Luna deep analysis — slate 20260927

Coverage receipt: COMPLETE · 9/8 sections; 14/14 games; 6/6 causes; 5/6 sources

1. Evidence receipt and gaps

Sources received

All seven source_manifest entries were checked:

The missing integrity-revision audit is material to remediation verification, although it is not evidence that the prediction model itself failed.

Games reviewed

The games array contains 14 games, all reviewed:

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

The game-level results reconcile to 9 winner hits and 5 misses. However, the evidence contains multiple coverage contradictions:

• games_reviewed: 14, hits: 9, and continuous_learning_evidence.summary reports 14 settled games.
• coverage_matrix.expected_games: 14.
• But canonical_artifacts_loaded: 15, continuous_learning_games: 15, forensic_deep_audit_games: 15, and structural_audits_completed: 15.
• complete_slate_audit_coverage states "15/14 games structurally audited".
• coverage_matrix.all_games_structurally_audited is true despite the 15-versus-14 mismatch.
• The pipeline-stage tables consistently operate on 14 games, while some slate-level summaries use 15.
• continuous_learning_evidence.summary.accuracy is 0.6429, which is 9/14, but also reports games: 15, misses: 5, and settled: 14.

This is an evidence-contract defect. It does not change the 14 game outcomes presented, but it prevents calling the slate fully clean.

Other gaps and contradictions

• Starter comparisons are reported for 14 games, but the coverage matrix reports 15 deep-audit and structural-audit games.
• Player-level forensics are unavailable for CIN@TOR, CLE@KC, COL@CHW, LAA@SEA, LAD@SF, NYM@WAS, PIT@DET, STL@MIL, and TEX@MIN because no matching settled per-player rows were found.
• Several starter lines contain implausible workload/result combinations, especially CHC@BOS, where Tanner Houck is represented as a 1.0-inning starter.
• TB@PHI has an unexplained calibrated-to-published probability transformation and no recency guardrail recorded.
• The artifact-integrity revision proof is missing.
• No source-labelled moneyline was available for market cross-checking.
• The report uses “high conviction” inconsistently with game conviction labels: the detailed games use moderate, lean, and toss_up; the aggregate spread analysis labels seven games as high-conviction.

2. Executive diagnosis

Winner accuracy

The slate produced 9 winner hits out of 14 settled games: 64.29%.

That exceeds:

• The reported walk-forward baseline accuracy of 0.570149.
• The historical model validation winner accuracy of 0.600414.
• Random guessing.

This is a positive directional result, but it is not sufficient evidence of a good overall prediction. Four of the nine winner hits had substantial score errors, and several hits were toss-ups that happened to resolve favorably.

The five misses were:

ARI@SD, CLE@KC, NYM@WAS, TB@PHI, TEX@MIN.

Calibration quality

Calibration was weak despite acceptable winner accuracy:

• Weighted absolute calibration error: 0.1031.
• 0.50–0.55 confidence bucket: average confidence 0.5114, observed hit rate 0.5714.
• 0.55–0.60 bucket: average confidence 0.5679, observed hit rate 0.7143.
• Both buckets were labeled underconfident under the supplied definition, but the overall calibration error remains material.
• High-confidence miss rate was 0.2857, or 2 of 7 games.
• Brier performance varied materially: best 0.1764, worst 0.331361.

The system’s published confidence was generally reduced by governors, score coherence, or recency controls. That does not mean the underlying probability path was always correct. TB@PHI has an unexplained publication transformation and is the critical probability-lineage issue.

A confidence edge of 0.12 is moderate, not near a coin flip. CLE@KC and COL@CHW both carried a 0.12 edge; CLE@KC lost and COL@CHW won. Those outcomes should not be treated as equivalent to the toss-up games.

Score accuracy

Score accuracy was the clearest model weakness:

• Mean predicted margin: 1.79.
• Mean actual margin: 3.36.
• Seven of 14 actual margins were at least four runs.
• Eleven of 14 predicted margins were two runs or less.
• Margin MAE: 3.0.
• Signed margin bias: -1.0, indicating under-dispersed predicted margins.
• Holdout run-model R²:
• Away runs: 0.030936
• Home runs: 0.002143
• Total runs: 0.002415

The model therefore had some directional value but almost no validated exact-run explanatory power. Correct winners with poor scores must not be counted as strong forecasts.

Starter accuracy

Starter performance was mixed and strongly contaminated by outliers and missing or implausible lines:

• The slate-level matchup analysis reports mean starter ERA error of -4.85 over six analyzed starts, but this is driven by a small number of extreme observations.
• CHC@BOS contains a structural error involving Tanner Houck’s 1.0-inning starter workload.
• CIN@TOR had large starter errors in both directions: Williamson +3 ER versus projection and Scherzer -3 ER.
• HOU@OAK had very large negative errors for both starters, but the winner was still correct.
• TB@PHI was materially affected by Zack Wheeler outperforming his projected ERA, while Nick Martinez underperformed.
• The persistent cross-slate starter ERA sign has not passed the persistence test: 7 of 12 slates shared the sign, with mean 0.713, standard error 0.47, and z=1.51.

The pitcher-form lever therefore should not be changed from this slate alone. The lever tension was 4 games helped versus 5 hurt.

Artifact integrity

Artifact integrity is not clean:

• Structural audit summary: 0 critical, 2 errors, 1 warning, 104 informational findings.
• The two errors are both on CHC@BOS and concern the implausible Tanner Houck workload/result line.
• The warning is on NYM@WAS: Sean Manaea pitch-count batters faced differ between summary and box score, 26 versus 24.
• The critical causal issue is not a structural game artifact but an unexplained probability publication delta on TB@PHI.
• The missing integrity-revision audit further limits post-remediation verification.

Causal issue register reconciliation

1. unexplained_probability_publication_delta — Critical
2. Affected game: TB@PHI.
3. Evidence: One game has an unexplained calibrated-to-published delta of at least two percentage points.
4. Action: Block promotion until the missing probability transformation is recorded and reconciled.
5. Guardrail: Do not assume the probability governor was bypassed until the pick-side math is traced.
6. Diagnosis: Supported. TB@PHI records calibrated probability 0.6056, governor output 0.6056, score-coherence probability 0.588988, and published pick probability 0.57564, with "no recorded transformation explains the publication delta". This is a contract failure, not proof that the governor was bypassed.

1. artifact_integrity_failure — High
2. Affected game: CHC@BOS.
3. Evidence: Two structural errors and one warning slate-wide.
4. Action: Exclude the slate from parameter learning until the artifact-generation step is fixed.
5. Guardrail: Do not blame accuracy changes on a malformed artifact.
6. Diagnosis: Supported. The CHC@BOS file defect is separate from the correct winner call.

1. score_margin_compression — High
2. Affected games: ARI@SD, CHC@BOS, CIN@TOR, HOU@OAK, TB@PHI.
3. Evidence: 11/14 predicted margins ≤2, 7/14 actual margins ≥4, predicted mean margin 1.79 versus actual 3.36.
4. Action: Check multiple recent out-of-sample slates before widening score distributions.
5. Guardrail: Do not widen every score prediction because of one blowout.
6. Diagnosis: Supported and now persistent for two consecutive slates. Cross-game evidence favors a bounded shadow increase in variance, not an unrestricted widening.

1. weak_run_model_signal — High
2. Affected: All 14 games.
3. Evidence: Near-zero holdout R².
4. Action: Investigate features and target through an isolated backtest.
5. Guardrail: Do not treat small positive R² as exact-score reliability.
6. Diagnosis: Strongly supported and escalated after 11 consecutive slates. This is a code/data/training problem, not a nightly tuning problem.

1. complete_slate_audit_coverage — Info
2. Evidence: The slate says every settled game received QA, but counts are internally inconsistent at 15 versus 14.
3. Action: Retain per-game audit receipts.
4. Diagnosis: The existence of per-game receipts is supported, but the coverage contract itself is not reconciled.

1. simulation_inert_on_slate — Info in the slate register, escalated across slates
2. Evidence: Zero flip candidates and zero published flips.
3. Action: Measure simulation contribution over rolling slates.
4. Guardrail: Zero flips alone is not proof that the simulator did no score work.
5. Diagnosis: As a single slate, this is only diagnostic. Across 11 consecutive slates, it is an established code/diagnostic failure requiring operator action.

3. Every game

The season, recent, pitcher, and composite edges below are home-minus-away edges. “Controls” includes governor, recency, starter-experience, or score-volatility handling. Player-level adjustments are reported only where supplied; absent data are not inferred.

Game-by-game table

Controls, score paths, and player adjustments

• ARI@SD: Governor and matchup-recency guardrail reduced the home probability to approximately 0.513 before publication; published home probability was 0.489. Score path moved from historical ARI 4.37–SD 3.56 to final ARI 6–SD 4, including a Savant/environment lift. Player adjustments included Machado (-2.7 predicted AB), Nootbaar (+2.686 RBI), Moreno (+2.362 RBI), and a 24-player material-change count. The audit passed with zero defects. The actual result was classified as variance rather than a material starter or bullpen break.
• ATL@MIA: Governor and recency guardrail applied; final probability remained 0.505. Score path ended at ATL 4–MIA 5. Player-level evidence identified Olson underperformance and Dubón overperformance. Audit passed.
• CHC@BOS: Governor and recency guardrail applied; published away probability was approximately 0.505. Score path was CHC 4–BOS 3; winner was right but score wrong. Bregman’s +3.6 run impact was the largest supplied player deviation. The artifact contained two errors, both tied to Houck’s implausible 1.0-inning starter line.
• CIN@TOR: Starter-experience shrink factor 0.9458 was applied; winner unchanged. Score path remained CIN 4–TOR 5. No player-level actuals were joined. Audit passed.
• CLE@KC: Governor, matchup-recency controls, and blowout-margin expansion applied. The score path expanded to CLE 6–KC 4, despite the actual low-scoring CLE 2–KC 3. Player adjustments had 23 material changes, including Tolbert and Halpin. Audit passed. This is a correct moderate-edge interpretation of the supplied pregame evidence but a wrong result.
• COL@CHW: Governor and matchup-recency guardrail applied; score path expanded to COL 5–CHW 8. Actual score was only 2–4, making this a correct-winner/wrong-score game. No player-level actuals were joined. Audit passed.
• HOU@OAK: Governor and matchup-recency guardrail applied; score path was HOU 8–OAK 6, actual 9–0. Walker was +2.06 runs versus projection. Audit passed. The actual starter outcomes were extreme in both directions.
• LAA@SEA: Governor and recency guardrail applied; score path LAA 3–SEA 5, actual 3–7. Audit passed. No player-level actuals were joined.
• LAD@SF: Governor and recency guardrail applied; score path LAD 6–SF 3, actual 5–1. The largest supplied starter deviation was Seymour at -4 ER. Audit passed.
• NYM@WAS: Governor, recency guardrail, and starter-experience shrink factor 0.4042 applied; the inexperienced-starter control changed probability but not the winner. Score path ended NYM 7–WAS 5. There was one warning for Manaea’s pitch-count mismatch. No player-level actuals were joined.
• PIT@DET: Governor, recency guardrail, and starter-experience control applied; score path PIT 4–DET 3, actual 4–2. Audit passed.
• STL@MIL: Governor and recency guardrail applied; score path STL 2–MIL 5, actual 4–6. Audit passed. No player-level actuals were joined.
• TB@PHI: Governor was applied, but recency guardrail was recorded as false. Score path changed TB 4–PHI 5 to TB 4–PHI 3 and then published TB 4–PHI 3; actual was 3–7. Player adjustments included Stott +1.89, Sosa +1.71, and Wheeler strikeout changes. Audit passed structurally, but probability lineage failed the critical contract.
• TEX@MIN: Governor and recency guardrail applied; score path TEX 5–MIN 4, actual 4–6. No player-level actuals were joined. Audit passed.

4. Miss reconstruction

ARI@SD

• Earliest wrong-direction movement: The historical model initially favored ARI, and variance application produced a tie. Enhanced prediction moved back toward ARI, and later score/environment stages increased ARI’s projected score to 6 versus 4.
• Bad input: Partial starter evidence: the detailed player feed lacked a settled pitching line for Michael Soroka, although the game-level starter line supplied a projection and actual.
• Bad weighting: Composite home edge was -0.258, while season and recent edges favored SD. The evidence was internally mixed; no clear weighting defect is proven.
• Calibration: Appropriate toss-up treatment was not the principal problem; published home probability was only 0.489.
• Starter error: Not material by ER: both starters had ER error 0.
• Offense/bullpen: SD’s offense was under-projected by 5 runs; the detailed forensics classify the result as closer to variance.
• Run environment: Total was under-projected by 3 and team-run absolute error was 7.
• Irreducible variance: Strong contributor. This was a low-confidence direction that flipped.

CLE@KC

• Earliest wrong-direction movement: Historical and variance stages favored CLE; enhanced prediction expanded CLE to 6–4. The score-volatility module then explicitly applied blowout-margin expansion.
• Bad input: No direct artifact defect. Player-level actuals were unavailable.
• Bad weighting: Season, recent, and composite edges all strongly favored CLE, so the direction was not an arbitrary small-edge mistake.
• Calibration: A published 0.12 edge is moderate, not near a coin flip. It was a meaningful commitment and lost.
• Starter error: Messick performed better than projected by 2 ER; Lynch was 1 ER worse. This modestly favored CLE and does not explain the miss.
• Offense/bullpen: CLE offense was over-projected by 4.
• Run environment: Primary cause. Ten projected total runs became five actual runs; total bias was +5.
• Irreducible variance: Some, but the model also overcommitted to a wide high-scoring outcome.

NYM@WAS

• Earliest wrong-direction movement: Historical ML favored NYM, and later stages increased NYM to 7–5. The starter-experience adjustment shrank confidence but did not change the winner.
• Bad input: No player-level actuals were joined. One pitch-count warning exists for Manaea.
• Bad weighting: Composite edge was only -0.193, despite season edge favoring NYM and recent edge favoring WAS. The lean-side call was not strongly supported after reconciliation.
• Calibration: Published away probability 0.531 and edge 0.0611 were appropriately lean rather than moderate.
• Starter error: Manaea was 1 ER worse than projected; Herz’s ER matched projection, though his ERA was worse because of workload.
• Offense/bullpen: NYM offense was over-projected by 3.
• Run environment: Total was over-projected by 2.
• Irreducible variance: Present, but the principal model issue was a small edge on the wrong side.

TB@PHI

• Earliest wrong-direction movement: Historical ML favored PHI, but the published pathway committed to TB. The probability-lineage evidence is incomplete, so the precise earliest probability-stage regression cannot be certified.
• Bad input: The probability publication record is incomplete; recency guardrail is recorded false despite the general matchup context. Starter feed is missing Nick Martinez’s settled player line.
• Bad weighting: Season and recent edges favored PHI, while the system backed TB. The composite home edge was -0.604, not a marginal disagreement.
• Calibration: Published away probability was 0.57564, a moderate 0.1513 edge. This was not a coin flip and was materially overconfident relative to the loss.
• Starter error: Martinez was 1 ER worse than projected; Wheeler was 2 ER better. Wheeler’s performance materially helped PHI.
• Offense/bullpen: PHI offense was under-projected by 4. The player audit identified Stott and Sosa as positive deviations. Bullpen ER was not the primary issue in supplied data.
• Run environment: Total was under-projected by 3; team-run error was 5.
• Irreducible variance: Wheeler’s outlier performance contributed, but the critical publication-lineage defect must be resolved before causality is finalized.

TEX@MIN

• Earliest wrong-direction movement: Historical and enhanced stages favored TEX. The recency and matchup controls did not reverse the side.
• Bad input: No player-level actuals were joined.
• Bad weighting: Recent edge favored MIN by +0.522, while season edge was nearly neutral and composite favored neither side strongly. The final TEX lean was not robust.
• Calibration: Published away probability 0.51197, edge 0.0239, correctly treated as toss-up.
• Starter error: Gore was 3 ER better than projected; Kremer was 1 ER better. The good pitching suppressed scoring and altered the close game.
• Offense/bullpen: No decisive lineup failure was documented.
• Run environment: Total bias was -1 and team-run error 3.
• Irreducible variance: Primary contributor.

Pipeline-stage quality across all stages

A later correct winner does not erase an earlier winner-direction regression.

Important transitions:

• Historical ML → variance applied improved margin MAE by 0.0948 and team-run MAE by 0.0215; it changed ARI and ATL to ties.
• Variance applied → enhanced prediction improved winner accuracy by one hit, but worsened margin MAE by 0.2143 and team-run MAE by 0.1785. It helped ATL@MIA and CHC@BOS but regressed TB@PHI. It also changed several winner directions, including CHC@BOS and TB@PHI.
• Enhanced → embedded Savant changed TB from a tie to a TB lead but did not improve winner accuracy and worsened team-run MAE.
• Embedded Savant → fundamental review again changed TB toward TB without improving winner accuracy.
• Fundamental review, capped target, simulation result, and published final were identical in winner and score metrics.

The pipeline improved winner accuracy after enhanced prediction, but the improvement came with worse score quality and a documented regression on TB@PHI. That intermediate regression remains a model-process finding.

5. Cross-game counterfactuals

Lever tension 1: widen run distributions

Lever: score.variance_multiplier / score.variance_floor

• Helps: CLE@KC, where a wider distribution should reduce overconfidence in the six-to-four projection and better represent a low-scoring reversal; TB@PHI, where the narrow score range failed to represent PHI’s possible offensive tail.
• Hurts: No games were listed as hurt in the supplied lever tension.
• Net evidence: Supports a bounded increase. The evidence is strengthened by two consecutive slates of margin compression, but the change should be shadowed because a single slate cannot prove the optimal magnitude.
• Guardrail: Do not widen all predictions aggressively based on HOU@OAK or one blowout.

Lever tension 2: pitcher form weights / prior-games weighting

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

• Helps: CIN@TOR, HOU@OAK, LAD@SF, and TEX@MIN according to starter-error attribution.
• Hurts: ARI@SD, CHC@BOS, LAA@SEA, NYM@WAS, and PIT@DET, where current starter projections were comparatively accurate or where reweighting could disturb useful results.
• Net evidence: 4 helped versus 5 hurt. Do not change this lever from the slate. Cross-slate persistence also does not support it.

Lever tension 3: recent/season run-mean weighting

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

• Helps: ARI@SD, CLE@KC, COL@CHW, HOU@OAK, NYM@WAS, TB@PHI.
• Hurts: ATL@MIA, LAA@SEA, PIT@DET.
• Net evidence: 6 helped versus 3 hurt, but the benefit is based primarily on large score errors and near-zero validated run-model signal. This supports diagnostic investigation, not immediate aggressive reweighting.

Tempting one-game fixes that must not be made

• Do not reverse the CLE@KC side solely because the 0.12 edge lost. The same magnitude of edge correctly selected COL@CHW, LAA@SEA, and HOU@OAK.
• Do not widen every score because HOU@OAK ended 9–0. The register expressly warns against learning from one blowout.
• Do not alter pitcher weights because of CIN@TOR or HOU@OAK. The cross-game count is 4 helped versus 5 hurt.
• Do not disable season/recent form because it failed in TB@PHI; the same directional structure helped multiple other games.
• Do not blame CHC@BOS’s correct winner or its model accuracy on the malformed artifact.
• Do not infer that the simulator was bypassed solely from zero winner flips; the persistent code-level evidence requires a diagnostic audit first.
• Do not lower all confidence caps based on the current slate’s moderate bucket results alone; persistent high-raw clustering is not present on this slate.

6. Data and contract defects

Audit findings by severity

Critical

• TB@PHI: Published probability differs materially from the calibrated/governed value without a recorded transformation. The lineage records:
• Raw model pick probability: 0.60695
• Calibrated probability: 0.6056
• Governor output: 0.6056
• Score-coherence probability: 0.588988
• Published pick probability: 0.57564
• Recency guardrail: false
• Transformation: no recorded explanation

This is a probability-contract failure. Promotion should remain blocked until reconciled.

Errors

• CHC@BOS: Structural audit error: BOS/Tanner Houck has an implausible 1.0-inning starter outing in both current and simulation artifacts.
• CHC@BOS: Same pitching-realism issue appears twice, once for current and once for simulation output.

These are artifact/pipeline representation errors, not proof of a wrong model winner.

Warning

• NYM@WAS: Sean Manaea pitch-count row differs from the pitching box score: batters faced 26 in the summary versus 24 in the box score.

Informational and coverage defects

• Structural audit summary reports 104 informational findings.
• Nine games have no matched player-level actual feed.
• Coverage reports 15 structural audits for an expected 14-game slate.
• The integrity-revision audit is missing.
• The simulation produced zero flip candidates and zero published flips.
• No market cross-check was available.

Prediction-model defects versus artifact bugs

Prediction-model/process defects

• Near-zero validated run-model signal: total R² 0.002415.
• Persistent margin compression.
• Persistent simulation inertness.
• Inconsistent or unexplained probability publication for TB@PHI.
• Under-dispersed score paths and poor hitter-level separation:
• H projection correlation 0.054, spread ratio 0.323.
• AB correlation -0.164.
• R correlation -0.070.
• HR correlation -0.142.
• Several metrics are beaten by publishing the slate mean for every hitter.

Presentation/artifact defects

• CHC@BOS implausible starter workload representation.
• NYM@WAS pitch-count mismatch.
• Missing integrity-revision audit.
• 15-versus-14 coverage and settled-game count mismatch.

These artifact defects must not be used to infer corresponding accuracy failures.

Complete probability-lineage audit

The pick-side lineage was checked for all 14 games:

• ARI@SD: raw 0.5176 → calibrated 0.5127 → governor 0.5127 → score coherence/recency → published 0.510815; transformation recorded, no material unexplained delta.
• ATL@MIA: raw 0.5071 → calibrated/governed/published 0.505; recorded.
• CHC@BOS: raw 0.5000/governor input 0.50075 → calibrated/governed 0.5071 → published 0.505149; recorded.
• CIN@TOR: raw 0.5332 → calibrated 0.5212 → score/recency → published 0.511591; recorded.
• CLE@KC: raw 0.6363 → calibrated 0.6785 → governor output 0.6785 → published 0.560; transformation recorded as score coherence/recency.
• COL@CHW: raw 0.6596 → calibrated 0.6745 → governor output 0.6745 → published 0.560; recorded.
• HOU@OAK: raw 0.6674 → calibrated 0.6476 → governor output 0.6476 → published 0.560; recorded.
• LAA@SEA: raw 0.6485 → calibrated 0.5973 → published 0.560; recorded.
• LAD@SF: raw 0.7304 → calibrated 0.6405 → published 0.580; recorded.
• NYM@WAS: raw 0.6402 → calibrated 0.6055 → published 0.530557; recorded.
• PIT@DET: raw 0.5182 → calibrated 0.5131 → published 0.505; recorded.
• STL@MIL: raw 0.7139 → calibrated 0.6314 → published 0.580; recorded.
• TB@PHI: raw 0.60695 → calibrated 0.6056 → governor 0.6056 → score coherence 0.588988 → published 0.57564; material transformation not recorded.
• TEX@MIN: raw 0.5247 → calibrated 0.5166 → published 0.51197; recorded.

The only declared material bypass/unexplained delta is TB@PHI. The other large raw-to-published changes have recorded transformations and should not be labeled bypasses.

Coverage reconciliation

The slate cannot be called clean because:

1. Expected games: 14.
2. Game records: 14.
3. Structural audit count: 15.
4. Canonical artifacts loaded: 15.
5. Continuous-learning games: 15.
6. Missing integrity-revision audit.
7. Settled count is 14 in one place and games count is 15 in another.

The per-game result set is complete for the 14 listed games, but the audit-count contract requires correction.

7. Cross-slate persistence

Raw structural evidence

Margin distribution

Current slate evidence shows:

• Actual margins ≥4: 7 of 14.
• Predicted margins ≤2: 11 of 14.
• Mean predicted margin: 1.79.
• Mean actual margin: 3.36.
• Current margin-bias series endpoint: -1.571.
• Trend direction: improving, but still materially negative.

The improvement does not remove the two-slate persistence of compression.

Pipeline-stage transitions

Across the current slate:

• Variance application improved run errors slightly.
• Enhanced prediction improved winner accuracy but worsened team-run and margin MAE and regressed TB@PHI.
• Savant and fundamental review did not improve winner accuracy after enhanced prediction.
• Capped target, simulation result, and published final were inert relative to the preceding stages.

Calibration monitoring

• Weighted absolute calibration error: 0.1031.
• Current metric-trend direction: improving.
• The current slate is better than several recent slates, but the level remains materially nonzero.
• High-raw cluster: 4 games, 4 hits, but the cluster is not persistent enough to support a cap change now.

High-raw cluster

Four games exceeded the raw cluster threshold and all hit:

COL@CHW, HOU@OAK, LAD@SF, STL@MIL.

The cluster is useful diagnostic evidence but not a persistent failure on this slate. It should not trigger a one-slate confidence-cap adjustment.

Simulation control

• Observed games: 15 in the simulation-control block.
• Flip candidates: 0.
• Published flips: 0.
• Status: diagnostic-only without independent support.
• The cross-slate escalation confirms this has persisted 11 consecutive slates.

Finding streaks

Every finding_streaks entry:

• high_raw_probability_cluster: consecutive-slate count 0; incidence 2/11; not present currently as a persistent finding.
• pipeline_winner_instability: consecutive-slate count 0; incidence 6/11; not present currently.
• score_margin_compression: consecutive-slate count 2; present currently.
• simulation_inert_on_slate: consecutive-slate count 11; present currently.
• single_slate_calibration_instability: consecutive-slate count 0; incidence 6/11; not present currently.
• weak_run_model_signal: consecutive-slate count 11; present currently.

Escalated findings

These are established failures, not single-slate noise.

1. Simulation inertness
2. Streak: 11 consecutive slates; incidence 11.
3. Target lever: simulation.allow_score, although the register marks it non-parameter-actionable.
4. Metric harmed: team_run_mae, total_run_mae.
5. Direction of fix: Confirm the simulator is actually running and independently estimate the winner distribution. The supplied self-assessment identifies a shared-draw, early-exit problem: the simulation uses the same target-realization draws for score construction, may stop after roughly two selected draws, and therefore cannot produce a meaningful win rate.

1. Weak run-model signal
2. Streak: 11 consecutive slates; incidence 11.
3. Target levers: training.xgb.n_estimators, training.xgb.max_depth, training.xgb.learning_rate, but the register marks the finding non-parameter-actionable.
4. Metric harmed: team_run_mae, total_run_mae.
5. Direction of fix: Isolated feature/target backtest. Add or validate market no-vig probability where source-labelled, park/weather environment, bullpen rest/fatigue, and platoon-adjusted lineup quality. Bump model schema when semantics change.

Top-confidence-pick loss streak

• top_confidence_pick.loss_streak: 0.
• The 11-slate sequence was: hit, hit, miss, hit, miss, miss, hit, hit, hit, miss, hit.
• The current slate’s top confidence pick was a hit.
• Therefore there is no current consecutive top-pick loss streak.

Metric-trend directions

All supplied metric trends are explicitly improving:

• Margin bias: improving.
• Mean team-run error on losses: improving.
• Weighted calibration error: improving.
• Winner accuracy: improving.

“Improving” does not mean solved. In particular, the run model remains near-zero-signal and margin compression remains present.

8. Adjustments for today

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

1. Simulation inertness — highest evidence strength

• Action type: Code/diagnostic fix, not a parameter promotion.
• Registered parameter: simulation.allow_score is the registered lever, but the finding explicitly says it is not parameter-actionable.
• Bounded change: Do not change the parameter blindly. Run a shadow implementation with an independent unsteered winner-distribution batch of approximately 200–500 draws, or a closed-form Poisson/Skellam difference using expected runs and variance. Remove the early-exit dependency from the probability batch.
• Expected metric: Improve the diagnostic validity of winner probability; potentially improve team_run_mae and total_run_mae.
• Helps: All games, especially games where the simulation currently produces no independent winner intervention.
• Hurts/risk: Additional compute and possible score/winner disagreements; rollback if independent simulation worsens rolling Brier or winner accuracy.
• Why first: Persistent 11-slate code-level failure and direct self-assessment evidence.

2. Weak run-model signal — persistent but operator-only

• Action type: Isolated training/backtest code and feature-contract work.
• Registered parameters: training.xgb.n_estimators, training.xgb.max_depth, and training.xgb.learning_rate are listed, but no bounded nightly parameter move is recommended because the finding is explicitly non-parameter-actionable.
• Bounded change: Do not tune these parameters in production today. Add the proposed feature-contract candidates in an isolated experiment, retrain with the same temporal fingerprint, and compare holdout R² before and after.
• Expected metric: Increase away/home/total run R² from approximately 0.031/0.002/0.002; reduce team_run_mae and total_run_mae.
• Helps: Broadly the full slate, particularly score-error games ARI@SD, CLE@KC, COL@CHW, HOU@OAK, NYM@WAS, and TB@PHI.
• Hurts/risk: Feature leakage, unstable schema changes, or deterioration in exact-run and winner metrics. Require schema-version increment and rollback if holdout performance does not improve.
• Why second: Persistent 11-slate failure and near-zero holdout signal.

3. Margin compression — strongest current single-slate driver with persistence

• Registered parameter: score.variance_multiplier.
• Direction: Increase modestly.
• Magnitude: Shadow increase of approximately 5% relative to the current value, with no simultaneous change to score.variance_floor.
• Expected metric: Reduce margin compression, improve tail coverage, and reduce margin MAE; potentially improve score calibration. The effect on winner Brier is uncertain.
• Helps: CLE@KC and TB@PHI are explicitly identified by lever tension; also protects against under-dispersed outcomes in ARI@SD, CIN@TOR, HOU@OAK, and STL@MIL.
• Hurts/risk: Could worsen already close-score games such as ATL@MIA, LAA@SEA, and PIT@DET, and can increase total-run MAE if applied indiscriminately.
• Gate: Keep as a shadow experiment with rollback if rolling total-run MAE increases materially or low-scoring-game calibration degrades.
• Why third: Present for two consecutive slates, directly supported by current margin distribution, and has a registered actionable lever.

4. Confidence caps — do not change today

• Registered levers: confidence.base_cap, confidence.moderate_cap, and confidence.empirical_max_weight.
• Reason for abstention: The current high-raw cluster went 4-for-4, the top-confidence pick hit, and the persistent high-raw finding is not present on this slate. Although the self-assessment proposes lower caps, the cross-game and cross-slate evidence does not isolate a current cap failure.
• Potential benefit: Lower Brier and calibration error if overconfidence persists.
• Potential harm: Unnecessarily suppressing useful moderate edges such as COL@CHW, LAA@SEA, LAD@SF, and STL@MIL.
• Status: Monitor, do not promote a change today.

5. Pitcher weighting — abstain

• Registered lever: player.pitcher.v4 form weights / player.pitcher_prior_games.
• Reason: 4 games helped and 5 hurt; cross-slate starter ERA bias is not separable from noise.
• Affected risk: A change could disturb accurate starter projections in ARI@SD, CHC@BOS, LAA@SEA, NYM@WAS, and PIT@DET.
• Action: No parameter move. Continue monitoring starter ER MAE and outs MAE.

9. Bottom line

Known

• The system correctly selected 9 of 14 winners, outperforming the supplied baseline.
• Winner accuracy was materially better than score accuracy.
• Score margins were compressed: predicted margins were too narrow and run-model holdout R² was near zero.
• Two failures are persistent across 11 slates: inert simulation behavior and weak run-model signal.
• Margin compression is present for two consecutive slates and supports a bounded variance shadow experiment.
• TB@PHI has a critical probability-lineage defect requiring reconciliation.
• CHC@BOS has artifact errors, and NYM@WAS has a pitch-count warning.
• The slate is not artifact-clean because of the 15-versus-14 coverage contradiction and the missing integrity-revision audit.

Uncertain

• Whether the TB@PHI publication delta reflects a governor bypass, an undocumented score-coherence transformation, or a reporting omission. The supplied guardrail correctly prohibits assuming bypass before tracing the math.
• Whether the observed starter ERA sign reflects a real bias or slate-to-slate noise.
• Whether recent/season run-mean weighting is genuinely beneficial, because the model’s underlying run features have almost no validated signal.
• The exact causal contribution of the simulation to score or winner probabilities, because its current batch is not independently informative.
• Player-level causes for the nine games without matched player actuals.

Single safest next action

Block promotion and run the operator-level probability-lineage and simulation contract audit first: reconcile TB@PHI’s raw-to-published probability transformation and replace the inert shared-draw simulation probability with an independent, non-early-exiting winner-distribution shadow. This improves artifact trust and diagnostic validity without making an unsupported model-parameter change.