# Luna deep analysis — slate 20260917 > Coverage receipt: **COMPLETE** · 9/8 sections; 9/9 games; 4/4 causes; 5/6 sources # 1. Evidence receipt and gaps ## Sources received All seven `source_manifest` entries were checked against the supplied evidence: | Source | Status | Purpose | |---|---:|---| | `prediction_review_20260917.json` — 229,581 bytes | Loaded | Official outcomes and scoring errors | | `PREDICTION_SYSTEM_DEEP_AUDIT_20260917.json` — 150,866 bytes | Loaded | Probability, provenance, starter, and decision contracts | | `continuous_learning_daily_20260917.json` — 1,555,702 bytes | Loaded | Pipeline stages, calibration, lineage, score distribution, starter comparisons | | `detailed_20260917.html` — 214,325 bytes | Loaded | Full game and simulation rendering | | `20260917_ultra_detailed_player_audit.html` — 20,306,685 bytes | Loaded | Player-level prediction trace | | `audit_report_20260917.html` — 1,143,113 bytes | Loaded | Structural and invariant findings | | `audit_report_20260917_integrity_revision.html` | **Missing** | Post-Luna remediation proof for immutable revision artifacts | The missing integrity-revision audit is an evidence gap. It does not invalidate the settled-game results, because the structural audits, prediction artifacts, and game-level audit receipts are present, but it prevents independent confirmation of post-remediation immutable-revision status. ## Slate coverage All nine expected games were reviewed: - BOS@TEX - DET@CHW - KC@HOU - LAD@CIN - MIL@PIT - MIN@LAA - OAK@TB - PHI@NYM - SD@COL Coverage was complete: - Expected games: 9 - Games structurally audited: 9 - Forensic deep-audit games: 9 - Continuous-learning games: 9 - Starter comparisons: 9 - Canonical artifacts loaded: 9 - Learning eligibility: eligible; all critical/error structural gates passed All nine game-level structural audits reported zero critical findings, zero errors, and zero warnings. They reported 70 informational findings in aggregate. ## Internal contradictions and evidence limitations 1. **BOS@TEX probability-lineage wording is internally awkward.** The row says `published_minus_calibrated = -0.009473`, the transformation says “no recorded transformation explains the publication delta,” but `unexplained_material_delta` is `false`. The delta is small and not material, so this is not a material bypass, but the wording should be normalized. 2. **The missing integrity-revision artifact remains unresolved.** It is the only missing manifest entry. 3. **Published probability and confidence edge use different representations.** For example, LAD@CIN has a governed/calibrated probability of 0.7106, but its published away-side probability is 0.56 and its displayed home-minus-away edge is -0.12. This is explainable by the named score-coherence and recency transformations, but reports should make the selected side and side-relative probability explicit. 4. **The pipeline-stage table contains a meaningful winner-direction regression that is not obvious from aggregate final accuracy.** BOS@TEX moves historical ML `TEX → TIE`, then enhanced prediction moves `TIE → TEX`. The later recovery does not erase the intermediate regression. 5. **No market cross-check was available.** There were zero source-labelled moneyline covers, so market agreement/disagreement cannot be evaluated. 6. **Player-level results are noisy.** Only 13 of 53 material player deviations were supported by the hitter’s own season contact quality, a learnable share of 24.5%. The other deviations should not be treated as feature evidence. --- # 2. Executive diagnosis ## Overall result - Winner accuracy: **4/9 = 44.44%** - Misses: 5 - Reported walk-forward accuracy: **56.47%** - Reported walk-forward Brier: **0.243312** - Slate winner Brier from game records: average approximately **0.2568** - Mean team-run error: approximately **3.67 runs** - Mean total-run bias across the slate: approximately **+1.11 runs** - High-conviction games: 4 - High-conviction misses: 2 - High-conviction miss rate: **50%** - Published confidence was not anti-predictive by the formal flag, but it was materially overconfident in the settled calibration buckets. This was not a good prediction slate merely because four winners were correct. Winner accuracy was below the stated walk-forward baseline and below coin-flip performance on this slate. Exact-score performance was substantially weaker than winner performance. ## Separate diagnoses ### Winner accuracy Poor on this slate: - 4/9 correct - 44.44% versus a 56.47% walk-forward baseline - Self-assessment: `beat_coinflip: false`, `beat_validated_baseline: false` - Trend: `winner_accuracy` is regressing; current slate is the lowest value in the displayed recent series at 44.44% The five misses were BOS@TEX, DET@CHW, MIL@PIT, MIN@LAA, and PHI@NYM. ### Calibration quality Calibration was weak and overconfident: - 0.50–0.55 bucket: average confidence 0.532, observed hit rate 0.400, calibration error **+0.132** - 0.55–0.60 bucket: average confidence 0.5635, observed hit rate 0.500, calibration error **+0.0635** - Settled games in calibration monitor: 9 - Weighted absolute calibration error: **0.1016** - High-confidence cluster: 2/2 hits, but the cluster’s empirical hit rates were only 0.5536 and 0.5846 despite raw probabilities above 0.65 The model did reduce high raw probabilities through score coherence and matchup-recency guardrails. That is good lineage behavior, but the residual published confidence still overstated reliability in the moderate and lean buckets. ### Score accuracy Score accuracy was materially weak: - Holdout R²: - Away runs: 0.02324 - Home runs: 0.002745 - Total runs: 0.005629 - The causal register correctly labels this a **high-severity** weak-run-model issue. - The model predicted margins of 2 runs or fewer in 6/9 games, while actual margins of at least 4 occurred in 4/9. - Actual margins: 1, 2, 4, 6, 3, 1, 9, 3, 7 - Predicted margins: 1, 1, 1, 4, 2, 1, 3, 1, 3 - Signed margin bias: **-0.778** - Cross-slate margin bias trend: **regressing** The system systematically compressed the margin distribution, particularly in OAK@TB and SD@COL. ### Starter accuracy Starter performance was mixed and often decisive: - Mean starter ERA error: **2.16** - Five of 17 listed starts exceeded projected ERA by at least 3.00: - Kyle Harrison: +16.2 ERA error - Brady Singer: +11.19 - Tanner Gordon: +10.2 - Jeffrey Springs: +8.49 - Seth Lugo: +5.4 - One thin-history starter was identified: Miguel Ullola, six prior appearances in the feed. - The starter lever had balanced cross-game evidence: 4 games helped, 4 games hurt. No broad starter-weight adjustment is justified from this slate alone. Venue-split evidence was specifically credible for several blown starters, especially Seth Lugo, Kyle Harrison, and Tanner Gordon. This supports a bounded shadow test of `context.venue_adjustment_weight`, but not an unvalidated production change. ### Artifact integrity Artifact integrity was strong for the settled game artifacts: - All nine game-level structural audits passed. - All reported zero critical, error, and warning findings. - Probability reversions: - Artifacts checked: 9 - Blend flips: 0 - Side survived to publication: 9 - Overwritten: 0 - Six material calibrated-to-published movements were explained by score-coherence and/or recency transformations. - Simulation produced zero winner flips on this slate. However, the missing integrity-revision audit prevents a fully closed artifact-revision receipt. ## Causal issue register reconciliation ### `weak_run_model_signal` — High - Severity: **high** - Evidence: home/away/total R² = 0.0027/0.0232/0.0056 - Affected games: all nine - Action: treat exact scores as low-confidence; improve features or target before tuning score tails - Guardrail: a small positive R² only means the model beats guessing the average; it does not make exact scores reliable This is a material model-validation risk, not a contract failure. The appropriate action is an isolated feature/target backtest, not a nightly score tweak. ### `complete_slate_audit_coverage` — Info - Severity: **info** - Evidence: 9/9 games structurally audited - Affected games: all nine - Action: retain per-game audit receipts - Guardrail: full audit coverage does not cancel within-game findings Correctly classified as informational. ### `probability_publication_delta_explained` — Info - Severity: **info** - Evidence: six material movements, all named as score-coherence and/or recency transformations - Affected games: DET@CHW, KC@HOU, LAD@CIN, MIL@PIT, OAK@TB, SD@COL - Action: continue displaying every probability layer and transformation - Guardrail: a published probability above the calibrated probability is not a governor bypass when the transformation is recorded This is supported. BOS@TEX has a small wording inconsistency but no material unexplained delta. ### `simulation_inert_on_slate` — Info for this slate, escalated persistent failure across slates - Severity in the register: **info** - Evidence on this slate: zero flip candidates and zero published flips - Affected games: all nine - Action on the slate: measure simulation contribution across rolling slates - Guardrail: zero winner flips is not itself a defect and does not prove that no score reconciliation occurred Cross-slate evidence changes the operational diagnosis: this has persisted for 11 consecutive slates and is now an established code/diagnostic failure pattern. --- # 3. Every game The game records use `confidence_home_minus_away`; negative values indicate an away-side edge. The published pick probability below is the probability of the selected side. ## BOS@TEX — miss; correct-winner/wrong-score: no - Predicted: TEX 4, BOS 3; predicted winner TEX - Actual: BOS 4, TEX 3; winner BOS - Published selected-side probability: TEX 0.553827 - System confidence edge: home-minus-away **+0.1077**, moderate - Brier: **0.306724** - Team-run absolute error: **2** - Total-run bias: **0** - Starter lines: - BOS Sonny Gray: projected 2 ER/18 outs, actual 3 ER/19 outs; ERA 3.00 projected vs 4.26 actual; ER error +1 - TEX Tyler Alexander: projected 1 ER/15 outs, actual 0 ER/3 outs; ERA 1.80 vs 0.00; ER error -1 - Edges: - Season home-minus-away: **-0.484** - Recent: **+0.262** - Pitcher: **-0.061** - Composite: **-0.154** - Controls: - Governor applied - Confidence pipeline reached 0.5633 before publication - No recency guardrail - No score-volatility change - Starter-experience shrinkage not applied - Score path: 3.915–4.294 historical ML → 4–4 variance baseline → 3–4 enhanced → 3–4 through Savant, target review, cap, and simulation. - Player adjustments: - Sonny Gray predicted strikeouts increased by 2 - Danny Jansen RBI projection worsened by 1.67 - Jake Burger RBI projection improved by 1.45 - Corey Seager AB projection worsened by 1.27 - Trevor Story strikeout projection worsened by 1.34 - Material player changes: 22 - Audit: PASS; zero critical/error/warning findings; eight informational notes. Saved artifact was not the cause. - Causes: - Primary: **wrong conviction side** - Secondary: none - Actual game-level explanation: BOS lineup scored below projection by 2, TEX lineup below projection by 1; result was close, but the directional side was wrong. - Important caution: this was a moderate 0.11 edge, not a near-coin-flip classification and not a high-conviction pick. ## DET@CHW — miss; correct-winner/wrong-score: no - Predicted: DET 5, CHW 4; predicted winner DET - Actual: DET 1, CHW 3; winner CHW - Published selected-side probability: DET approximately **0.530** - System confidence edge: home-minus-away **-0.060**, lean - Brier: **0.280900** - Team-run absolute error: **5** - Total-run bias: **+5** - Starter lines: - DET Framber Valdez: projected 3 ER/18 outs, actual 3 ER/15 outs; projected ERA 4.50 vs actual 5.40; ER error 0 - CHW Erick Fedde: projected 5 ER/14 outs, actual 1 ER/14 outs; projected ERA 9.64 vs actual 1.93; ER error -4 - Edges: - Season: **-0.442** - Recent: **-0.770** - Pitcher: **+0.084** - Composite: **-0.431** - Controls: - Governor and recency guardrail applied - Confidence pipeline reached 0.5837 before later controls - Starter-experience shrinkage not applied - `supported_blowout_margin_expansion` applied to the away side - Score path: 4.539–4.383 historical → 5–4 variance baseline → 5–4 enhanced → 5–4 through final simulation. - Player adjustments: - Javier Báez RBI projection improved by 3.66 - Randal Grichuk AB projection worsened by 2.50 - Chase Meidroth strikeout projection worsened by 2.07 - Miguel Vargas hit projection worsened by 1.89 - Material player changes: 25 - Audit: PASS; zero critical/error/warning findings; seven informational notes. - Causes: - Primary: **starter projection** - Secondary: **run environment over** - Fedde allowed four fewer ER than projected. DET’s offense was over-projected by four. This was a large score miss and a wrong side, not merely a one-run variance outcome. ## KC@HOU — hit; correct-winner/wrong-score: yes - Predicted: HOU 5, KC 4; predicted winner HOU - Actual: HOU 6, KC 2; winner HOU - Published selected-side probability: HOU **0.529531** - System confidence edge: home-minus-away **+0.0591**, lean - Brier: **0.221341** - Team-run absolute error: **3** - Total-run bias: **+1** - Starter lines: - KC Seth Lugo: projected 1 ER/15 outs, actual 4 ER/15 outs; ERA 1.80 vs 7.20; ER error +3 - HOU Miguel Ullola: projected 4 ER/9 outs, actual 1 ER/2 outs; ERA 12.00 vs 13.50; ER error -3 - Edges: - Season: **+0.118** - Recent: **-0.234** - Pitcher: **+0.011** - Composite: **+0.211** - Controls: - Governor and recency guardrail applied - Thin-history starter shrinkage applied to Ullola: six prior appearances, shrink factor 0.675; probability changed from 0.5606 to 0.540874 without changing the winner - No score-volatility change - Large Savant run-environment swing flag - Score path: 4.539–4.383 historical → 5–4 variance → 5–4 enhanced → 5–4 final. - Player adjustments: - LaMonte Wade Jr. exceeded projection by 1.85 runs - Taylor Trammell by 1.26 - Jeremy Peña by 1.09 - Daulton Varsho by 0.93 - Jac Caglianone underperformed by 0.78 - Audit: PASS; zero critical/error/warning findings; seven informational notes. - Causes: no registered primary/secondary driver. The winner was correct, but the score was wrong. - Forensic interpretation: Seth Lugo’s credible away split was underweighted. The recommended shadow lever is `context.venue_adjustment_weight`, not an immediate production change. ## LAD@CIN — hit; correct-winner/wrong-score: yes - Predicted: LAD 9, CIN 5; predicted winner LAD - Actual: LAD 8, CIN 2; winner LAD - Published selected-side probability: LAD **0.56** - System confidence edge: home-minus-away **-0.12**, moderate - Brier: **0.193600** - Team-run absolute error: **4** - Total-run bias: **+4** - Starter lines: - LAD Justin Wrobleski: projected 3 ER/15 outs, actual 0 ER/9 outs; ERA 5.40 vs 0.00; ER error -3 - CIN Brady Singer: projected 4 ER/14 outs, actual 7 ER/10 outs; ERA 7.71 vs 18.90; ER error +3 - Edges: - Season: **-1.929** - Recent: **-1.414** - Pitcher: **-0.059** - Composite: **-2.264** - Controls: - Governor, score coherence, and matchup-recency guardrail applied - Confidence pipeline reached 0.7106 before publication - `supported_blowout_margin_expansion` - High raw cluster member; published probability was reduced to 0.56 - Score path: approximately 5–4 historical → 6–4 variance → 8–5 enhanced → 9–5 final. - Player adjustments: - Kyle Tucker exceeded projection by 2.63 runs, supported by above-average xwOBA 0.330 - Miguel Rojas by 1.87 - Sal Stewart by 1.40, supported by elite xwOBA 0.346 - Elly De La Cruz underperformed by 1.73 - Audit: PASS; zero critical/error/warning findings; eight informational notes. - Causes: none registered; this was a solid winner call with score overstatement. - Important classification: **correct winner, wrong score**. Brady Singer’s strong-offense matchup and large ERA miss were directionally useful, but the exact 9–5 forecast was too high. ## MIL@PIT — miss; correct-winner/wrong-score: no - Predicted: MIL 6, PIT 4; predicted winner MIL - Actual: MIL 4, PIT 7; winner PIT - Published selected-side probability: MIL **0.56** - System confidence edge: home-minus-away **-0.12**, moderate - Brier: **0.313600** - Team-run absolute error: **5** - Total-run bias: **-1** - Starter lines: - MIL Kyle Harrison: projected 3 ER/15 outs, actual 4 ER/5 outs; ERA 5.40 vs 21.60; ER error +1, ERA error +16.2 - PIT Wilber Dotel: projected 1 ER/15 outs, actual 1 ER/6 outs; ERA 1.80 vs 4.50; ER error 0 - Edges: - Season: **-1.170** - Recent: **-1.138** - Pitcher: **-0.061** - Composite: **-1.423** - Controls: - Governor and recency guardrail applied - `supported_blowout_margin_expansion` toward MIL - Confidence pipeline reached 0.6562 before publication - Score path: 5–4 historical → 5–4 variance → 6–4 enhanced → 6–4 final. - Player adjustments: - Brandon Lowe improved by 2.69 projected hits - Joey Ortiz AB improved by 1.85 - Sal Frelick RBI improved by 1.82 - Jackson Chourio underperformed by 1.36 - Bo Naylor exceeded RBI projection by 1.75 - Audit: PASS; zero critical/error/warning findings; eight informational notes. - Causes: - Primary: **wrong conviction side** - Secondary: **lean-side miss** in the registered labels, despite the displayed moderate 0.12 edge - Kyle Harrison’s credible away split was underweighted; PIT offense was under-projected by 3. - A 0.12 edge is moderate, not a near coin flip. The problem was not that the edge was “almost zero”; it was that a moderate directional signal was treated too confidently relative to the calibration evidence. ## MIN@LAA — miss; correct-winner/wrong-score: no - Predicted: MIN 4, LAA 3; predicted winner MIN - Actual: MIN 4, LAA 5; winner LAA - Published selected-side probability: MIN **0.516481** - System confidence edge: home-minus-away **-0.033**, toss-up - Brier: **0.266753** - Team-run absolute error: **2** - Total-run bias: **-2** - Starter lines: - MIN Taj Bradley: projected 3 ER/18 outs, actual 3 ER/21 outs; ERA 4.50 vs 3.86; ER error 0 - LAA Walbert Ureña: projected 1 ER/17 outs, actual 0 ER/21 outs; ERA 1.59 vs 0.00; ER error -1 - Edges: - Season: **+0.079** - Recent: **+0.407** - Pitcher: **-0.034** - Composite: **-0.110** - Controls: - Governor and recency guardrail applied - Confidence pipeline reached 0.5228 - No score-volatility change - No experience shrinkage - Score path: 4.523–3.893 historical → 5–4 variance → 4–3 enhanced → 4–3 final. - Player adjustments: - Walker Jenkins exceeded projection by 2.01 runs - Oswald Peraza by 0.94 - Zach Neto underperformed by 1.40 - Mike Trout strikeout projection worsened by 1.17 - Audit: PASS; zero critical/error/warning findings; eight informational notes. - Causes: - Primary: **toss-up variance** - Secondary: none - This is not evidence for a model change. Both starters were close to the projected ER lines, and the actual one-run result was consistent with ordinary variance. ## OAK@TB — hit; correct-winner/wrong-score: yes - Predicted: TB 7, OAK 4; predicted winner TB - Actual: TB 10, OAK 1; winner TB - Published selected-side probability: TB **0.58** - System confidence edge: home-minus-away **+0.16**, moderate - Brier: **0.176400** - Team-run absolute error: **6** - Total-run bias: **0** - Starter lines: - OAK Jeffrey Springs: projected 4 ER/14 outs, actual 6 ER/10 outs; ERA 7.71 vs 16.20; ER error +2 - TB Drew Rasmussen: projected 2 ER/18 outs, actual 0 ER/15 outs; ERA 3.00 vs 0.00; ER error -2 - Edges: - Season: **+1.683** - Recent: **+1.202** - Pitcher: **+0.052** - Composite: **+1.880** - Controls: - Governor and recency guardrail applied - Confidence pipeline reached 0.6502 - `supported_blowout_margin_expansion` - Large Savant run-environment swing flag - Score path: roughly 5–4 historical → 5–4 variance → 6–4 enhanced → 7–4 final. - Player adjustments: - Lawrence Butler exceeded projection by 1.07 runs, supported by above-average xwOBA 0.321 - Richie Palacios by 0.97 - Junior Caminero underperformed by 1.07 - Zack Gelof underperformed by 0.93 - Audit: PASS; zero critical/error/warning findings; eight informational notes. - Causes: none registered; correct winner, wrong score. - Important classification: **correct winner, wrong score**. The model identified the side but materially understated the margin. ## PHI@NYM — miss; correct-winner/wrong-score: no - Predicted: NYM 4, PHI 3; predicted winner NYM - Actual: PHI 3, NYM 0; winner PHI - Published selected-side probability: NYM **0.544183** - System confidence edge: home-minus-away **+0.0884**, lean - Brier: **0.296135** - Team-run absolute error: **4** - Total-run bias: **+4** - Starter lines: - PHI Aaron Nola: projected 4 ER/18 outs, actual 0 ER/21 outs; ERA 6.00 vs 0.00; ER error -4 - NYM Nolan McLean: projected 2 ER/18 outs, actual 1 ER/18 outs; ERA 3.00 vs 1.50; ER error -1 - Edges: - Season: **-0.312** - Recent: **+0.358** - Pitcher: **+0.068** - Composite: **+0.247** - Controls: - Governor and recency guardrail applied - Confidence pipeline reached 0.5612 - No score-volatility change - Large Savant run-environment swing flag - Score path: 4–5 historical → 4–5 variance → 3–4 enhanced → 3–4 final. - Player adjustments: - Bryce Harper exceeded projection by 0.94 runs, supported by elite xwOBA 0.385 - Kyle Schwarber underperformed by 1.40 - Trea Turner AB projection worsened by 2.94 - Francisco Alvarez RBI projection improved by 1.59 - Audit: PASS; zero critical/error/warning findings; eight informational notes. - Causes: - Primary: **starter projection** - Secondary: **lean-side miss** - Nola’s four-ER projection error was decisive. The forensic attribution calls the start an outlier with no credible venue, opponent, or experience explanation. No model change is warranted for Nola’s individual result. ## SD@COL — hit; correct-winner/wrong-score: yes - Predicted: SD 8, COL 5; predicted winner SD - Actual: SD 9, COL 2; winner SD - Published selected-side probability: SD **0.54** - System confidence edge: home-minus-away **-0.08**, lean - Brier: **0.211600** - Team-run absolute error: **4** - Total-run bias: **+2** - Starter lines: - SD Michael King: projected 3 ER/19 outs, actual 2 ER/18 outs; ERA 4.26 vs 3.00; ER error -1 - COL Tanner Gordon: projected 6 ER/15 outs, actual 7 ER/9 outs; ERA 10.80 vs 21.00; ER error +1 - Edges: - Season: **-1.504** - Recent: **-1.654** - Pitcher: **-0.064** - Composite: **-1.922** - Controls: - Governor and recency guardrail applied - Confidence pipeline reached 0.6972 - `supported_blowout_margin_expansion` - High raw cluster member; published probability reduced to 0.54 - Score path: roughly 5–4 historical → 5–4 variance → 7–4 enhanced → 8–5 final. - Player adjustments: - Xander Bogaerts exceeded projection by 2.47 - Dustin Harris by 0.94, supported by above-average xwOBA 0.336 - Manny Machado underperformed by 1.73 - Tanner Gordon’s credible home weakness was underweighted - Audit: PASS; zero critical/error/warning findings; eight informational notes. - Causes: none registered; solid correct winner call with wrong score. - Important classification: **correct winner, wrong score**. ## Game-level summary Correct winners: - KC@HOU - LAD@CIN - OAK@TB - SD@COL Correct winner but wrong score: - KC@HOU - LAD@CIN - OAK@TB - SD@COL Misses: - BOS@TEX — wrong side - DET@CHW — starter and run-environment miss - MIL@PIT — wrong side, underweighted venue split - MIN@LAA — toss-up variance - PHI@NYM — starter projection miss and lean-side miss There were no exact-score hits reported. --- # 4. Miss reconstruction ## Earliest wrong-direction stages ### BOS@TEX - Earliest adverse movement: historical ML projected TEX ahead, but variance application moved the direction to a tie; enhanced prediction restored TEX. - The decisive problem was not a late artifact or simulation flip. The fundamental evidence was contradictory: - Season edge favored BOS by 0.484 - Recent form favored TEX by 0.262 - Pitcher edge favored BOS by 0.061 - Composite favored BOS by 0.154 - Diagnosis: **bad weighting / wrong conviction side**, with irreducible close-game variance also present. - The 0.11 edge was moderate, not near-coin-flip confidence. ### DET@CHW - Historical ML already favored DET, consistent with the eventual wrong direction. - Enhanced, Savant, target review, and simulation preserved DET. - The earliest wrong evidence was the run/starter input structure that projected DET for 5 and CHW for 4, especially the CHW starter line of 5 ER/9.64 ERA. - Actual Fedde allowed 1 ER, four fewer than projected. - Diagnosis: **bad starter input/projection**, compounded by **run-environment overprojection** and overconfidence in the away side. ### MIL@PIT - Historical ML favored MIL; variance preserved that direction. - Enhanced prediction increased MIL’s expected runs from 5 to 6 and maintained the wrong side. - Savant adjustments slightly reduced both teams but did not correct direction. - Kyle Harrison’s away split was credible and materially adverse, yet the venue-role weakness was underweighted. - Diagnosis: **bad weighting**, specifically venue-split weighting, plus a large starter realization error and PIT offense underprojection. ### MIN@LAA - Historical ML favored MIN; variance, enhanced prediction, and final simulation retained MIN. - The evidence was nearly balanced: - Composite home edge only -0.110 - Published away probability about 0.516 - Displayed edge only 0.033 - Both starters were broadly accurate by ER, and the actual one-run result flipped the predicted side. - Diagnosis: **irreducible variance**, not a validated model defect. ### PHI@NYM - Historical ML favored NYM and the enhanced score path retained NYM. - The strongest adverse realization was Aaron Nola: projected 4 ER, actual 0 ER. - No credible contextual signal explains the outlier. - Diagnosis: **starter projection miss**, specifically irreducible/outlier variance, plus a lean-side miss. ## Loss-cause classification Across five losses: - Starter projection: 2 registered primary drivers — DET@CHW and PHI@NYM - Toss-up variance: 1 — MIN@LAA - Wrong conviction side: 2 — BOS@TEX and MIL@PIT - Bad input: most clearly CHW starter projection line; no structural identity error was found - Bad weighting: venue-role weighting for MIL@PIT, and possibly BOS@TEX conflicting season/recent weighting - Calibration: confidence was too high relative to observed performance, but calibration did not independently create each wrong side - Offense/bullpen miss: - PIT offense under-projected by 3 in MIL@PIT - DET offense over-projected by 4 in DET@CHW - NYM offense underperformed by 4 in PHI@NYM - Run-environment miss: strongest in DET@CHW and PHI@NYM; total runs were over-projected by 5 and 4 respectively - Irreducible variance: strongest in MIN@LAA and PHI@NYM’s Nola outlier A 0.12 confidence edge is moderate. It should not be described as a near coin flip, but neither should it be treated as high-confidence certainty. The appropriate response is calibration and distribution control, not removal of all directional signal. ## Pipeline-stage quality across all stages | Stage | Winner accuracy | Team-run MAE | Total-run MAE | Margin MAE | |---|---:|---:|---:|---:| | historical_ml | 4/9 = 44.44% | 2.0863 | 2.1738 | 3.5824 | | variance_applied | 4/9 | 2.0556 | 2.1111 | 3.6667 | | enhanced_prediction | 4/9 | 1.8889 | 2.0000 | 3.5556 | | embedded_savant_run_environment | 4/9 | 1.9444 | 2.1111 | 3.4444 | | fundamental_edge_target_review | 4/9 | 1.9444 | 2.1111 | 3.4444 | | capped_target | 4/9 | 1.9444 | 2.1111 | 3.4444 | | simulation_result | 4/9 | 1.9444 | 2.1111 | 3.4444 | | published_final | 4/9 | 1.9444 | 2.1111 | 3.4444 | Transitions: 1. Historical ML → variance applied: - Margin MAE regressed +0.0843 - Team-run MAE improved -0.0307 - Winner accuracy unchanged - BOS moved TEX → TIE: an intermediate winner-direction regression 2. Variance applied → enhanced prediction: - Margin MAE improved -0.1111 - Team-run MAE improved -0.1667 - BOS moved TIE → TEX: later recovery - This recovery does not erase the intermediate regression 3. Enhanced → embedded Savant: - Margin MAE improved -0.1112 - Team-run MAE worsened +0.0555 - Winner accuracy unchanged 4. Embedded Savant → fundamental review: - No change 5. Fundamental review → capped: - No change 6. Capped → simulation: - No change 7. Simulation → published final: - No change The system’s later stages improved run-error metrics relative to historical ML, but none improved winner accuracy, and simulation made no measurable winner-direction intervention. --- # 5. Cross-game counterfactuals ## Lever tension: widen run distributions **Lever:** `score.variance_multiplier / score.variance_floor` **Targets:** Brier score, calibration error ### Helps - BOS@TEX: lost a 0.11-confidence pick; wider tails would reduce overconfidence in a close matchup. - MIL@PIT: lost a 0.12-confidence pick; wider tails would better represent the possibility of a moderate edge failing. ### Hurts - No harmed games were recorded in the lever-tension entry. ### Net evidence - 2 games helped, 0 hurt on this slate. - Cross-slate evidence shows lineup runs were below projection on 8 of 12 slates, but the registered bias is not statistically separable from noise. - A bounded shadow experiment is reasonable, but widening variance cannot repair the near-zero run-model signal or guarantee better exact scores. ## Lever tension: pitcher form weights / prior games **Lever:** `player.pitcher.v4 form weights / player.pitcher_prior_games` **Targets:** starter ER MAE, starter outs MAE ### Helps - DET@CHW: Fedde was four ER better than projected. - KC@HOU: Lugo was three ER worse and Ullola three ER better than projected. - LAD@CIN: Singer was three ER worse than projected. - PHI@NYM: Nola was four ER better than projected. ### Hurts - BOS@TEX: starter projection was relatively accurate; ER error 1. - MIL@PIT: starter ER error 1 for Harrison under the registered absolute-ER measure. - MIN@LAA: starter ER errors 0 and -1. - SD@COL: starter ER errors -1 and +1. ### Net evidence - 4 helped, 4 hurt. - No broad form-weight change is supported. - The venue-specific problem is narrower than the general pitcher-form lever. Use `context.venue_adjustment_weight` in a shadow test for credible split cases instead of changing global pitcher priors. ## Lever tension: recent/season run-mean weighting **Lever:** `recent.* / season.* run-mean weighting` **Targets:** team-run MAE, total-run MAE, accuracy ### Helps - DET@CHW: reduced reliance on an incorrect high run environment would help. - LAD@CIN: score was directionally correct; better weighting may reduce the 4-run total overprojection. - MIL@PIT: PIT offense was under-projected. - OAK@TB: winner was right, but margin was understated. - PHI@NYM: lower run environment would help. - SD@COL: margin was understated. ### Hurts - BOS@TEX: team-run error was only 2; more reweighting could disturb a close score. - MIN@LAA: team-run error was only 2; no evidence supports a broad adjustment from this game. ### Net evidence - 6 helped, 2 hurt. - This is the strongest slate-level bounded model lever among the listed tensions, but cross-slate evidence does not support an unvalidated production change because run-model R² is near zero. A shadow change is preferable. ## Tempting fixes that should not be made - Do not globally increase pitcher form weighting from these nine games; the evidence is 4–4. - Do not tune to individual outlier starts such as Aaron Nola, Fedde, or Singer. - Do not treat all 53 player deviations as learnable; only 13 had supporting season contact-quality evidence. - Do not increase confidence simply because LAD, OAK, or SD won; those were correct winners but wrong scores. - Do not remove moderate edges entirely because BOS and MIL lost. A 0.12 edge is moderate, not a coin-flip label. - Do not force a parameter change for the persistent simulation failure; the registered action is a code/diagnostic fix. - Do not tune score tails as a substitute for fixing the near-zero run-model features and target. --- # 6. Data and contract defects ## Critical findings No critical audit findings were reported in the nine game audits or the structural audit summary. The missing integrity-revision artifact is a manifest gap, but it was not reported as a critical settled-game contract failure. ## Error findings No error audit findings were reported. ## Warning findings No warning audit findings were reported. ## Informational findings The aggregate structural audit reported **70 informational findings**. The principal informational findings were: - All nine games structurally audited. - All nine game artifacts passed. - Six material calibrated-to-published probability movements had named transformations. - Simulation produced zero winner flips. - Numerous game-level informational notes documented confidence controls, player changes, score-volatility decisions, and recency handling. ## Model defects versus artifact bugs ### Prediction-model defects 1. **Weak run-model signal — high severity** - Total-run R² 0.005629 - Home-run R² 0.002745 - Away-run R² 0.02324 - Requires feature/target investigation and isolated retraining backtest. 2. **Simulation inertness — persistent code/diagnostic defect** - 11 consecutive slates - Zero flip candidates and zero published flips on this slate - The self-assessment identifies the likely cause: simulation uses the same target-realization draws, early exits after finding a play-by-play realization, and then treats roughly two selected draws as a probability distribution. - This is a code/diagnostic problem, not a parameter-tuning problem. 3. **Margin compression** - Actual margins ≥4 in 4/9 games, predicted margins ≤2 in 6/9. - Cross-slate margin-bias trend is regressing. 4. **Calibration overconfidence** - Weighted absolute calibration error 0.1016 - 50% high-conviction miss rate - Moderate and lean buckets both overconfident. 5. **Player projection separation** - Hit and strikeout spread ratios only 0.402 and 0.423. - RBI per-player model lost to a flat slate mean: MAE 0.642 versus 0.597. - HR model also lost to the flat mean: 0.189 versus 0.171. ### Presentation/artifact defects 1. BOS@TEX lineage wording says no transformation explains a delta while marking it non-material and not unexplained. This should be clarified. 2. The missing integrity-revision audit is a presentation/receipt gap unless independently shown to reflect an actual missing immutable-revision check. 3. Probability layers should consistently identify: - raw model probability - calibrated probability - governor input - governor output - score-coherence probability - recency adjustment - published selected-side probability ## Coverage reconciliation The coverage matrix is clean for the nine settled games: - `all_games_structurally_audited: true` - `canonical_artifacts_loaded: 9` - `continuous_learning_games: 9` - `expected_games: 9` - `forensic_deep_audit_games: 9` - `games_with_starter_comparisons: 9` - `structural_audits_completed: 9` The slate should not be called fully artifact-clean because the integrity-revision audit is missing, but it is structurally complete and learning-eligible. ## Probability-lineage audit Materially relevant rows: - DET@CHW: calibrated 0.5837 → published 0.53; named score-coherence and recency guardrail - KC@HOU: 0.5606 → 0.529531; named transformations - LAD@CIN: 0.7106 → 0.56; score coherence plus matchup-recency guardrail - MIL@PIT: 0.6562 → 0.56; named transformations - OAK@TB: 0.6502 → 0.58; named transformations - SD@COL: 0.6972 → 0.54; named transformations These six material movements have recorded explanations; no material governor bypass is established. BOS@TEX moved from calibrated 0.5633 to published 0.553827, a delta of only -0.009473. The row’s wording is inconsistent, but this is not a material unexplained delta. MIN@LAA and PHI@NYM had small deltas and named recency/score-coherence handling. No probability reversion or overwritten side was identified. --- # 7. Cross-slate persistence A single slate is not enough to establish most model patterns, but several findings have already escalated and must be treated as established failures. ## Escalated findings ### `simulation_inert_on_slate` - Consecutive-slate count: **11** - Incidence in window: 11/11 - Present now: yes - Target lever listed: `simulation.allow_score` - Parameter actionable: no - Metric group: run model - Metrics harmed: team-run MAE and total-run MAE through lack of independent simulation signal - Direction of fix: confirm the simulation module is running and not silently no-oping - Required action: code/diagnostic repair, not a forced parameter move The self-assessment gives the concrete suspected defect: separate target-realization search from independent win-probability simulation; prevent early exit; use a registered batch size such as 200–500 draws or an independent closed-form distribution. ### `weak_run_model_signal` - Consecutive-slate count: **11** - Incidence in window: 11/11 - Present now: yes - Target levers listed: `training.xgb.n_estimators`, `training.xgb.max_depth`, `training.xgb.learning_rate` - Parameter actionable: no, despite listed training levers - Metrics harmed: team-run MAE and total-run MAE - Direction of fix: investigate features and target with an isolated training backtest - Required action: add or validate market no-vig probability where source-labelled, park/weather environment, bullpen rest/fatigue, and platoon-adjusted lineup quality; compare same-fingerprint holdout R²; bump model schema if semantics change This is an established failure, not single-slate noise. ## Non-escalated finding streaks | Finding | Consecutive count | Present now | Interpretation | |---|---:|---:|---| | high_raw_probability_cluster | 0 | No | Occurred once in the 11-slate window; this slate had two high-raw games, but the registered streak is not current | | pipeline_winner_instability | 0 | No | Incidence 2 in window, but no current streak | | score_margin_compression | 0 | No | Incidence 1 in window under this finding label; current margin compression is visible but not escalated under this streak record | | simulation_inert_on_slate | 11 | Yes | Established persistent failure | | single_slate_calibration_instability | 0 | No | Seven occurrences in window but no consecutive current streak | | weak_run_model_signal | 11 | Yes | Established persistent failure | ## Metric trends Each direction is explicit: - `margin_bias`: **regressing**, mean -1.5357 - `mean_team_run_error_on_losses`: **improving**, mean 5.5165, with current slate at 3.6 - `weighted_calibration_error`: **improving**, mean 0.1781, current 0.1016 - `winner_accuracy`: **regressing**, mean 0.5961, current 0.4444 “Improving” calibration and loss-run error do not mean they are good; they mean the recent series is moving in a favorable direction. Current calibration remains materially imperfect. ## Top-confidence-pick loss streak - Reported loss streak: **0** - Per-slate sequence from 20260907 through 20260917: - miss, hit, hit, hit, miss, miss, hit, miss, hit, hit, hit - Current top-confidence pick: hit - Therefore, no persistent top-confidence loss streak is present. ## Other structural persistence - Margin compression appears repeatedly in the raw margin distribution and is reflected in the regressing margin-bias trend, but the formal `finding_streaks` entry reports only zero consecutive slates. It should be monitored rather than falsely labeled an escalated streak. - The run model’s weak holdout R² and simulation inertness are the strongest persistent structural signals. - The player-level spread problem is a current-slate diagnostic, not an escalated cross-slate finding in the supplied streak structure. --- # 8. Adjustments for today Recommendations are ranked by persistence, scope, expected metric impact, affected games, and regression risk. ## 1. Repair and independently measure simulation — highest evidence strength - Pattern: `simulation_inert_on_slate`, persistent for 11 slates - Exact registered target: `simulation.allow_score` - Parameter direction: **do not force a parameter value; repair the code path** - Bounded action: - Confirm `simulation.allow_score` is active. - Separate target-realization search from win-probability simulation. - Remove early exits from the probability batch. - Register an independent simulation batch size, initially 200–500 draws. - Expected metric: - Increase nonzero simulation contribution and flip-candidate rate - Improve winner-direction diagnostics - Potentially improve Brier and winner accuracy, but causal effect is not guaranteed - Helps: - All nine games in principle, especially close/moderate-edge games BOS@TEX, MIL@PIT, and MIN@LAA - Hurts: - No specific harmed game is established; risk is compute cost and introducing unvalidated probability noise - Gate: - Shadow only until independent simulation contribution is measurable. - Roll back if calibration error or winner accuracy worsens across a predefined rolling sample. ## 2. Run an isolated run-model feature/target backtest — persistent high-severity issue - Pattern: `weak_run_model_signal`, persistent for 11 slates - Exact registered levers: `training.xgb.n_estimators`, `training.xgb.max_depth`, `training.xgb.learning_rate` - Parameter direction: **abstain from direct nightly parameter movement** - Bounded action: - Add/test the registered feature-contract candidates: source-labelled no-vig market probability, park/weather environment, bullpen rest/fatigue, and platoon-adjusted lineup quality. - Compare same-fingerprint holdout R² and MAE. - Do not promote without improvement over the current R² baseline. - Expected metric: - Increase home, away, and total-run R² - Reduce team-run MAE, total-run MAE, and margin compression - Helps: - Broadly all games, with greatest relevance to DET@CHW, MIL@PIT, OAK@TB, PHI@NYM, and SD@COL - Hurts: - Potential overfit risk in BOS@TEX and MIN@LAA, where current team-run errors were only 2 - Gate: - Isolated backtest, schema-version update, then shadow deployment. No direct production tuning from this slate. ## 3. Bounded confidence-cap reduction for the strongest single-slate driver - Strongest single-slate driver: **confidence overstatement relative to outcome**, evidenced by 44.44% winner accuracy, 0.1016 weighted calibration error, and a 50% high-conviction miss rate. - Exact registered parameter: `confidence.moderate_cap` - Direction: **decrease** - Small bounded magnitude: **reduce by 0.02 from its current registered value** - Expected metric: - Lower weighted calibration error and Brier score - Reduce high-conviction miss severity - No expected direct improvement in raw winner accuracy - Helps: - BOS@TEX and MIL@PIT, the moderate-edge losses - Potentially DET@CHW and PHI@NYM, the lean-side losses - Hurts: - Could reduce useful confidence in correct moderate calls such as LAD@CIN and OAK@TB - Could understate SD@COL, which was a correct lean call - Gate: - Apply only as a shadow calibration experiment alongside existing score-variance controls. - Roll back if rolling Brier or calibration error worsens. ## 4. Optional shadow test of run-mean weighting - Exact lever: `recent.* / season.* run-mean weighting` - Direction: modestly increase adaptive recent/season run-environment weighting only within a shadow branch; do not change production from this slate alone. - Expected metric: - Reduce team-run MAE, total-run MAE, and margin compression - Helps: - DET@CHW, LAD@CIN, MIL@PIT, OAK@TB, PHI@NYM, SD@COL - Hurts: - BOS@TEX and MIN@LAA, where scores were comparatively close - This is ranked below the persistent code and run-model actions because the lever is not tied to an escalated finding and the underlying run model has near-zero predictive signal. ## Changes to abstain from today - Do not directly alter global pitcher form weights; the lever tension is 4–4. - Do not promote `context.venue_adjustment_weight` globally yet. It is a sensible shadow experiment for Lugo, Harrison, and Gordon, but it is not a registered escalated pattern. - Do not tune to the individual Nola, Fedde, Singer, or Springs outcomes. - Do not force a simulation parameter change in place of the required code repair. --- # 9. Bottom line ## Known - The slate went 4–9, or 44.44%, below the walk-forward baseline and below coin flip. - Four correct winners were also wrong on exact score. - The score model has near-zero holdout R² and materially compresses winning margins. - Confidence was overestimated in the 0.50–0.60 buckets; weighted calibration error was 0.1016. - Five losses were attributable to two wrong-conviction-side calls, two starter-projection failures, and one genuine toss-up variance result. - All nine game artifacts passed structural QA with no critical, error, or warning findings. - Six material probability movements were documented through named transformations; no material probability bypass is established. - Simulation inertness and weak run-model signal are persistent across 11 consecutive slates and are established failures. - The top-confidence pick did not have a loss streak; it was a hit on the current slate. ## Uncertain - Whether venue-split weighting will improve starter forecasts across slates; this slate provides suggestive cases but balanced global starter evidence. - Whether widening score variance will improve Brier/calibration without worsening useful correct calls. - Whether recent/season run-mean weighting can overcome the underlying lack of run-model feature signal. - The causal impact of an independent simulation distribution until the code path is repaired and measured. - The missing integrity-revision artifact’s effect on immutable revision verification. ## Single safest next action Repair and instrument the simulation probability path before changing model parameters: separate independent win-probability draws from target-realization search, remove early exit, register the batch size, and run it in shadow with rollback monitoring. This directly addresses an 11-slate established failure, preserves the existing forecast path, and produces the most valuable evidence for the next slate without overfitting today’s nine games.