact defects - Missing integrity-revision audit artifact. - `STL@LAD` pitch-count contract warning. - `TOR@CLE` unexplained publication delta. - Possible score-path versus published-score reconciliation issue in `CHW@HOU`: the stored score evolution ends at 3–5 while the published prediction is 6–5. This should be checked against the canonical artifact before calling it a defect; the supplied JSON does not provide enough information to determine whether this is a display-path omission or a real artifact mismatch. ## Coverage reconciliation The slate cannot be called completely clean despite full coverage: - Coverage is complete: 15/15. - Critical/error gates are clear: yes. - Warning-free: no. - Complete revision evidence: no, because one manifest artifact is missing. - Complete probability lineage: no, because TOR has an unexplained material delta. ## Probability-lineage audit The required chain is raw model → calibrated → governor → score coherence/recency → published. ### Games with generally recorded lineage The following games show named transformations and no unexplained material delta: `ATL@WAS`, `BAL@COL`, `CHW@HOU`, `DET@MIN`, `MIA@KC`, `MIL@CHC`, `NYM@TB`, `NYY@LAA`, `OAK@TEX`, `PHI@ARI`, `SD@CIN`, `SEA@BOS`, `SF@PIT`, and `STL@LAD`. The continuous-learning finding states that nine material movements were explained by named score-coherence and/or recency transformations. The high-raw cluster specifically records all four movements as explained. ### Bypass/defect - **`TOR@CLE`:** calibrated/governed probability 0.545; published probability 0.555236; delta +0.010236; transformation text says “no recorded transformation explains the publication delta.” - This qualifies as a material lineage defect under the supplied contract, although it did not alter the selected winner. --- # 7. Cross-slate persistence Four prior slates were loaded, creating a five-slate window from 2026-08-28 through 2026-09-01. ## Finding streaks | Finding | Consecutive-slate count | Present now? | Interpretation | |---|---:|---|---| | `high_raw_probability_cluster` | 1 | Yes | Current slate has the cluster, but the streak itself is not persistent. | | `pipeline_winner_instability` | 0 | No | Not present now; incidence in window is 2 but no current consecutive streak. | | `score_margin_compression` | 2 | Yes | Current and prior consecutive slate; persistent emerging run-model issue. | | `simulation_inert_on_slate` | 5 | Yes | Established failure across all five slates. | | `single_slate_calibration_instability` | 2 | Yes | Current and prior consecutive slate; persistent emerging calibration issue. | | `weak_run_model_signal` | 5 | Yes | Established failure across all five slates. | The `escalated_findings` entries must be treated as established failures, not single-slate noise: 1. **Simulation inertness** - Streak: 5 consecutive slates. - Target lever listed: `simulation.allow_score`. - Metric group: run model. - Metrics hurt: `team_run_mae`, `total_run_mae`. - Direction: confirm simulation is running and not silently no-oping. - Parameter actionability: false; this requires code/diagnostic action rather than forced tuning. 2. **Weak run-model signal** - Streak: 5 consecutive slates. - Target levers listed: `training.xgb.n_estimators`, `training.xgb.max_depth`, `training.xgb.learning_rate`. - Metrics hurt: `team_run_mae`, `total_run_mae`. - Direction: investigate features and target with an isolated training backtest. - Parameter actionability: false in the registered finding; do not force a hyperparameter change from this report. ## Metric trends All reported directions must be stated explicitly: - **Margin bias:** **Regressing.** - Mean: −2.4068. - Series: −1.333, −1.882, −2.786, −3.5, −2.533. - Interpretation: increasingly underpredicted margins, consistent with compression. - **Mean team-run error on losses:** **Regressing.** - Mean: 6.075. - Series: 5.0, 3.0, 7.0, 8.25, 7.125. - **Weighted calibration error:** **Regressing.** - Mean: 0.2194. - Series: 0.2179, 0.3961, 0.0693, 0.1661, 0.2476. - **Winner accuracy:** **Regressing.** - Mean: 0.6083. - Series: 0.4667, 0.9412, 0.5, 0.6667, 0.4667. ## Top-confidence-pick streak The top-confidence pick history is: - 2026-08-28: hit. - 2026-08-29: hit. - 2026-08-30: miss. - 2026-08-31: hit. - 2026-09-01: miss. Current loss streak: **1**. The top pick is not on a multi-slate losing streak, but the current miss is consistent with the separate high-confidence cluster concern. ## Raw structural blocks - **Margin distribution:** persistent negative margin bias and current compression are present. - **Pipeline transitions:** variance application improved run errors slightly but reduced winner accuracy; enhancement materially worsened margin MAE; later recovery came at fundamental review; simulation and publication were inert. - **Calibration monitoring:** current high-confidence buckets are severely overconfident, while low-confidence buckets are underconfident. - **High-raw cluster:** current 0–4; only one-slate consecutive streak but incidence in the five-slate window is 2. - **Simulation control:** five-slate persistence, diagnostic-only status, 0 flips in 15 current games. --- # 8. Adjustments for today Recommendations are ranked by evidence strength and persistence. These are bounded shadow or diagnostic actions, not promotions. ## 1. Diagnose the inert simulation path — highest persistence, code action - **Pattern:** `simulation_inert_on_slate`, escalated for 5 consecutive slates. - **Registered lever:** `simulation.allow_score`. - **Action:** Do **not** change the parameter blindly. Run a code/diagnostic verification that `simulation.allow_score` is actually read, that score-level simulation executes, and that its output can differ from the capped target. - **Expected metric:** If the module is currently a no-op, restoring execution should change `team_run_mae` and `total_run_mae`; direction should be improvement. - **Helps:** Potentially all score forecasts, especially compressed-margin games: `DET@MIN`, `SEA@BOS`, `SF@PIT`, `STL@LAD`. - **Hurts/regression risk:** Any game where simulation adds unjustified tail variance, including relatively close-score successes such as `NYY@LAA` and `TOR@CLE`. - **Gate:** Require observable nonzero simulation contribution on shadow output and rollback if rolling team-run or total-run MAE worsens. ## 2. Run an isolated run-model feature/target backtest — highest persistent model weakness - **Pattern:** `weak_run_model_signal`, escalated for 5 consecutive slates. - **Registered targets:** `training.xgb.n_estimators`, `training.xgb.max_depth`, `training.xgb.learning_rate`. - **Action:** Code/diagnostic action, not an immediate parameter move. Test the existing features and target definition in an isolated temporal backtest before selecting a bounded hyperparameter shadow. - **Expected metric:** Improve team-run MAE and total-run MAE; R² should improve without degrading winner Brier. - **Helps:** All run forecasts, especially `DET@MIN`, `SEA@BOS`, `SF@PIT`, `STL@LAD`, and `PHI@ARI`. - **Hurts/regression risk:** Increased tail flexibility could worsen close-score games such as `NYY@LAA`, `NYM@TB`, and `TOR@CLE`. - **Gate:** Do not promote unless improvement persists across rolling temporal folds and does not materially worsen winner accuracy or Brier. ## 3. Bounded score-tail shadow experiment — persistent compression - **Pattern:** `score_margin_compression`, present for 2 consecutive slates; margin bias is regressing. - **Exact lever:** `score.variance_multiplier` — increase slightly, for example **+0.05** in shadow only. Keep `score.variance_floor` unchanged initially to isolate the effect. - **Expected metric:** Reduce negative margin bias and improve `team_run_mae`/`total_run_mae`; expected direction is improvement. - **Helps:** The lever-tension list identifies `BAL@COL`, `MIA@KC`, `MIL@CHC`, `SD@CIN`, and `STL@LAD`. - **Hurts:** No hurt games were listed in the direct tension record, but regression risk exists for close or already-accurate games, especially `NYY@LAA`, `NYM@TB`, and `TOR@CLE`. - **Gate:** Shadow against the unchanged multiplier; rollback if margin MAE or Brier worsens over the registered rolling evaluation. ## 4. Bounded calibration-confidence shadow — current calibration and high-raw cluster - **Patterns:** `single_slate_calibration_instability` has a 2-slate streak and is present; `high_raw_probability_cluster` is present with a current 0–4 result but only a one-slate consecutive streak. - **Exact lever:** `confidence.base_cap` — reduce modestly, for example **0.01** in shadow. Do not simultaneously change all three confidence levers. - **Expected metric:** Lower weighted calibration error and improve Brier by reducing overconfidence in the 0.60+ region. - **Helps:** Current high-confidence misses `BAL@COL`, `MIL@CHC`, `SD@CIN`, and `STL@LAD`. - **Hurts:** Correct high-conviction or moderate calls could become underconfident, especially `MIA@KC`’s opposing-side calibration context, and low-confidence successes could lose useful separation. - **Gate:** Monitor bucket calibration over the rolling window; rollback if low-confidence underconfidence worsens or Brier deteriorates. ## 5. Pitcher-form/venue shadow — material but less persistent than run-model failures - **Pattern:** Starter projection was the primary driver in four losses; the lever-tension count is 10 helped versus 0 hurt. - **Exact lever:** `player.pitcher.v4 form weights` — increase the recent-form component slightly in shadow, without changing `player.pitcher_prior_games` simultaneously. - **Expected metric:** Improve `starter_er_mae` and `starter_outs_mae`. - **Helps:** The tension record names `BAL@COL`, `CHW@HOU`, `DET@MIN`, `MIA@KC`, `MIL@CHC`, `NYM@TB`, `SD@CIN`, `SEA@BOS`, `SF@PIT`, and `TOR@CLE`. - **Hurts:** No direct hurt games were listed, but recent-form overweighting could damage stable starter baselines and exacerbate `SEA@BOS`-type reversals. - **Important companion diagnostic:** Separately shadow `context.venue_adjustment_weight` for credible venue-role cases, especially `MIA@KC` and `MIL@CHC`; do not bundle this with the form-weight experiment. - **Gate:** Require improvement in starter MAE without worsening winner Brier or introducing excess score volatility. ## 6. Single strongest current-slate driver: run-environment underprojection The strongest single-slate aggregate driver is run-environment compression/underdispersion, because it affected the largest errors and is supported by a two-slate streak and five-slate regressing margin bias. The bounded `score.variance_multiplier` recommendation above is therefore the safest direct single-slate parameter experiment. Do not separately tune a one-game DET/MIN-specific factor. --- # 9. Bottom line ## Known - The slate was 7–8, with poor winner accuracy and especially poor calibration at higher confidence. - Score margins were materially compressed: 13 predicted margins ≤2 versus 9 actual margins ≥4. - The enhancement stage was the worst measured stage for margin and team-run MAE. - Starter projection errors materially contributed to four losses, while several extreme deviations are best classified as variance. - Four high-raw-probability selections lost. - Simulation produced no winner-direction intervention and has done so for five consecutive slates. - Weak run-model signal has persisted for five consecutive slates. - All 15 games were reviewed and structurally audited. - There were no critical or error-level structural findings. - There was one structural warning, one missing manifest artifact, and one unexplained material probability-publication delta. ## Uncertain - Whether widening score variance will improve future slates rather than merely fit this run of high-margin outcomes. - Whether pitcher-form or venue adjustments are causal improvements rather than retrospective explanations. - Whether the high-confidence cluster reflects calibration failure, winner variance, or a short-lived sample anomaly. - Whether the simulation is genuinely inert or whether its score contribution is simply not represented in winner-flip metrics. - Whether the `CHW@HOU` score-path discrepancy is a true artifact defect or an omitted intermediate rendering detail. - The individual content of the 108 informational audit findings is not included. ## Safest next action The single safest next action is: > **Run a non-promoting diagnostic/shadow audit of the simulation path and probability/score lineage, while testing a small increase to `score.variance_multiplier` only in shadow.** This addresses the five-slate simulation/run-model persistence, directly tests the current score-compression failure, preserves rollback ability, and avoids making an irreversible one-game calibration or winner-side correction.