Luna deep analysis — slate 20260912

Coverage receipt: COMPLETE · 9/8 sections; 15/15 games; 6/6 causes; 6/6 sources

1. Evidence receipt and gaps

Sources received

Six of seven manifest entries were loaded:

1. Prediction review — /app/output/reports/prediction_review_20260912.json, 377,165 bytes; official outcomes and scoring errors.
2. Forensic deep audit — /app/output/reports/PREDICTION_SYSTEM_DEEP_AUDIT_20260912.json, 240,873 bytes; probability, provenance, starter, and decision contracts.
3. Continuous learning — /app/output/learning/continuous_learning_daily_20260912.json, 2,581,062 bytes; pipeline-stage quality, calibration, probability lineage, score distribution, and starter comparisons.
4. Detailed report — /app/output/reports/detailed_20260912.html, 354,592 bytes; full game and simulation rendering.
5. Ultra-detailed player audit — /app/output/reports/20260912_ultra_detailed_player_audit.html, 34,118,143 bytes; player-level prediction trace.
6. Structural audit — /app/output/reports/audit_report_20260912.html, 1,868,598 bytes; invariant findings.

The following manifest item is missing:

1. Integrity revision audit — /app/output/reports/audit_report_20260912_integrity_revision.html; intended purpose: post-Luna remediation proof for immutable revision artifacts.

This missing artifact is a receipt gap, but it is not itself evidence that the forecasts were altered. The game-level integrity audits report zero critical and error findings.

Coverage confirmation

All 15 expected games were reviewed:

BAL@TOR, CHW@STL, CIN@MIL, CLE@MIN, COL@DET, HOU@TB, KC@BOS, LAA@WAS, LAD@MIA, NYM@NYY, PHI@ATL, PIT@CHC, SD@SF, SEA@OAK, TEX@ARI.

Coverage is complete at the structural level:

• Canonical artifacts loaded: 15/15
• Continuous-learning games: 15/15
• Forensic deep-audit games: 15/15
• Starter comparisons: 15/15
• Structural audits: 15/15
• Games reviewed: 15
• Hits: 8
• Misses: 7

Important gaps and contradictions

• The player feed contains no rows for the entire 20260912 slate. Consequently, player-level actual-performance attribution is unavailable for every game, even where player adjustments were applied pregame.
• No source-labelled moneyline was available, so market agreement and market cross-checks could not be evaluated.
• The missing integrity-revision audit prevents confirmation of that specific immutable-revision artifact, although the available structural audits all passed.
• Two probability publications contain unexplained calibrated-to-published deltas: CHW@STL and COL@DET. This is a material contract gap, not a demonstrated governor bypass.
• The causal register says the score-compression issue affects five named games, while the underlying evidence is slate-wide: 12/15 predicted margins were ≤2 runs, versus 8/15 actual margins ≥4. The register's affected-game list should therefore be treated as the most directly implicated subset, not the complete set of games exhibiting compressed margins.
• The pipeline metrics show a later recovery from variance_applied to enhanced_prediction, but that recovery does not erase the intermediate winner-direction regressions.

2. Executive diagnosis

Overall result

The slate was weak as a prediction product:

• Published slate accuracy: 8/15 = 53.33%
• Walk-forward baseline accuracy: 55.94%
• Historical holdout winner accuracy: 59.82%
• Slate Brier mean, from the listed game scores: approximately 0.258
• Weighted absolute calibration error: 0.1711
• High-conviction games: 4
• High-conviction misses: 3
• High-conviction miss rate: 75%
• Mean team-run error on losses: 6.143
• Mean total-run bias on losses: -1.286
• Model margin MAE at the published stage: 4.133
• Model signed margin bias: +0.667 using the reported margin orientation

The headline is therefore: winner accuracy was barely above chance and below the validated baseline; calibration was poor and anti-predictive at high conviction; score accuracy was materially poor; starter projections were mixed; artifact integrity was mostly clean except for two unresolved probability-publication lineage failures and one missing revision-audit artifact.

A correct winner does not make the prediction good. Several correct calls had very large score errors, including:

• CIN@MIL: winner correct, team-run error 6
• COL@DET: winner correct, team-run error 7
• SEA@OAK: winner correct, team-run error 17

Winner accuracy

The system selected the correct winner in 8 of 15 games. Four of seven losses were low-edge or toss-up decisions, but three were materially dangerous confidence failures:

• CHW@STL: published home probability 0.591; STL lost.
• LAD@MIA: published away probability 0.600; LAD lost.
• TEX@ARI: published home probability 0.559; ARI lost.

The published confidence did not rank outcomes reliably. The four high-conviction games had only one winner hit: COL@DET was correct; LAD@MIA, CHW@STL, and TEX@ARI were misses.

Calibration quality

Calibration was poor:

• 0.50–0.55 bucket: average confidence 0.5257, observed hit rate 0.6364; the bucket was underconfident by the supplied definition.
• 0.55–0.60 bucket: average confidence 0.5829, observed hit rate 0.3333; overconfident.
• 0.60–0.65 bucket: average confidence 0.6000, observed hit rate 0.0; overconfident.
• Weighted absolute calibration error: 0.1711.

The register correctly labels this as a high model/process risk, not a critical contract failure. One slate is too small to redesign calibration confidently, but the cross-slate evidence has escalated the run-model and simulation findings; calibration instability itself has persisted for two consecutive slates and five occurrences in the 11-slate window.

Score accuracy

Score prediction was the largest performance problem:

• Historical ML team-run MAE: 2.8836
• Published team-run MAE: 3.0000
• Published total-run MAE: 4.2667
• Mean predicted margin: 1.67
• Mean actual margin: 4.33
• 12/15 predicted margins were ≤2; 8/15 actual margins were ≥4.
• Actual margins included 10, 10, and 18 runs, while the model predicted margins of 2, 2, and 1 in those games.

The validated run models explain very little holdout variance:

• Away-run R²: 0.021038
• Home-run R²: 0.011791
• Total-run R²: 0.020666

This is an established run-model weakness, not a one-game anomaly.

Starter accuracy

Starter performance was mixed and contributed directly to several misses:

• LAD@MIA: Glasnow projected 0 ER, actual 4 ER; primary starter-projection miss.
• TEX@ARI: Rocker projected 5 ER, actual 1; the overprojection favored the wrong home side.
• LAA@WAS: Ureña projected 1 ER, actual 4.
• CLE@MIN: Espino projected 3 ER, actual 0, while the opposing starter was also overprojected.
• CIN@MIL: both starters were underprojected, but the stronger-side winner survived.
• SD@SF: Perdomo was underprojected by 3 ER, but the winner was correct.

The lever-tension evidence says pitcher reweighting would help six games and risk three, but the cross-slate register does not establish a persistent separable starter bias. The correct posture is bounded investigation, not an unrestricted pitcher-weight increase.

Artifact integrity

Artifact quality is mostly good:

• Every game received structural QA.
• Every game-level structural audit reported zero critical, zero error, and zero warnings.
• Starter identities matched in all 15 comparisons.
• Forecast-file integrity was reported clean for every game.

However:

• CHW@STL and COL@DET have unexplained material calibrated-to-published probability deltas.
• The integrity-revision audit is missing.
• The player feed is empty for the slate.
• The probability lineage is therefore not fully contract-complete, even though most individual transformations are documented.

Causal issue register reconciliation

Critical: unexplained_probability_publication_delta

• Severity: Critical.
• Evidence: Two games have unexplained calibrated-to-published deltas of at least two percentage points.
• Affected games: CHW@STL, COL@DET.
• Action: Block promotion until the missing probability transformation is recorded and reconciled.
• Guardrail: Do not assume the probability governor was bypassed until pick-side math is traced.
• Conclusion: This is a contract/evidence failure. It is not proof that the published pick was numerically wrong or that the governor was bypassed.

High: score_margin_compression

• Severity: High.
• Evidence: 12/15 predicted margins ≤2; 8/15 actual margins ≥4; predicted mean margin 1.67 versus actual 4.33.
• Affected games: BAL@TOR, NYM@NYY, PHI@ATL, SEA@OAK, TEX@ARI, with wider slate-level relevance.
• Action: Evaluate recent out-of-sample tail coverage and margin MAE before widening score distributions.
• Guardrail: Do not widen every score prediction because of one blowout.
• Conclusion: Supported by both this slate and a regressing cross-slate margin trend.

High: single_slate_calibration_instability

• Severity: High.
• Evidence: WACE 0.1711 over 15 games.
• Affected games: all 15.
• Action: Monitor the same buckets over the rolling window before changing calibration weight.
• Guardrail: Do not infer a stable calibration curve from one slate.
• Conclusion: The current slate is poor, and the pattern has now persisted for two consecutive slates, but the calibration lever still needs bounded shadow testing.

High: weak_run_model_signal

• Severity: High.
• Evidence: home/away/total R² of 0.0118/0.0210/0.0207.
• Affected games: all 15.
• Action: Investigate features and target in an isolated backtest.
• Guardrail: Do not treat small positive R² as exact-score reliability.
• Conclusion: Escalated across 11 consecutive slates. This is a model/data design problem, not a nightly variance-tuning problem.

Info: complete_slate_audit_coverage

• Severity: Informational.
• Evidence: 15/15 structurally audited.
• Action: Retain the per-game audit receipt.
• Guardrail: Full coverage does not cancel internal findings.
• Conclusion: Coverage is complete, but the slate is not clean.

Info: simulation_inert_on_slate

• Severity: Informational on this slate, escalated across slates.
• Evidence: zero flip candidates and zero published flips.
• Affected games: all 15.
• Action: Measure simulation contribution over rolling slates and separate score realization from winner intervention.
• Guardrail: Zero flips is not itself a defect and does not prove no score reconciliation occurred.
• Conclusion: The cross-slate evidence now establishes a persistent diagnostic/code concern.

3. Every game

Notation: p is the published probability of the selected side; edge is the published home-minus-away confidence edge. Starter format is projected ER/ERA/outs → actual ER/ERA/outs. “Controls” includes governor, recency, and experience handling. Player actuals were unavailable on the entire slate.

BAL@TOR — winner hit, score miss

• Prediction: TOR 4–3; TOR winner.
• Actual: TOR 7–3; TOR winner.
• p / edge / bucket: home p 0.530; edge +0.060; lean.
• Errors: Brier 0.2209; team-run error 3; total bias -3; predicted margin 1, actual margin 4.
• Starters: Bradish 4.0/6.75/16 → 5.0/15.00/9; Miles 2.0/3.60/15 → 0/0/6.
• Edges: season +0.068, recent +0.330, pitcher -0.143, composite -0.015.
• Controls: governor reached 0.5871; recency guardrail applied; no starter-experience shrink; final publication reduced home p to 0.530; score volatility no-change.
• Score path: approximately 3–4 historical/baseline, published 3–4; no meaningful tail expansion.
• Player adjustment: no player actual feed; no pregame material player-adjustment list reported.
• Audit: PASS; zero critical/error/warning, 102 good, 7 info. Artifact clean.
• Cause: Correct directional call; total and TOR scoring were underprojected. Not a winner miss.

CHW@STL — winner miss and wrong-score miss

• Prediction: STL 4–3; STL winner.
• Actual: CHW 6–5; CHW winner.
• p / edge / bucket: home p 0.591; edge +0.1816; moderate. A 0.18 edge is meaningful but not certainty.
• Errors: Brier 0.349039; team-run error 4; total bias -4; predicted margin 1, actual margin 1 in absolute size but direction reversed.
• Starters: Newcomb 2.0/3.60/15 → 1.0/6.75/4; Leahy 0/0/15 → 0/0/9.
• Edges: season -0.037, recent +0.834, pitcher -0.034, composite +0.294; recent form favored STL.
• Controls: governor output 0.6339; no recency guardrail; no experience shrink; publication 0.590795. Score volatility no-change.
• Score path: historical about 3.6–4.7; baseline 4–5; enhanced/published 3–4. The model remained on STL despite recent and composite support.
• Player adjustment: 24 material changes reported, including Benintendi improvement for CHW and several STL/CHW AB, strikeout, and hit changes; no actual player outcomes available.
• Audit: PASS, zero critical/error/warning. Probability-lineage critical finding: calibrated 0.6339 to published 0.590795 has no recorded transformation.
• Primary cause: wrong conviction side.
• Secondary cause: underprojected total and CHW offense.
• Earliest wrong-direction evidence: enhanced prediction and the confidence path retained STL; no later stage recovered the side.

CIN@MIL — winner hit, score miss

• Prediction: MIL 10–6; MIL winner.
• Actual: MIL 13–9; MIL winner.
• p / edge / bucket: home p 0.540; edge +0.080; lean, despite the internal high-margin signal.
• Errors: Brier 0.2116; team-run error 6; total bias -6; predicted margin 4, actual margin 4.
• Starters: Singer 6/10.8/15 → 8/18.0/12; Drohan 2/3.6/15 → 5/12.27/11.
• Edges: season +2.038, recent +1.126, pitcher +0.057, composite +2.247.
• Controls: governor reached 0.7235; recency guardrail applied; no experience shrink; score action supported blowout margin expansion; publication fell to 0.540.
• Score path: historical about 4–5; enhanced approximately 6–10; published 6–10.
• Player adjustment: no player actual feed; no listed material player-adjustment table in this record.
• Audit: PASS; zero critical/error/warning.
• Cause: Right winner, wrong score. Both starters were worse than projected, creating an underdispersed total.

CLE@MIN — winner miss and wrong-score miss

• Prediction: CLE 6–4; CLE winner.
• Actual: MIN 4–3; MIN winner.
• p / edge / bucket: published home p 0.459, selected-away p 0.541; edge -0.0823; lean.
• Errors: Brier 0.292863; team-run error 3; total bias +3; predicted margin 2 for CLE, actual margin 1 for MIN.
• Starters: Espino 3/9.0/9 → 0/0/6; Prielipp 4/7.2/15 → 3/5.06/16.
• Edges: season -0.179, recent +0.066, pitcher +0.067, composite -0.295.
• Controls: governor 0.5578; recency guardrail and inexperienced-starter shrink applied; final home p 0.459; score volatility no-change.
• Score path: historical 5–3; baseline 5–3; enhanced 6–5; Savant reduced MIN to 4; published 6–4.
• Player adjustment: 25 material changes, including a +4.43 projected-RBI adjustment for Rocchio and other lineup changes; no actual player feed.
• Audit: PASS.
• Primary cause: mixed errors.
• Secondary cause: both starter and run-environment estimates were off.
• Earliest wrong-direction evidence: enhanced prediction moved the away side to a 6–5 advantage; the Savant stage reduced MIN further and preserved the wrong side.

COL@DET — winner hit, score miss

• Prediction: DET 7–4; DET winner.
• Actual: DET 11–7; DET winner.
• p / edge / bucket: home p 0.599; edge +0.1978; moderate.
• Errors: Brier 0.160881; team-run error 7; total bias -7; predicted margin 3, actual margin 4.
• Starters: Gordon 3/5.4/15 → 4/9.0/12; Sears 3/6.75/12 → 4/8.31/13.
• Edges: season +1.208, recent +0.814, pitcher +0.057, composite +1.534.
• Controls: governor 0.7296; no recency guardrail; starter-experience shrink applied because one starter had three career games; final p 0.5989; blowout-margin expansion supported.
• Score path: historical/baseline approximately 4–5; enhanced and final 4–7.
• Player adjustment: no actual player feed; no detailed player adjustment list reported.
• Audit: PASS. Probability-lineage critical finding: calibrated 0.7296 to published 0.5989 has no recorded transformation.
• Cause: Right winner, wrong score. Both offenses exceeded projection.

HOU@TB — winner hit, score miss

• Prediction: TB 5–4; TB winner.
• Actual: TB 3–2; TB winner.
• p / edge / bucket: home p 0.532; edge +0.0631; lean.
• Errors: Brier 0.219447; team-run error 4; total bias +4; predicted margin 1, actual margin 1.
• Starters: Lambert 1/1.69/16 → 3/4.50/18; Seymour 2/3.60/15 → 2/3.00/18.
• Edges: season +0.615, recent +0.240, pitcher -0.022, composite +0.529.
• Controls: governor 0.5437; recency guardrail applied; no experience shrink; volatility no-change.
• Score path: roughly 4–5 throughout; final 4–5.
• Player adjustment: no actual player feed.
• Audit: PASS.
• Cause: Right winner, wrong score; HOU offense and total environment were overprojected.

KC@BOS — winner hit, score miss

• Prediction: BOS 6–4; BOS winner.
• Actual: BOS 5–1; BOS winner.
• p / edge / bucket: home p 0.530; edge +0.060; lean.
• Errors: Brier 0.2209; team-run error 4; total bias +4; predicted margin 2, actual margin 4.
• Starters: Dobnak 3/4.50/18 → 3/4.76/17; Suárez 3/5.06/16 → 1/1.29/21.
• Edges: season +0.857, recent +0.251, pitcher +0.041, composite +0.984.
• Controls: governor 0.5523; recency guardrail applied; no experience shrink; volatility no-change.
• Score path: historical/baseline around 4–6; final 4–6.
• Player adjustment: no actual player feed.
• Audit: PASS.
• Cause: Right winner, wrong score; BOS pitching and KC offense were misestimated.

LAA@WAS — winner miss and wrong-score miss

• Prediction: LAA 4–3; LAA winner.
• Actual: WAS 6–5; WAS winner.
• p / edge / bucket: selected away p 0.506; edge -0.0113; toss-up.
• Errors: Brier 0.25568; team-run error 4; total bias -4; predicted margin 1, actual margin 1 opposite direction.
• Starters: Ureña 1/1.50/18 → 4/6.00/18; Alvarez 0/0/15 → 0/0/21.
• Edges: season +0.275, recent -0.022, pitcher -0.035, composite -0.111.
• Controls: governor 0.5078; recency guardrail applied; no experience shrink; volatility no-change.
• Score path: historical near 5–5; enhanced 4–3; final 4–3.
• Player adjustment: 23 material changes; no actual player feed.
• Audit: PASS.
• Primary cause: toss-up variance.
• Secondary cause: small edge on losing side and Ureña starter miss.
• Pipeline note: historical ML changed WAS to a tie at variance application, then enhanced prediction moved to LAA; that is an intermediate winner-direction regression even though later stages retained LAA.

LAD@MIA — winner miss, wrong-score miss, correct-winner/wrong-score not applicable

• Prediction: LAD 7–4; LAD winner.
• Actual: MIA 4–3; MIA winner.
• p / edge / bucket: selected away p 0.600; edge -0.200; strong.
• Errors: Brier 0.36; team-run error 4; total bias +4; predicted margin 3, actual margin 1 opposite direction.
• Starters: Glasnow 0/0/18 → 4/9.00/12; Phillips 3/5.40/15 → 2/4.91/11.
• Edges: season -1.178, recent -0.868, pitcher -0.056, composite -1.528; all supported LAD.
• Controls: governor 0.6956; recency guardrail and starter-experience shrink applied; score path used supported blowout expansion; final p 0.600.
• Score path: historical 5–3; enhanced 7–4; final 7–4.
• Player adjustment: 20 material changes, including LAD and MIA hit/RBI/strikeout adjustments; no actual feed.
• Audit: PASS.
• Primary cause: starter projection.
• Secondary cause: wrong conviction side.
• Earliest wrong-direction evidence: the historical and fundamental edges strongly supported LAD; the decisive failure was the zero-ER Glasnow projection, not a late publication flip.

NYM@NYY — winner miss and severe score miss

• Prediction: NYY 7–5; NYY winner.
• Actual: NYM 12–2; NYM winner.
• p / edge / bucket: home p 0.539; edge +0.0784; lean.
• Errors: Brier 0.290715; team-run error 12; total bias -2; predicted margin 2, actual margin 10 opposite direction.
• Starters: Thornton 0/0/16 → 2/2.57/21; Cole 3/4.50/18 → 5/10.38/13.
• Edges: season +0.882, recent -0.012, pitcher -0.014, composite +0.812.
• Controls: governor 0.5392; no recency or experience shrink; volatility no-change.
• Score path: historical 4–5; enhanced 5–6; Savant 5–7; final 5–7.
• Player adjustment: 25 material changes; no actual feed.
• Audit: PASS.
• Cause: lean-side miss, amplified by offense/pitching miss. The model overprojected NYY and underprojected NYM by 5 and 7 runs respectively.
• Earliest wrong-direction evidence: season-quality adjustment moved the score toward NYY despite recent form being nearly neutral; no later stage corrected it.

PHI@ATL — winner miss and severe score miss

• Prediction: PHI 4–3; PHI winner.
• Actual: ATL 12–2; ATL winner.
• p / edge / bucket: selected away p 0.505; edge -0.010; toss-up.
• Errors: Brier 0.255025; team-run error 11; total bias -7; predicted margin 1, actual margin 10 opposite direction.
• Starters: Mayza 1/1.80/15 → 0/0/9; Mahle 3/4.76/17 → 1/1.29/21.
• Edges: season +0.365, recent +0.078, pitcher -0.012, composite +0.013.
• Controls: governor 0.5065; recency guardrail applied; no experience shrink; volatility no-change.
• Score path: historical 4–5; enhanced 4–3; final 4–3.
• Player adjustment: 23 material changes; no actual feed.
• Primary cause: run environment under/underdispersion.
• Secondary cause: toss-up variance.
• Pipeline note: variance_applied favored ATL, but enhanced prediction flipped to PHI. This intermediate winner-direction regression is not erased by the final score path.

PIT@CHC — winner hit, modest score miss

• Prediction: CHC 5–4; CHC winner.
• Actual: CHC 4–3; CHC winner.
• p / edge / bucket: home p 0.530; edge +0.060; lean.
• Errors: Brier 0.2209; team-run error 2; total bias +2; predicted margin 1, actual margin 1.
• Starters: Skenes 3/5.06/16 → 2/2.57/21; Holmes 2/3.18/17 → 2/3.18/17.
• Edges: season +0.438, recent -0.094, pitcher -0.020, composite +0.353.
• Controls: governor 0.5547; recency guardrail; no experience shrink; volatility no-change.
• Score path: broadly 4–5; final 4–5.
• Player adjustment: no actual feed.
• Audit: PASS.
• Cause: Right winner; score was close. This is a counterexample to broad run-weight changes.

SD@SF — winner hit, score miss

• Prediction: SD 5–4; SD winner.
• Actual: SD 7–6; SD winner.
• p / edge / bucket: selected away p 0.505; edge -0.010; toss-up.
• Errors: Brier 0.245025; team-run error 4; total bias -4; predicted margin 1, actual margin 1.
• Starters: King 3/4.50/18 → 3/3.86/21; Perdomo 2/3.60/15 → 5/9.64/14.
• Edges: season -0.632, recent -0.518, pitcher -0.016, composite -0.693.
• Controls: governor 0.5078; recency guardrail and inexperienced-starter shrink applied; volatility no-change.
• Score path: historical/baseline around 5–4; final 5–4.
• Player adjustment: no actual feed.
• Audit: PASS.
• Cause: Right winner; SF starter underperformance caused the score miss.
• Pipeline note: variance stage favored SF, then enhanced prediction moved to SD. The eventual recovery does not erase that intermediate regression.

SEA@OAK — winner hit, extreme score miss

• Prediction: SEA 6–5; SEA winner.
• Actual: SEA 19–1; SEA winner.
• p / edge / bucket: selected away p 0.525; edge -0.0499; lean.
• Errors: Brier 0.22565; team-run error 17; total bias -9; predicted margin 1, actual margin 18.
• Starters: Woo 2/3.18/17 → 0/0/18; Jump 4/8.31/13 → 5/22.50/6.
• Edges: season -0.415, recent +0.586, pitcher +0.021, composite -0.215.
• Controls: governor 0.5346; recency guardrail; no experience shrink; volatility no-change despite a large Savant flag.
• Score path: approximately 4–5 through final; no tail expansion.
• Player adjustment: no actual feed.
• Audit: PASS.
• Cause: Right winner, severe score miss; run-environment and OAK starter tail were not captured.

TEX@ARI — winner miss and wrong-score miss

• Prediction: ARI 5–4; ARI winner.
• Actual: TEX 6–2; TEX winner.
• p / edge / bucket: home p 0.559; edge +0.1179; moderate. A 0.12 edge is moderate, not a near coin flip, but also not strong conviction.
• Errors: Brier 0.312417; team-run error 5; total bias +1; predicted margin 1, actual margin 4 opposite direction.
• Starters: Rocker 5/9.64/14 → 1/2.25/12; Pfaadt 4/5.68/19 → 6/8.10/20.
• Edges: season +0.368, recent +0.216, pitcher +0.071, composite +0.593.
• Controls: governor 0.5834; recency guardrail; no experience shrink; volatility no-change.
• Score path: historical 4–5; enhanced 4–5; final 4–5.
• Player adjustment: 22 material changes; no actual feed.
• Primary cause: starter projection.
• Secondary cause: wrong conviction side.
• Earliest wrong-direction evidence: the home-side composite and season edges supported ARI, but the decisive starter forecast overstated Rocker’s risk by four ER and understated Pfaadt’s by two.

4. Miss reconstruction

Earliest wrong-direction stage for every loss

Loss-type distinctions

• Bad input / artifact: No direct evidence that an input file or player identity was corrupted. Starter identities matched. However, the empty player feed is a data-availability defect that prevents postgame player attribution.
• Bad weighting: CHW@STL, CLE@MIN, LAA@WAS, NYM@NYY, and PHI@ATL show evidence of weighting or stage behavior that preserved or introduced the wrong side.
• Calibration: CHW@STL, LAD@MIA, and TEX@ARI were materially overconfident relative to the outcome. The high-conviction miss rate was 75%.
• Starter error: Most clearly LAD@MIA and TEX@ARI; materially present in LAA@WAS and CLE@MIN.
• Offense/bullpen miss: NYM@NYY, PHI@ATL, SEA@OAK, and CHW@STL had large team-scoring misses. Bullpen-specific attribution cannot be made from the supplied data.
• Run-environment miss: PHI@ATL, NYM@NYY, SEA@OAK, and CHW@STL are particularly severe. The system repeatedly predicted narrow totals and margins.
• Irreducible variance: LAA@WAS and part of CLE@MIN/PHI@ATL were genuinely low-edge or toss-up situations. That explains outcome uncertainty but not the score compression.

The 0.12 home-away confidence edge in TEX@ARI is moderate, not a near coin flip and not a strong call. It should not be treated as equivalent to the 0.20 strong LAD@MIA edge, nor as evidence that the outcome was almost random.

Pipeline-stage audit across all stages

Transition audit

• historical_ml → variance_applied: LAA@WAS moved WAS → TIE; winner accuracy fell by one game. This is a regression even though total-run MAE improved.
• variance_applied → enhanced_prediction: LAA@WAS moved TIE → LAA and PHI@ATL moved ATL → PHI; PIT@CHC and SD@SF improved. Net winner accuracy recovered by one, but PHI@ATL was a direct regression.
• enhanced_prediction → embedded_savant_run_environment: no winner changes; run MAE improved modestly.
• All later transitions: no winner-direction changes and no score changes.

The simulation did not intervene in winner direction on this slate. A later recovery in aggregate metrics does not erase LAA@WAS and PHI@ATL's intermediate winner regressions.

5. Cross-game counterfactuals

Lever tension: widen run distributions

Lever: score.variance_multiplier / score.variance_floor

• Helps: CHW@STL, LAD@MIA, TEX@ARI; the register says wider tails could lower overconfidence in these losing high-confidence calls.
• Hurts: COL@DET, where a strong correct call at 0.20 could be softened.
• Net evidence: 3 helps versus 1 hurts. The cross-slate margin-bias trend is regressing, so a small bounded shadow increase is supported. It must be evaluated on tail coverage and margin MAE, not on one slate alone.

Lever tension: pitcher form/prior weights

Lever: player.pitcher.v4 form weights / player.pitcher_prior_games

• Helps: CIN@MIL, CLE@MIN, LAA@WAS, LAD@MIA, SD@SF, TEX@ARI.
• Hurts: CHW@STL, COL@DET, PIT@CHC, where starter projections were accurate within one ER and reweighting could damage useful performance.
• Net evidence: 6 helps versus 3 hurts. This supports a bounded shadow experiment, not a large live reweight, especially because cross-slate starter bias remains statistically inseparable from noise.

Lever tension: recent/season run-mean weighting

Lever: recent.* / season.* run-mean weighting

• Helps: CHW@STL, CIN@MIL, COL@DET, HOU@TB, KC@BOS, LAA@WAS, LAD@MIA, NYM@NYY, PHI@ATL, SD@SF, SEA@OAK, TEX@ARI.
• Hurts: PIT@CHC, where the team-run estimate was already close.
• Net evidence: 12 helps versus 1 hurts. This is directionally attractive for team-run MAE and total-run MAE, but the near-zero holdout R² means changing weights alone is unlikely to solve the underlying problem.

Tempting one-game fixes that should not be made

• Do not globally widen all scores because of SEA@OAK.
• Do not remove recent form because LAD@MIA and TEX@ARI lost; recent form was informative in other games.
• Do not force more confidence toward home teams because of PHI@ATL or NYM@NYY.
• Do not increase pitcher weight indiscriminately; CHW@STL, COL@DET, and PIT@CHC had accurate starter projections.
• Do not treat TEX@ARI's 0.12 edge as a coin flip or as proof that all moderate edges should be suppressed.
• Do not infer a governor bypass from CHW@STL/COL@DET before tracing the missing publication transformation.
• Do not promote any calibration change solely from this 15-game slate.

6. Data and contract defects

Audit finding inventory

Critical

1. Unexplained probability publication delta
2. Affects CHW@STL and COL@DET.
3. Calibrated-to-published differences are material and not documented.
4. Required action: block promotion until reconciled.

Errors

• No game-level structural audit reported an error.
• The structural audit summary reported 0 errors.

Warnings

• No game-level structural audit reported a warning.
• The structural audit summary reported 0 warnings.

Informational findings

• Complete 15/15 structural audit coverage.
• Simulation produced zero winner flips.
• Each game generally reported 7–8 informational notes.
• The global structural summary reported 110 info findings.
• The player feed had no rows for the slate, preventing player-level actual attribution.
• The integrity-revision audit artifact was missing from the manifest.

Model defects versus presentation/artifact defects

Prediction-model defects

• Near-zero run-model signal: total R² 0.020666, home 0.011791, away 0.021038.
• Margin compression and poor tail coverage.
• High-conviction confidence was anti-predictive: 3/4 high-conviction calls missed.
• Simulation winner intervention was inert across 11 consecutive slates.
• Starter projections produced material errors in several losses.

Presentation or artifact defects

• Missing immutable revision-audit artifact.
• Missing player-level actual feed.
• Two undocumented probability-publication transformations.
• No source-labelled market data, preventing market cross-checks.

These artifact issues should not be conflated with the model's poor predictive performance. The available evidence says the forecast files themselves were structurally clean, while the probability lineage contract was incomplete.

Complete probability lineage audit

The intended path is:

raw model → calibrated → governor → score coherence/recency → published.

Games with documented transformations

• BAL@TOR: documented score coherence followed by matchup-recency guardrail.
• CIN@MIL: documented score coherence followed by matchup-recency guardrail.
• CLE@MIN: documented score coherence and recency controls.
• HOU@TB: documented score coherence and recency controls.
• KC@BOS: documented score coherence and recency controls.
• LAA@WAS: published probability matches calibrated/governed value apart from negligible rounding.
• LAD@MIA: documented score coherence followed by matchup-recency guardrail.
• NYM@NYY: published probability matches calibrated/governed value.
• PHI@ATL: published probability matches calibrated/governed value apart from negligible rounding.
• PIT@CHC: documented score coherence and recency controls.
• SD@SF: published probability matches calibrated/governed value apart from negligible rounding.
• SEA@OAK: documented score coherence and recency controls.
• TEX@ARI: documented score coherence and recency controls.

Bypasses or unexplained deltas

A material bypass-like gap is present only where the recorded transformation is absent:

• CHW@STL: calibrated/governor 0.6339; score coherence 0.606818; published 0.590795. The record states no transformation explains the publication delta.
• COL@DET: calibrated/governor 0.7296; score coherence 0.692975; published 0.5989. The record states no transformation explains the publication delta.

These are not yet proven governor bypasses. They are unexplained transformations requiring code-level lineage tracing.

Coverage-matrix conclusion

The coverage matrix is complete, but the slate is not “clean”:

• Structural coverage: complete.
• Probability-contract coverage: incomplete for two games.
• Player-forensics coverage: unavailable for all 15 games.
• Revision-integrity coverage: missing one artifact.

7. Cross-slate persistence

A single slate is noise; escalated findings are established patterns in this evidence.

Finding streaks

• high_raw_probability_cluster: consecutive-slate count 0; present now false. It occurred once in the 11-slate window and should not drive a persistent calibration change by itself.
• pipeline_winner_instability: consecutive-slate count 0; present now false. Incidence 3/11, but not currently streaking.
• score_margin_compression: consecutive-slate count 1; present now true; incidence 2/11. This is emerging and supported by the current slate.
• simulation_inert_on_slate: consecutive-slate count 11; present now true. Established failure pattern.
• single_slate_calibration_instability: consecutive-slate count 2; present now true; incidence 5/11. Persistent enough for a bounded shadow experiment, not a wholesale calibration redesign.
• weak_run_model_signal: consecutive-slate count 11; present now true. Established failure pattern.

Escalated findings

Simulation inertness

• Streak: 11 consecutive slates.
• Target lever: simulation.allow_score, with code/diagnostic action rather than a parameter move.
• Metric harmed: team-run MAE and total-run MAE; also removes meaningful simulation-based winner intervention.
• Direction: confirm the module is actually running and not silently no-oping.
• Evidence: zero flip candidates and zero published flips this slate; self-assessment identifies early exit and reuse of target-realization draws as the likely root cause.

Weak run-model signal

• Streak: 11 consecutive slates.
• Target levers: training.xgb.n_estimators, training.xgb.max_depth, training.xgb.learning_rate are registered, but the finding is explicitly non-parameter-actionable.
• Metric harmed: team-run MAE and total-run MAE.
• Direction: investigate feature and target design through isolated training backtest.
• Evidence: near-zero holdout R² over 2,165 rows and persistence across 11 slates.

Metric trends

Every supplied trend is regressing:

• Margin bias: regressing; mean -1.659 across the supplied series.
• Mean team-run error on losses: regressing; mean 5.7784.
• Weighted calibration error: regressing; mean 0.1573.
• Winner accuracy: regressing; mean 0.5812 across the trend window.
• Top-confidence pick loss streak: 2 consecutive slates, 20260911 and 20260912.
• Slate winner accuracy: current 0.5333, below the 0.559444 walk-forward baseline.
• Confidence status: anti-predictive; high-conviction miss rate 0.75.

The escalated simulation and run-model failures should not be dismissed as single-slate noise.

8. Adjustments for today

Recommendations are ranked by persistence, scope, expected metric benefit, and regression risk.

1. Repair and independently validate the simulation module

• Pattern: simulation inertness, 11-slate streak.
• Registered parameter/code target: simulation.allow_score; however, the evidence says this is a code/diagnostic issue, not a safe parameter change.
• Bounded change: Do not tune the parameter. Run a diagnostic shadow path with an independent unsteered winner-distribution batch of approximately 200–500 draws, without early exit, while retaining the existing target-realization search for score generation.
• Expected metric: improve meaningful simulation contribution; potentially improve winner calibration and score-tail evaluation.
• Helps: games with large unexplained tails such as SEA@OAK, NYM@NYY, PHI@ATL, and CHW@STL.
• Hurts/risk: compute cost and possible overreaction in well-behaved games such as PIT@CHC and COL@DET; use a rollback monitor on winner Brier, calibration error, and team/total-run MAE.
• Evidence strength: strongest; 11 consecutive slates and an identified code-path mechanism.

2. Run an isolated run-model feature/target backtest

• Pattern: weak run-model signal, 11-slate streak.
• Registered target: training feature contract and training process; the listed XGBoost levers are not appropriate as a direct nightly adjustment.
• Bounded change: add or test, in shadow only, source-labelled no-vig market probability where available, park/weather environment, bullpen rest/fatigue, and platoon-adjusted lineup quality; compare same-fingerprint holdout R² and MAE before promotion.
• Expected metric: increase home, away, and total-run R²; lower team-run MAE, total-run MAE, and margin MAE.
• Helps: broad slate, especially NYM@NYY, PHI@ATL, SEA@OAK, LAD@MIA, and CHW@STL.
• Hurts/risk: feature leakage, dependence on unavailable market/weather inputs, and possible degradation in already-close PIT@CHC. Require same-fingerprint backtest and rollback.
• Evidence strength: strongest model-performance evidence; 11-slate persistence and R² near zero.

3. Apply a small shadow reduction to confidence caps

• Pattern: calibration instability and current anti-predictive high-confidence cluster.
• Exact registered parameter: confidence.base_cap and confidence.moderate_cap.
• Bounded change: reduce each by approximately 0.02 in shadow only; do not change the full calibration curve yet.
• Expected metric: lower Brier score and weighted calibration error, especially in moderate and strong selections.
• Helps: CHW@STL, LAD@MIA, and TEX@ARI by reducing overcommitment; also protects against future high-raw clusters.
• Hurts/risk: may soften correct strong calls such as COL@DET and could reduce useful separation in genuinely strong matchups.
• Evidence strength: moderate; current slate is poor and calibration streak is two slates, but the register correctly warns against a one-slate overhaul.

4. Run a bounded pitcher-prior shadow experiment

• Pattern: starter errors helped six lever-tension games and hurt three.
• Exact registered lever: player.pitcher.v4 form weights or player.pitcher_prior_games.
• Bounded change: increase the prior-history requirement modestly, for example by one registered prior-game unit, or reduce the form-weight contribution by a small 5% shadow adjustment; test both separately rather than combining them.
• Expected metric: starter ER MAE and starter outs MAE.
• Helps: CIN@MIL, CLE@MIN, LAA@WAS, LAD@MIA, SD@SF, TEX@ARI.
• Hurts/risk: CHW@STL, COL@DET, PIT@CHC, where starter projections were comparatively accurate.
• Evidence strength: moderate-to-low for a live change because cross-slate starter bias is not persistent; shadow only.

5. Resolve probability lineage before promotion

• Target: no safe model parameter.
• Action: reconstruct and record the missing calibrated → score-coherence/recency → published transformations for CHW@STL and COL@DET, and produce the missing integrity-revision artifact.
• Expected metric: artifact integrity and audit completeness, not winner accuracy directly.
• Risk: none if diagnostic-only; promotion risk is unacceptable until resolved.
• Evidence strength: critical contract evidence.

9. Bottom line

Known

• The slate went 8–7, 53.33%, below the 55.94% walk-forward baseline.
• Confidence was poorly calibrated and strongly anti-predictive at the high end: three of four high-conviction picks lost.
• Scores and margins were materially compressed, especially in NYM@NYY, PHI@ATL, SEA@OAK, and several moderate-confidence misses.
• Run-model holdout signal is near zero and has persisted for 11 consecutive slates.
• Simulation winner intervention has been inert for 11 consecutive slates.
• Starter projection error materially contributed to LAD@MIA, TEX@ARI, LAA@WAS, and CLE@MIN.
• Structural game audits covered all 15 games and found no game-level critical, error, or warning findings.
• Two probability lineages are contract-incomplete, and one integrity-revision artifact is missing.

Uncertain

• Whether the two unexplained probability deltas reflect a governor bypass, an undocumented legitimate transformation, or a serialization/reporting defect.
• Whether pitcher reweighting would improve future slates without damaging accurate starter projections.
• Whether the current calibration instability warrants a larger cap adjustment; the evidence supports only a bounded shadow test.
• Player-level causal attribution for every game, because the entire player actual feed is absent.
• The causal value of simulation until an independent, non-early-exiting distribution is measured.

Single safest next action

Do not promote the current slate or tune broad model weights. First run a diagnostic code fix/shadow for the simulation distribution and independently reconcile the two missing probability transformations. This is the safest next action because it addresses the strongest persistent process failure and the critical contract gap without making an unsupported live parameter change.