# Luna deep analysis — slate 20260903 > Coverage receipt: **COMPLETE** · 9/8 sections; 9/9 games; 5/5 causes; 6/6 sources # 1. Evidence receipt and gaps ## Sources received All six available source-manifest artifacts were marked loaded: 1. **Prediction review** — `/app/output/reports/prediction_review_20260903.json`, 225,604 bytes; official outcomes and scoring errors. 2. **Forensic deep audit** — `/app/output/reports/PREDICTION_SYSTEM_DEEP_AUDIT_20260903.json`, 151,096 bytes; probability, provenance, starter, and decision contracts. 3. **Continuous learning** — `/app/output/learning/continuous_learning_daily_20260903.json`, 1,175,839 bytes; pipeline-stage quality, calibration, probability lineage, score distributions, and official starter comparisons. 4. **Detailed report** — `/app/output/reports/detailed_20260903.html`, 215,805 bytes; game and simulation rendering. 5. **Ultra-detailed player audit** — `/app/output/reports/20260903_ultra_detailed_player_audit.html`, 18,544,069 bytes; player-level prediction trace. 6. **Structural audit** — `/app/output/reports/audit_report_20260903.html`, 1,125,883 bytes; invariant findings. One manifest entry was missing: - **Integrity revision audit** — `/app/output/reports/audit_report_20260903_integrity_revision.html`; intended to provide post-remediation proof for immutable revision artifacts. Bytes were null and status was `missing`. That missing artifact is a receipt gap. It does not establish that the predictions were wrong, but it prevents a complete independent verification of revision immutability and post-Luna remediation claims. ## Slate coverage All nine expected games were reviewed: - BOS@BAL - CHW@HOU - MIA@KC - MIL@CHC - OAK@SEA - SF@PIT - STL@LAD - TB@TEX - TOR@CLE Coverage is internally complete for the available evidence: - Expected games: **9** - Games reviewed: **9** - Canonical artifacts loaded: **9** - Continuous-learning games: **9** - Forensic deep-audit games: **9** - Games with starter comparisons: **9** - Structural audits completed: **9** - Learning eligibility: **eligible**, because all artifacts passed critical/error structural gates. ## Contradictions and unresolved gaps 1. **MIA@KC probability publication delta:** calibrated pick probability was `0.5271`, while published pick probability was `0.547642`, a `+0.020542` delta. The lineage explicitly says there is “no recorded transformation” explaining it. This is the registered critical issue. 2. **MIA@KC probability representation:** the game-level diagnostic says published home probability was `0.452`, consistent with a `0.547642` away-pick probability. Thus, the numerical values may be complementary side representations rather than a mathematical contradiction. The problem is that the transformation and side orientation were not recorded clearly enough to reconcile the value. 3. **MIL@CHC structural warning:** Logan Henderson’s pitch-count summary recorded 3 hits allowed versus 2 in the pitching box score. This is a warning-level source inconsistency, not a demonstrated cause of the miss. 4. **Artifact warning:** the aggregate structural audit reports one warning and 68 informational findings, but the supplied aggregate does not identify the exact warning. The per-game evidence identifies the MIL pitch-count warning; the canonical artifact warning should still be reconciled explicitly. 5. **No player-level actuals:** MIA@KC, MIL@CHC, OAK@SEA, SF@PIT, and TB@TEX state that per-player actuals were unavailable. Player-level attribution for those games is therefore limited to projected adjustments and team/starter outcomes. 6. **No market cross-check:** no source-labelled moneyline was available for any game. Market agreement or disagreement cannot be assessed. 7. **No actual causal proof from one slate:** several causal claims are appropriately framed as drivers or risks rather than proven causal effects. Persistent findings in the cross-slate section are different: those have been escalated and must be treated as established operational failures. --- # 2. Executive diagnosis ## Outcome summary The system went **5–4**, with winner accuracy of **55.56%**. That result is only modestly above guessing and below the validated historical walk-forward baseline: - Slate winner accuracy: **0.5556** - Walk-forward accuracy: **0.560345** - Historical validation winner accuracy: **0.596659** - Walk-forward Brier: **0.243487** - Historical validation Brier: **0.239497** - Slate weighted absolute calibration error: **0.1474** The self-assessment correctly says the slate beat coin-flipping but did **not** beat the validated baseline. ## Winner accuracy - Five winners were correct: BOS, HOU, PIT, LAD, and TOR. - Four winners were wrong: MIA@KC, MIL@CHC, OAK@SEA, and TB@TEX. - The two highest-conviction games included one miss: OAK@SEA was a strong-confidence miss at 0.637 home probability. - High-conviction miss rate: **0.5**. Winner accuracy is therefore weak-to-flat, not a success merely because five calls were correct. ## Calibration quality Calibration was unstable and poor in the most consequential bucket: - 0.50–0.55 bucket: 7 games, 4 hits, average confidence 0.5359, observed hit rate 0.5714, calibration error `-0.0355`. - 0.55–0.60 bucket: 1 game, 1 hit, calibration error `-0.4409`; this is a one-game sample and should not be generalized. - 0.60–0.65 bucket: 1 game, 0 hits, calibration error `+0.6373`; this was OAK@SEA. - Weighted absolute calibration error: **0.1474**. The confidence system was not demonstrated to be globally anti-predictive, but the strong-confidence loss and the cross-slate confidence findings justify bounded confidence reduction and wider score distributions. A published home-away edge of **0.12**, such as CHW@HOU, is moderate—not a near-coin-flip justification for ignoring all directional information, and not strong evidence of a robust edge. ## Score accuracy Score accuracy was the weakest model dimension. Slate-level facts: - Every predicted margin was between 1 and 2 runs. - Actual margins ranged from 1 to 6 runs. - Three games had actual margins of at least 4. - Model margin MAE: **2.8889** - Signed margin bias: **−1.333**, indicating compressed predicted margins. - Losses had mean team-run error of **6.75** and mean total-run bias of **+2.75**. - MIL@CHC was particularly poor: 10 team runs of absolute error and total overprojection of 10. The run models have near-zero holdout explanatory power: - Away-run R²: **0.026799** - Home-run R²: **0.023403** - Total-run R²: **0.026176** - Training rows: **2,093** This supports treating exact scores as low-confidence estimates. It does not prove that every specific score miss came from the run model, nor does it justify arbitrary score-tail tuning without backtesting. ## Starter accuracy Starter projections were materially wrong in several losses: - MIA@KC: Sandy Alcantara projected 1 ER, actual 4 ER; error **+3 ER**. - MIL@CHC: Kevin Gausman projected 5 ER, actual 1 ER; error **−4 ER**. - TB@TEX: Cal Quantrill projected 3 ER, actual 0 ER; error **−3 ER**. - TOR@CLE: José Soriano projected 0 ER, actual 3 ER; error **+3 ER**, although TOR still won. The starter-weighting lever has a mixed slate result: four games identified as helped versus three games potentially harmed. That is not strong enough to justify an aggressive global reweighting from this slate alone, especially because the persistent cross-slate run-model findings are stronger than this one-slate starter balance. ## Artifact integrity Artifact integrity was mostly good but not clean: - Aggregate structural audit: **0 critical, 0 error, 1 warning, 68 info**. - Per-game audits passed except MIL@CHC, which had one warning. - No artifact defect was identified as the cause of any winner miss. - The missing integrity revision audit remains a contract-evidence gap. - MIA@KC has a **critical probability-lineage failure**, even though its structural game artifact passed. The distinction is important: prediction artifacts can pass structural QA while still containing a material probability-publication contract defect. ## Causal issue register reconciliation ### Critical: unexplained published probability delta - **ID:** `unexplained_probability_publication_delta` - **Severity:** critical - **Domain:** probability lineage - **Affected game:** MIA@KC - **Evidence:** one game has an unexplained calibrated-to-published delta of at least 2 percentage points; calibrated pick probability `0.5271`, published pick probability `0.547642`, delta `0.020542`. - **Action:** block promotion until the missing transformation is recorded and reconciled. - **Guardrail:** do not assume the probability governor was bypassed before tracing pick-side math. This is a contract/evidence failure, not proof that the wrong side was selected because of the publication delta. ### High: artifact-integrity warning - **ID:** `artifact_integrity_warning` - **Severity:** high - **Domain:** artifact integrity - **Affected games:** none assigned - **Evidence:** aggregate totals of 0 critical, 0 error, 1 warning, 68 info. - **Action:** reconcile the flagged stat or presentation contract before treating the slate as clean evidence. - **Guardrail:** a warning does not prove the model picked the wrong side. - **Known concrete warning:** MIL@CHC’s Logan Henderson pitch-count row reports 3 hits allowed versus 2 in the box score. ### High: weak run-model signal - **ID:** `weak_run_model_signal` - **Severity:** high - **Domain:** model validation - **Affected games:** all nine. - **Evidence:** home/away/total R² of `0.0234/0.0268/0.0262`. - **Action:** improve features or target definition and use isolated backtesting; do not tune score tails as though exact scores were reliable. - **Guardrail:** low positive R² means the model slightly beats the average-run guess; it does not make exact scores reliable. ### Informational: complete audit coverage - **ID:** `complete_slate_audit_coverage` - **Severity:** info - **Evidence:** 9/9 games structurally audited. - **Action:** preserve per-game audit receipts. - **Guardrail:** coverage does not cancel internal game findings. ### Informational: simulation inertness - **ID:** `simulation_inert_on_slate` - **Severity:** info on this slate, but escalated across slates. - **Evidence:** zero flip candidates and zero published flips. - **Action:** measure simulation contribution over rolling slates and separate score realization from winner flipping. - **Guardrail:** zero winner flips is not itself a defect and does not prove that no score reconciliation occurred. --- # 3. Every game ## BOS@BAL — correct winner, wrong score - **Predicted:** BOS 5, BAL 4; winner BOS. - **Actual:** BOS 6, BAL 5; winner BOS. - **Published probability:** away/BOS pick probability `0.53`; published home probability `0.470`. - **System confidence edge:** home-minus-away `−0.06`; published system edge 0.06, **lean**. - **Brier:** `0.2209`. - **Team-run absolute error:** `2`. - **Total-run bias:** `−2`, meaning total was underprojected by two. - **Conviction:** lean. ### Evidence edges and controls - Season edge, home minus away: `−0.2541`. - Recent edge: `+0.610`. - Pitcher edge: `−0.0592`. - Composite edge: `−0.2975`. - Season quality edge was below action threshold; no season margin adjustment. - Confidence path: raw `0.52165` → governed/calibrated `0.5082` → published pick probability `0.53`. - Governor applied; recency guardrail applied; score coherence and matchup-recency guardrail were recorded. - Score volatility action: `no_change`. ### Starters - Jake Bennett, BOS: projected 3 ER/17 outs; actual 3 ER/18 outs. ER error `0`. - Brandon Young, BAL: projected 4 ER/17 outs; actual 3 ER/20 outs. ER error `−1`. - Starter identities matched. ### Score path Historical ML was approximately BOS 4.95, BAL 4.03; variance stage rounded to 5–4. Later stages did not change the published score. The score missed by one run for each team but retained the correct one-run margin. ### Player adjustments and audit No player-level adjustments were recorded. Savant adjustment was applied. Structural audit: **PASS**, zero defects, seven informational findings. Forecast-file integrity was clean. - **Primary cause:** none assigned; correct directional call. - **Secondary cause:** score underestimation, but not a registered causal miss. - This is explicitly a **correct-winner/wrong-score** game. --- ## CHW@HOU — correct winner, wrong score - **Predicted:** CHW 4, HOU 5; winner HOU. - **Actual:** CHW 2, HOU 6; winner HOU. - **Published probability:** home/HOU pick probability `0.559116`. - **System confidence edge:** home-minus-away `+0.1182`; **moderate**. - **Brier:** `0.194379`. - **Team-run absolute error:** `3`. - **Total-run bias:** `+1`. - **Conviction:** moderate. ### Evidence edges and controls - Season edge: `−0.4151`. - Recent edge: `+0.066`. - Pitcher edge: `+0.0402`. - Composite edge: `+0.0566`. - Season quality edge was below action threshold. - Confidence path: raw/governor input `0.569255` → governed `0.5786` → published `0.559116`. - Governor applied; recency guardrail applied. - Score volatility: `no_change`. ### Starters - Luis Castillo, CHW: projected 4 ER/14 outs; actual 6 ER/12 outs. ER error `+2`. - Hunter Brown, HOU: projected 3 ER/17 outs; actual 2 ER/18 outs. ER error `−1`. - Starter identities matched. ### Score path and audit Predicted margin was one run; actual margin was four. No player-level adjustments were listed. Savant adjustment was applied. Structural audit: **PASS**, zero defects, eight informational findings. - **Primary cause:** none; winner call was correct. - **Secondary cause:** starter/run-score error, especially Castillo’s two extra ER and the compressed one-run margin. - This is a **correct-winner/wrong-score** game. - The 0.1182 edge is moderate, not a near-coin-flip edge. --- ## MIA@KC — winner miss - **Predicted:** MIA 5, KC 4; winner MIA. - **Actual:** MIA 3, KC 7; winner KC. - **Published probability:** lineage gives away/MIA pick probability `0.547642`; game diagnostic gives published home probability `0.452`. - **System confidence edge:** home-minus-away `−0.0953`; displayed edge 0.10, **lean**. - **Brier:** `0.299912`. - **Team-run absolute error:** `5`. - **Total-run bias:** `−1`. - **Conviction:** lean. ### Evidence edges and controls - Season edge: `−0.4839`. - Recent edge: `−0.134`. - Pitcher edge: `−0.0209`. - Composite edge: `−0.5256`. - Recent-form data actually favored KC modestly: KC recent run differential `+8`, MIA `+1`; KC recent win percentage `.600`, MIA `.500`. - Head-to-head evidence consisted of two prior KC home losses to MIA, both listed as blowouts. - Season quality moved the score toward MIA with a margin adjustment of `−0.2032`. - Confidence path: raw/governor input `0.509365` → calibrated/governed `0.5271`; recency guardrail was not applied. - The published pick probability was `0.547642`, with no recorded transformation explaining the calibrated-to-published increase. - Score volatility: `no_change`. ### Starters - Sandy Alcantara, MIA: projected 1 ER/19 outs; actual 4 ER/18 outs. ER error `+3`. - Michael Wacha, KC: projected 3 ER/19 outs; actual 3 ER/18 outs. ER error `0`. - Starter identities matched. ### Score path Historical ML: MIA approximately 4.95, KC 4.03. The model retained MIA 5–4 through variance, enhancement, Savant, fundamental review, cap, and simulation. The simulation did not alter the winner or score. ### Player adjustments and audit Twenty-four material player changes were recorded. Examples include: - Michael Wacha: better, predicted strikeouts delta `+5`. - Craig Kimbrel: worse, predicted strikeouts delta `−2`. - Griffin Conine: worse, predicted AB delta `+1.985`. - Jakob Marsee: better, predicted doubles delta `+1.878`. - Agustín Ramírez: better, predicted RBI delta `+1.74`. - Kyle Stowers: worse, predicted strikeouts delta `−1.34`. Actual player-level results were unavailable, so these adjustments cannot be validated individually. Structural audit: **PASS**, zero defects, seven informational findings; forecast integrity clean. - **Primary cause:** `lean_side_miss` — small edge on the losing side. - **Secondary cause:** starter projection, specifically Alcantara’s three-ER underprojection; KC offense was under-projected by three. - **Additional artifact issue:** critical probability-lineage defect, not proven to have caused the miss. --- ## MIL@CHC — winner miss and largest score failure - **Predicted:** MIL 7, CHC 6; winner MIL. - **Actual:** MIL 1, CHC 2; winner CHC. - **Published probability:** away/MIL pick probability `0.533686`; game diagnostic gives home probability `0.466`. - **System confidence edge:** home-minus-away `−0.0674`; displayed edge 0.07, **lean**. - **Brier:** `0.284821`. - **Team-run absolute error:** `10`. - **Total-run bias:** `+10`. - **Conviction:** lean. ### Evidence edges and controls - Season edge: `−0.1714`. - Recent edge: `+0.290`. - Pitcher edge: `−0.1494`. - Composite edge: `−0.5304`. - Recent form showed CHC with +34 run differential and 8.6 runs scored per game despite a .400 win percentage; MIL had +9 run differential. - Confidence path: raw/governor input `0.54` → governed `0.5082` → published `0.533686`. - Governor applied; recency guardrail applied. - Score volatility: `no_change`, despite a recent-environment total of `12.05` and low-score probability `0.5003`. ### Starters - Logan Henderson, MIL: projected 3 ER/18 outs; actual 2 ER/21 outs. ER error `−1`. - Kevin Gausman, CHC: projected 5 ER/17 outs; actual 1 ER/21 outs. ER error `−4`. - Identities matched. ### Score path This was the clearest score-path deterioration: - Historical ML: MIL 4.82, CHC 4.04. - Variance stage: 5–4. - Enhanced prediction: expanded to 7–6. - Savant stage retained 7–6. - Fundamental review, cap, simulation, and publication all retained 7–6. - Actual result was 1–2. The transition from variance to enhanced prediction increased the score total by four runs while preserving the wrong MIL direction. Later stages did not recover the error. ### Player adjustments and audit Twenty-two material player changes were recorded, including: - Christian Yelich: predicted RBI delta `+2.6`. - David Hamilton: predicted hits delta `+2.35`. - William Contreras: predicted AB delta `−2.13`. - Kevin Gausman: neutral starter adjustment with predicted strikeouts delta `+2`. - Seiya Suzuki: predicted RBI delta `+1.32`. - Michael Busch: predicted hits delta `+1.28`. Actual player-level results were unavailable. Structural audit: **WARNING**, risk score 10. The concrete warning was a pitch-count mismatch for Logan Henderson: summary hits allowed `3`, box score `2`. It was not identified as the reason for the winner miss. - **Primary cause:** `starter_projection`; Gausman was projected for five ER and allowed one. - **Secondary cause:** `run_environment_over`; total runs were overprojected by ten. - The model also materially expanded the score at the enhanced stage without evidence that the winner direction improved. --- ## OAK@SEA — strong-confidence winner miss - **Predicted:** OAK 4, SEA 6; winner SEA. - **Actual:** OAK 7, SEA 4; winner OAK. - **Published probability:** home/SEA `0.637251`. - **System confidence edge:** home-minus-away `+0.2745`; **strong**. - **Brier:** `0.406089`, worst on the slate. - **Team-run absolute error:** `5`. - **Total-run bias:** `−1`. - **Conviction:** strong. ### Evidence edges and controls - Season edge: `+0.6111`. - Recent edge: `+0.006`. - Pitcher edge: `−0.0394`. - Composite edge: `+0.6109`. - Season quality edge was `+1.05` after adjustment; season margin adjustment `+0.2567`. - Recent form was weak for both teams, but SEA’s recent run differential was `−21` versus OAK’s `−6`. - Historical head-to-head listed several SEA wins, which likely reinforced the home side. - Confidence path: raw/governor input `0.62335` → governed `0.6045` → published `0.637251`. - No recency guardrail was applied. - Score volatility: `no_change`. ### Starters - Jack Perkins, OAK: projected 3 ER/12 outs; actual 3 ER/9 outs. ER error `0`. - Kade Anderson, SEA: projected 3 ER/13 outs; actual 4 ER/15 outs. ER error `+1`. - Starter projections were broadly accurate in ER terms. ### Score path - Historical ML: OAK 4.02, SEA 5.04. - Variance: 4–5. - Enhanced prediction: 4–6. - Savant and later stages retained 4–6. - Actual: 7–4 OAK. The score and winner moved toward SEA before the later stages, and no later stage corrected that direction. ### Player adjustments and audit Twenty-six material player changes were recorded. Examples: - Cole Young: worse, predicted RBI delta `+3.67`. - Julio Rodríguez: worse, predicted AB delta `−2.14`. - Michael Stefanic: worse, predicted BB delta `+1.667`. - Cal Raleigh: worse, predicted AB delta `−1.65`. - Brian Serven: boosted, predicted RBI delta `+1.31`. - Lawrence Butler: worse, predicted AB delta `+1.25`. Actual player-level results were unavailable. Structural audit: **PASS**, zero defects, seven informational findings; forecast integrity clean. - **Primary cause:** `wrong_conviction_side` — the system backed SEA with genuine conviction and lost. - **Secondary cause:** `lean_side_miss` is also recorded, but it is subordinate to the strong published confidence. - This is the slate’s clearest calibration failure and a correct-winner/wrong-score classification does **not** apply because the winner was wrong. --- ## SF@PIT — correct winner, wrong score - **Predicted:** SF 5, PIT 6; winner PIT. - **Actual:** SF 2, PIT 5; winner PIT. - **Published probability:** home/PIT `0.53`. - **System confidence edge:** home-minus-away `+0.06`; **lean**. - **Brier:** `0.2209`. - **Team-run absolute error:** `4`. - **Total-run bias:** `+4`. - **Conviction:** lean. ### Evidence edges and controls - Season edge: `+0.3866`. - Recent edge: `−0.438`. - Pitcher edge: `+0.0199`. - Composite edge: `+0.3269`. - Confidence path: raw/governor input `0.53` → governed `0.5082` → published `0.53`. - Governor and recency guardrail applied. - Score volatility: `no_change`. ### Starters - Blade Tidwell, SF: projected 2 ER/15 outs; actual 2 ER/15 outs. Error `0`. - Lake Bachar, PIT: projected 1 ER/15 outs; actual 0 ER/7 outs. Error `−1`. - Identities matched. ### Score path and audit Prediction remained 5–6; actual PIT won 5–2. No player-level actuals were available. Savant adjustment applied. Structural audit: **PASS**, zero defects, seven informational findings; forecast integrity clean. - **Primary cause:** none; winner correct. - **Secondary cause:** SF offense overprojection and total overprojection. - This is a **correct-winner/wrong-score** game. --- ## STL@LAD — correct winner, wrong score - **Predicted:** STL 4, LAD 5; winner LAD. - **Actual:** STL 2, LAD 3; winner LAD. - **Published probability:** home/LAD `0.53`. - **System confidence edge:** home-minus-away `+0.06`; **lean**. - **Brier:** `0.2209`. - **Team-run absolute error:** `4`. - **Total-run bias:** `+4`. - **Conviction:** lean. ### Evidence edges and controls - Season edge: `+0.8594`. - Recent edge: `−0.062`. - Pitcher edge: `+0.0015`. - Composite edge: `+0.6405`. - Season quality edge after adjustment: `+1.2424`; margin adjustment `+0.3609`. - Confidence path: raw/governor input `0.53` → governed `0.5786` → published `0.53`. - Governor and recency guardrail applied. - Diagnostic flag: `large_savant_run_environment_swing`. - Score volatility action: `supported_blowout_margin_expansion`. ### Starters - Quinn Mathews, STL: projected 3 ER/14 outs; actual 1 ER/15 outs. Error `−2`. - Tarik Skubal, LAD: projected 0 ER/18 outs; actual 2 ER/17 outs. Error `+2`. - Identities matched. ### Score path and audit The system correctly selected LAD but projected a five-run total versus the actual five-run total? The predicted total was 9 and actual total 5, so total bias was +4. Structural audit: **PASS**, zero defects, seven informational findings; forecast-file integrity clean. - **Primary cause:** none; winner correct. - **Secondary cause:** run-environment/score-tail miss despite a supported blowout-margin expansion action. - This is a **correct-winner/wrong-score** game. --- ## TB@TEX — winner miss - **Predicted:** TB 5, TEX 4; winner TB. - **Actual:** TB 0, TEX 6; winner TEX. - **Published probability:** away/TB `0.533686`; game diagnostic gives home probability `0.466`. - **System confidence edge:** home-minus-away `−0.0674`; **lean**. - **Brier:** `0.284821`. - **Team-run absolute error:** `7`. - **Total-run bias:** `+3`. - **Conviction:** lean. ### Evidence edges and controls - Season edge: `−0.654`. - Recent edge: `−0.3767`. - Pitcher edge: `−0.0494`. - Composite edge: `−0.7425`. - Both season and recent evidence favored TB, but actual game conditions strongly contradicted the score. - Confidence path: raw/governor input `0.49665` → governed `0.5082` → published `0.533686`. - Governor and recency guardrail applied. - Score volatility: `no_change`; low-score probability was `0.6951`. ### Starters - Shane McClanahan, TB: projected 2 ER/16 outs; actual 4 ER/15 outs. Error `+2`. - Cal Quantrill, TEX: projected 3 ER/17 outs; actual 0 ER/21 outs. Error `−3`. - Identities matched. ### Score path and audit Prediction remained TB 5–4 from historical ML through publication. Actual was TB 0–TEX 6. No player-level actuals were available. Twenty-four material player changes were recorded, including Jorge Mateo, Ezequiel Duran, Cal Quantrill, Corey Seager, and others, but their actual effects cannot be validated. Structural audit: **PASS**, zero defects, seven informational findings; forecast integrity clean. - **Primary cause:** `lean_side_miss`. - **Secondary cause:** starter projection, especially Quantrill’s three-ER overprojection and McClanahan’s two-ER underprojection. - This is not a score-only miss; the winner was wrong. --- ## TOR@CLE — correct winner, wrong score - **Predicted:** TOR 4, CLE 3; winner TOR. - **Actual:** TOR 6, CLE 3; winner TOR. - **Published probability:** away/TOR `0.546553`; game diagnostic gives home probability `0.453`. - **System confidence edge:** home-minus-away `−0.0931`; **lean**. - **Brier:** `0.205614`. - **Team-run absolute error:** `2`. - **Total-run bias:** `−2`. - **Conviction:** lean. ### Evidence edges and controls - Season edge: `+0.2137`. - Recent edge: `+0.034`. - Pitcher edge: `−0.0401`. - Composite edge: `−0.1422`. - Season quality edge below action threshold. - Confidence path: raw/governor input `0.545724` → governed `0.5439` → published `0.546553`; lineage says the published value matches the calibrated/governed value. - Governor and recency guardrail applied. - Score volatility: `no_change`. ### Starters and audit - José Soriano, TOR: projected 0 ER/18 outs; actual 3 ER/14 outs. Error `+3`. - Tanner Bibee, CLE: projected 3 ER/18 outs; actual 5 ER/12 outs. Error `+2`. - Both starter identities matched. - Structural audit: **PASS**, zero defects, nine informational findings; forecast-file integrity clean. - **Primary cause:** none; winner correct. - **Secondary cause:** score underprojection despite errors by both starters. - This is a **correct-winner/wrong-score** game. --- # 4. Miss reconstruction ## MIA@KC ### Earliest wrong-direction stage The historical ML stage already pointed toward the wrong result: - Historical expected score: MIA approximately 4.95, KC 4.03. - Every subsequent score stage retained MIA ahead, ending at 5–4. - No later stage changed the winner direction. The earliest model movement toward the wrong result was therefore the **historical model / pre-enhancement directional baseline**, reinforced by season-quality and head-to-head treatment. ### Error classification - **Bad input:** not established. Starter identities matched and structural audit passed. - **Bad weighting:** plausible. Recent-form data favored KC modestly, while season and head-to-head evidence pushed toward MIA. - **Calibration:** the published away probability was materially higher than the calibrated value without a recorded transformation. This is a contract failure, but causation to the miss is unproven. - **Starter error:** established as a material score contributor; Alcantara was three ER worse than projected. - **Offense/bullpen miss:** KC offense was under-projected by three. Bullpen-specific actual evidence is absent. - **Run-environment miss:** not primary; total was underprojected by one. - **Irreducible variance:** possible, especially given the lean edge and unavailable player-level actuals. The `−0.0953` home-minus-away edge was a lean, not strong conviction. It should not be interpreted as a near coin flip, but it also did not support aggressive confidence. ## MIL@CHC ### Earliest wrong-direction stage Historical ML already favored MIL: - Historical score: MIL 4.82, CHC 4.04. - Variance stage retained MIL ahead at 5–4. - Enhanced prediction expanded the score to 7–6 while retaining MIL. - Savant, fundamental review, cap, simulation, and publication made no further directional change. The first clear **score-quality regression** occurred at the variance-to-enhanced transition: total score rose from 9 to 13 runs, while winner direction remained wrong. The winner-direction error itself was present at historical ML. ### Error classification - **Bad input:** one pitch-count warning exists for Henderson, but it does not explain the Gausman miss and is not proven causal. - **Bad weighting:** plausible; recent-form and H2H signals supported an extreme run environment despite the eventual three-run total. - **Calibration:** lean probability only; not the main issue. - **Starter error:** established and major. Gausman was projected for 5 ER and allowed 1. - **Offense/bullpen miss:** both offenses were overprojected; bullpen effects are not separately available. - **Run-environment miss:** established and dominant. The model forecast 13 total runs and actual total was 3. - **Irreducible variance:** possible, but the scale of the score miss is also consistent with a weak run model. ## OAK@SEA ### Earliest wrong-direction stage The historical ML stage favored SEA 5.04–4.02. Variance retained SEA at 5–4, and enhancement widened SEA’s margin to 6–4. No later stage corrected it. Thus the earliest wrong-direction evidence was the **historical model**, with later enhancement and season-quality weighting reinforcing the error. ### Error classification - **Bad input:** not established. Structural audit passed. - **Bad weighting:** strongly plausible. Season edge and historical head-to-head favored SEA, while recent form was essentially neutral-to-negative for both teams. The model published 0.637 despite limited recent separation and a small pitcher edge against SEA. - **Calibration:** established as a high-confidence failure; Brier `0.406089`, strong pick lost. - **Starter error:** not primary; Perkins was exact in ER and Anderson was only one ER worse. - **Offense/bullpen miss:** OAK offense was underprojected by three; bullpen/actual player effects unavailable. - **Run-environment miss:** secondary; total was underprojected by one, but the main problem was choosing SEA. - **Irreducible variance:** possible, but insufficient to explain away the conviction error. ## TB@TEX ### Earliest wrong-direction stage Historical ML favored TB approximately 4.56–3.63. All stages retained TB at 5–4. No stage moved toward TEX. The wrong direction therefore entered at the **historical model**, and the later controls failed to reverse it. ### Error classification - **Bad input:** not established; structural audit passed. - **Bad weighting:** plausible. Season and recent edges both favored TB, but the model did not adequately represent the possibility of a TEX run/starting-pitching outcome. - **Calibration:** lean, not strong; no special high-confidence failure. - **Starter error:** established; Quantrill was three ER better than projected and McClanahan two ER worse. - **Offense/bullpen miss:** TB was shut out relative to five projected runs; bullpen/player actuals unavailable. - **Run-environment miss:** material; actual total 6 versus predicted 9, with a completely different allocation. - **Irreducible variance:** meaningful possibility, especially in a lean game. ## Pipeline-stage audit across all stages and transitions | Stage | Winner accuracy | Winner hits | Team-run MAE | Total-run MAE | Margin MAE | Change versus historical | |---|---:|---:|---:|---:|---:|---:| | historical_ml | 55.56% | 5 | 2.0006 | 2.5017 | 2.7246 | baseline | | variance_applied | 44.44% | 4 | 2.0556 | 2.5556 | 2.7778 | worse | | enhanced_prediction | 55.56% | 5 | 2.3333 | 3.1111 | 2.8889 | worse | | embedded_savant_run_environment | 55.56% | 5 | 2.3333 | 3.1111 | 2.8889 | unchanged from enhanced | | fundamental_edge_target_review | 55.56% | 5 | 2.3333 | 3.1111 | 2.8889 | unchanged | | capped_target | 55.56% | 5 | 2.3333 | 3.1111 | 2.8889 | unchanged | | simulation_result | 55.56% | 5 | 2.3333 | 3.1111 | 2.8889 | unchanged | | published_final | 55.56% | 5 | 2.3333 | 3.1111 | 2.8889 | unchanged | Transitions: 1. **Historical → variance:** BOS moved BOS → TIE. No games improved. Winner accuracy fell by one hit, team-run MAE rose `0.055`, and margin MAE rose `0.0532`. 2. **Variance → enhanced:** BOS recovered TIE → BOS. No games regressed in winner direction, but team-run MAE rose `0.2777` and margin MAE rose `0.1111`. 3. **Enhanced → embedded Savant:** no changes. 4. **Embedded Savant → fundamental review:** no changes. 5. **Fundamental review → capped target:** no changes. 6. **Capped target → simulation:** no changes. 7. **Simulation → published final:** no changes. The BOS recovery does not erase the intermediate variance-stage regression. Likewise, later unchanged stages do not demonstrate that they validated the earlier score expansion. The simulation contributed no winner-direction intervention on this slate. --- # 5. Cross-game counterfactuals ## Lever tension: widen score distributions - **Lever:** `score.variance_multiplier / score.variance_floor` - **Helps:** OAK@SEA, where a 0.27-confidence pick lost and wider tails could reduce overconfidence and better represent OAK’s upset/run outcome. - **Hurts:** none identified in the supplied tension entry. - **Net evidence:** 1 helped, 0 hurt. This supports a small, bounded shadow experiment, not a large production change. It primarily targets calibration and tail coverage; exact score MAE could improve or worsen. ## Lever tension: pitcher form weights and prior games - **Lever:** `player.pitcher.v4 form weights / player.pitcher_prior_games` - **Helps:** MIA@KC, MIL@CHC, TB@TEX, TOR@CLE; starter ER errors were 3, 4, 3, and 3 respectively. - **Hurts:** BOS@BAL, OAK@SEA, SF@PIT; starter ER projections were accurate to within one ER and reweighting could disturb those calls. - **Net evidence:** 4 helped versus 3 hurt. - **Net conclusion:** evidence supports investigation but not aggressive reweighting. The balance is too close, and starter outcomes include substantial variance. ## Lever tension: recent/season run-mean weighting - **Lever:** `recent.* / season.* run-mean weighting` - **Helps:** MIA@KC, MIL@CHC, OAK@SEA, SF@PIT, STL@LAD, TB@TEX; six games had material team-run errors where altered run-mean weighting could plausibly help. - **Hurts:** BOS@BAL and TOR@CLE, where team-run error was only two. - **Net evidence:** 6 helped versus 2 hurt. - **Net conclusion:** directional evidence supports review, but the near-zero run-model R² means the proper response is feature/target investigation, not simply turning a nightly weighting knob. ## One-game fixes that must not be made - Do not globally reverse the historical/season-quality preference because OAK@SEA lost; that would damage games such as STL@LAD and the correct PIT/LAD/TOR calls. - Do not aggressively increase pitcher recency weight solely because Gausman, Alcantara, and Quantrill missed; BOS, OAK, and SF had comparatively accurate starter ER projections. - Do not widen all score distributions dramatically based only on MIL@CHC. The run-model issue is persistent, but the correct response is a bounded shadow test plus code/data investigation. - Do not treat MIA’s unexplained probability delta as proof that the governor was bypassed or that the away pick should have been removed until side-oriented lineage is reconciled. - Do not infer that simulation was useless merely because it produced zero flips; the evidence shows no winner intervention, not necessarily zero score-realization work. - Do not promote the slate as clean while the missing integrity revision audit and MIA lineage defect remain unresolved. --- # 6. Data and contract defects ## Critical findings ### Unexplained probability publication delta - Affected game: MIA@KC. - Calibrated probability: `0.5271`. - Published pick probability: `0.547642`. - Unexplained delta: `0.020542`. - Transformation text: “no recorded transformation explains the publication delta.” - Required action: trace raw model → calibration → governor → score coherence/recency → published value and block promotion until reconciled. This is a prediction-contract/evidence defect, not necessarily a model-side selection defect. ## Error findings - Aggregate structural audit reports **zero errors**. - No per-game structural audit reported an error severity. - Therefore, there are no supplied error-level findings to enumerate. ## Warning findings 1. **MIL@CHC pitch-count mismatch** - Logan Henderson’s summary reports hits allowed `3`; pitching box score reports `2`. - Rule: `PITCH_COUNT_PITCHER_MATCH`. - Source: simulation. - Risk score: 10. - This is a data/audit consistency defect. It was not demonstrated to cause the prediction miss. 2. **Aggregate artifact warning** - Structural totals contain one warning. - The supplied evidence identifies the MIL mismatch as the concrete warning, but the aggregate canonical warning should be reconciled explicitly. 3. **Missing integrity revision audit** - Not marked as a structural warning in the totals, but it is a receipt and artifact-integrity gap because the manifest explicitly marks the revision-proof artifact missing. ## Informational findings - Aggregate structural audit reports **68 info** findings. - Per-game information counts are: - BOS: 7 - CHW: 8 - MIA: 7 - MIL: 9 - OAK: 7 - SF: 7 - STL: 7 - TB: 7 - TOR: 9 - The supplied JSON does not expose the text of every informational finding, so no additional descriptions should be invented. - `complete_slate_audit_coverage` and `simulation_inert_on_slate` are explicitly informational at the slate level. ## Model defects versus presentation/artifact defects ### Prediction-model/process defects - Near-zero run-target holdout signal: total R² `0.026176`, home `0.023403`, away `0.026799`. - Persistent score-margin compression: every predicted margin was 1–2 runs while actual margins included 4, 4, 4, and 6. - Simulation winner-direction intervention was inert for seven consecutive slates and requires code/diagnostic verification. - Starter projection misses were material in four games, but the one-slate lever balance is mixed. ### Presentation/artifact/contract defects - MIA probability lineage lacks a recorded material transformation. - MIL pitch-count summary conflicts with the box score. - Integrity revision audit is missing. - Per-game forecast files were otherwise reported clean; these bugs are not established as causes of winner misses. ## Complete pick-side probability lineage audit The evidence provides the following chain for all nine games: - **Raw model pick probability** - **Governor input** - **Calibrated/governor output** - **Score coherence probability** - **Recency guardrail state** - **Published pick probability** - Transformation description and unexplained-delta flag Eight games have a recorded transformation and `unexplained_material_delta: false`. MIA@KC is the sole material bypass/lineage exception. Important examples: - BOS: calibrated `0.5082` → published pick `0.53`, with score coherence and recency guardrail recorded. - OAK: calibrated `0.6045` → published `0.637251`, explicitly attributed to score coherence alignment. - TOR: calibrated `0.5439` → published `0.546553`, described as matching the calibrated/governed value. - MIA: calibrated `0.5271` → published `0.547642`, with no transformation recorded. A bypass should be called only for MIA because its material delta lacks a recorded transformation. The other deltas are documented, even where the values appear surprising. Coverage matrix reconciliation supports complete game coverage, but not a clean slate: full structural coverage coexists with the MIA contract defect, the MIL warning, and the missing integrity revision artifact. --- # 7. Cross-slate persistence Six prior slates were loaded, forming a seven-slate window from 2026-08-28 through 2026-09-03. Persistent escalations are not single-slate noise. ## Finding streaks | Finding | Consecutive-slate count | Present now? | Interpretation | |---|---:|---|---| | high_raw_probability_cluster | 0 | No | Occurred in 2 of 7 slates, but not currently persistent | | pipeline_winner_instability | 0 | No | Occurred in 3 of 7, not currently present | | score_margin_compression | 0 | No | Occurred in 3 of 7, not currently present as a registered streak, though this slate shows the same structural symptom | | simulation_inert_on_slate | 7 | Yes | Escalated established failure | | single_slate_calibration_instability | 0 | No | Occurred in 5 of 7, but no current consecutive streak | | weak_run_model_signal | 7 | Yes | Escalated established failure | The current slate independently exhibits margin compression, but the finding-streak record reports `score_margin_compression` with zero consecutive slates. Those two statements should not be conflated: the raw current-slate margin data show compression, while the registered streak counter does not classify it as currently persistent. ## Escalated finding: simulation inertness - **Streak:** 7 consecutive slates. - **Present now:** yes. - **Target lever:** `simulation.allow_score`. - **Metric group:** run model. - **Metrics hurt:** team-run MAE and total-run MAE; also potentially winner-probability agreement diagnostics. - **Direction:** confirm the simulation module is running and not silently no-oping. - **Evidence:** seven-slate persistence; this slate had zero flip candidates and zero published flips. - **Required interpretation:** established code/diagnostic failure, not a parameter-tuning invitation. The self-assessment identifies a concrete root cause: the same draws are used to search for a target-realizing play-by-play, and the loop may early-exit after finding one, with a default draw count around two. That makes the resulting win rate non-probabilistic and the flip signal vacuous. This requires operator/code action. ## Escalated finding: weak run-model signal - **Streak:** 7 consecutive slates. - **Present now:** yes. - **Target levers listed:** `training.xgb.n_estimators`, `training.xgb.max_depth`, `training.xgb.learning_rate`. - **Param actionable:** no. - **Metrics hurt:** team-run MAE and total-run MAE. - **Direction:** investigate features and target through an isolated training backtest. - **Evidence:** persistent seven-slate finding plus holdout R² near zero in all run targets. The correct action is not to blindly alter the three training hyperparameters. The registered response is a feature-contract and target investigation involving market no-vig probability where available, park/weather, bullpen rest/fatigue, and platoon-adjusted lineup quality, followed by same-fingerprint holdout comparison. ## Metric trends All directions must be stated explicitly: - **Margin bias:** improving. Seven-slate series mean `−2.4113`; current value `−1.778`. - **Mean team-run error on losses:** regressing. Mean `6.2179`; series ends at `6.75`. - **Weighted calibration error:** improving. Seven-slate mean `0.2078`; current `0.1474`. - **Winner accuracy:** flat. Seven-slate mean `0.5615`; current `0.5556`. - **Top-confidence pick:** loss streak of **3**. The sequence was hit, hit, miss, hit, miss, miss, miss across the seven dates, with misses on 2026-09-01, 2026-09-02, and 2026-09-03. The top-confidence loss streak is a material warning even though the high-raw cluster is not present this slate. ## Raw structural evidence - **Margin distribution:** all nine predicted margins were ≤2, while three actual margins were ≥4. This confirms current compression. - **Pipeline transitions:** variance application regressed BOS from a winner to a tie and reduced winner accuracy; enhancement later recovered BOS but worsened run MAE. Later recovery does not erase the intermediate regression. - **Calibration monitoring:** current weighted absolute calibration error is still high at `0.1474`; the 0.60–0.65 bucket had zero hits. - **High raw cluster:** 0 games at the 0.65 threshold this slate; no current cluster. - **Simulation control:** zero observed games with a simulation flip candidate and zero published flips; diagnostic-only status. --- # 8. Adjustments for today Recommendations are ranked by persistence and evidence strength first, then scope and regression risk. ## 1. Repair and independently validate the simulation path - **Pattern:** seven-slate persistent simulation inertness. - **Registered parameter/lever:** `simulation.allow_score`. - **Recommended direction:** do not numerically tune it first; confirm it is enabled and functioning. If disabled or bypassed, restore the score-realization path. - **Bounded change:** run an independent, unsteered probability batch of approximately 200–500 draws, or a closed-form Skellam/Poisson-difference calculation, without early exit. Keep target-realization search separate from winner-probability estimation. Register batch size. - **Expected metric:** improve the validity of `team_run_mae`/`total_run_mae` diagnostics and make simulation agreement/flip metrics informative; direction for winner accuracy and Brier is expected to improve, but not guaranteed. - **Helps:** all games where simulation could have independently challenged a compressed or wrong winner, especially MIA@KC, MIL@CHC, OAK@SEA, and TB@TEX. - **Hurts/regression risk:** a small risk of altering correct calls such as BOS@BAL, CHW@HOU, SF@PIT, STL@LAD, and TOR@CLE. Use a shadow gate and rollback monitor. - **Action type:** code/diagnostic fix, not a forced parameter move. ## 2. Investigate the persistent near-zero run model - **Pattern:** seven consecutive slates with weak run-model signal. - **Registered target levers:** `training.xgb.n_estimators`, `training.xgb.max_depth`, `training.xgb.learning_rate`. - **Recommended direction:** no direct hyperparameter move; the evidence explicitly marks this as not parameter-actionable. - **Bounded change:** add or validate feature-contract candidates—source-labelled no-vig market probability where available, park/weather environment, bullpen fatigue/rest, and platoon-adjusted lineup quality—then run an isolated same-fingerprint backtest and bump the model schema version if semantics change. - **Expected metric:** team-run MAE and total-run MAE should decrease; home/away/total holdout R² should increase from approximately 0.02. If they do not, roll back the feature change. - **Helps:** broadly MIA@KC, MIL@CHC, OAK@SEA, SF@PIT, STL@LAD, and TB@TEX, where run-score errors were material. - **Hurts/regression risk:** BOS@BAL and TOR@CLE currently had only two runs of team error; changes must be shadowed to ensure their score quality does not deteriorate. - **Action type:** operator-led code/data/backtest work. Do not force a nightly XGBoost knob adjustment. ## 3. Apply a small score-variance shadow increase for the strongest single-slate driver - **Single-slate driver:** OAK@SEA, a 0.637251 published home probability that lost, with a five-run team-score error. - **Registered lever:** `score.variance_multiplier` or, preferably for a bounded floor test, `score.variance_floor`. - **Recommended direction:** increase slightly, approximately **5% relative** in shadow mode; do not make a large production jump. - **Expected metric:** lower high-confidence overstatement and improve Brier/calibration error; improve tail coverage and potentially reduce margin compression. Exact run MAE direction is uncertain. - **Helps:** OAK@SEA directly; lever-tension evidence identifies one helped game and no harmed games. - **Hurts/regression risk:** wider tails can reduce apparent sharpness in correct close calls such as BOS@BAL, SF@PIT, and TOR@CLE. Monitor those games and roll back if Brier or team-run MAE deteriorates. - **Rationale:** the high-raw-confidence cluster is not currently persistent, so this is a small shadow experiment rather than a broad confidence-cap overhaul. ## 4. Reconcile MIA probability lineage before any promotion - **Registered issue:** `unexplained_probability_publication_delta`. - **Parameter move:** none. - **Action:** reconstruct the pick-side math and record the exact calibrated → governor → score-coherence → recency → published transformation, including whether each value is home-side or pick-side. - **Expected metric:** zero unexplained material deltas and a clean probability contract. - **Affected game:** MIA@KC. - **Risk:** none from a prediction lever; the risk is continued publication of untraceable probabilities. - **Promotion status:** block until reconciled. ## 5. Do not globally reweight pitcher form today - **Lever considered:** `player.pitcher.v4 form weights / player.pitcher_prior_games`. - **Evidence:** 4 potentially helped versus 3 potentially harmed. - **Reason for abstention:** mixed one-slate evidence and no persistent escalated finding targeting this lever. - **Next step:** evaluate across a larger registered backtest, particularly starter ER MAE and outs MAE, before moving the parameter. --- # 9. Bottom line ## Known - The slate was **5–4**, with winner accuracy **55.56%**, below the walk-forward baseline of **56.03%**. - Exact-score performance was weak: predicted margins were compressed to 1–2 runs, while actual margins reached 6. - Run-model holdout signal is persistently near zero: total R² `0.026176`. - Four misses were MIA@KC, MIL@CHC, OAK@SEA, and TB@TEX. - OAK@SEA was the strongest single-slate calibration failure: 0.637 home probability, wrong winner, Brier `0.406089`. - Starter projection errors materially contributed to MIA@KC, MIL@CHC, and TB@TEX. - Simulation produced no winner-direction intervention and this inertness has persisted for seven slates. - The probability lineage has one critical unresolved defect at MIA@KC. - Structural coverage was complete, but the slate is not artifact-clean because of the MIL warning and missing integrity revision audit. ## Uncertain - Whether MIA’s unexplained publication delta changed the selected side or only reflects an unrecorded side-orientation transformation. - Whether pitcher-form reweighting would improve future starter accuracy; this slate showed 4-versus-3 mixed evidence. - Whether wider variance will improve exact-run MAE rather than only calibration and tail coverage. - The exact player-level causes of several offensive misses, because actual player-level audit data were unavailable. - Whether individual winner misses were bad weighting or irreducible game variance once the run-model and starter errors are separated. ## Safest next action The single safest action is to **repair and independently verify the simulation/probability pipeline before promotion**: separate score-target realization from an independently sampled winner-probability distribution, eliminate the early-exit/no-op behavior, register the draw-count parameter, and reconcile MIA’s published probability lineage. This directly addresses a seven-slate established failure and a critical evidence-contract defect without making an unvalidated global model adjustment.