Luna deep analysis — slate 20260830

Coverage receipt: COMPLETE · 9/8 sections; 14/14 games; 7/7 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_20260830.json, 349,718 bytes; official outcomes and scoring errors.
2. Forensic deep audit — /app/output/reports/PREDICTION_SYSTEM_DEEP_AUDIT_20260830.json, 226,049 bytes; probability, provenance, starter, and decision contracts.
3. Continuous learning — /app/output/learning/continuous_learning_daily_20260830.json, 2,264,904 bytes; pipeline-stage quality, calibration, lineage, score distribution, and starter comparisons.
4. Detailed report — /app/output/reports/detailed_20260830.html, 331,974 bytes.
5. Ultra-detailed player audit — /app/output/reports/20260830_ultra_detailed_player_audit.html, 2,061,684 bytes.
6. Structural audit — /app/output/reports/audit_report_20260830.html, 138,926 bytes.

The following manifest item is explicitly missing:

1. Integrity revision audit — /app/output/reports/audit_report_20260830_integrity_revision.html; intended to provide post-remediation proof for immutable revision artifacts.

Slate coverage

All 14 expected games were reviewed:

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

Coverage is structurally complete: 14/14 canonical artifacts loaded, 14/14 continuous-learning games, 14/14 forensic deep audits, 14/14 starter comparisons, and 14/14 structural audits.

Gaps and contradictions

• The integrity revision audit is missing. Therefore, immutable revision-artifact remediation cannot be independently confirmed.
• The structural audit reports 0 critical, 0 error, 1 warning, and 100 info findings, but the causal register also records a high-severity “canonical artifact warning requiring reconciliation.” The named underlying presentation/stat contract is not supplied in the evidence, so its exact defect cannot be identified.
• CHW@MIN has an unexplained calibrated-to-published probability delta of at least 2 percentage points. The evidence explicitly warns not to assume a governor bypass without tracing pick-side semantics.
• MIA@WAS also says “no recorded transformation explains the publication delta,” but the causal register identifies only CHW@MIN as the material unexplained probability-publication failure. This is an internal inconsistency in the evidence packaging: MIA’s delta is approximately 1.64 points, below the stated ≥2-point causal threshold, while its transformation field is still incomplete.
• CIN@CHC has a pitch-count warning: Chase Burns’ simulation summary says H=3 and BF=21, while the pitching box score says H=2 and BF=20.
• The data marks the slate learning-eligible because there were no critical/error structural gates, but that does not mean the probability lineage, presentation contract, or intermediate pipeline transitions are clean.

2. Executive diagnosis

Winner accuracy

• Published winner accuracy: 7/14, 50.0%.
• The continuous-learning summary independently reports 7 hits and 7 misses.
• Historical walk-forward winner accuracy: 56.6962%.
• The slate therefore underperformed the historical reference by approximately 6.7 percentage points.
• Brier score: 0.236911 for the model-validation reference and 0.24228 for the walk-forward reference; slate-level per-game Brier values average approximately 0.253. The slate was not merely a winner miss problem: several high-confidence misses materially worsened probabilistic quality.

Correct winner does not imply a good prediction:

• KC@CLE was a correct winner call but had a 9-run team error and a 7-ER starter error.
• LAD@DET was a correct winner call but missed the score by 4 total team runs.
• COL@ATL was correct but had a 4-run team error.
• Conversely, PHI@LAA was both a correct winner and a highly accurate score call.

Calibration quality

Calibration was poor in every reported confidence bucket:

Overall weighted absolute calibration error was 0.0693 across 14 settled games. All three buckets were overconfident.

The high-raw cluster was especially concerning:

• Three games had raw pick probability at least 0.65.
• Only 1/3 won.
• The losses were CHW@MIN and CIN@CHC.
• PHI@LAA was the successful high-raw pick.

The evidence does not support a new bucket-specific shrinkage formula from one slate, but it does support retaining and auditing raw → calibrated → governed → published lineage.

Score accuracy

Score quality was weak:

• Home-run R²: 0.042885
• Away-run R²: 0.025261
• Total-run R²: 0.022432
• These models explain little holdout variance beyond a mean-like baseline.
• Mean team-run error on losses: 7.0 runs.
• Losses had mean total-run bias −0.714, but this aggregate hides large opposing errors: BOS@NYY was 8 runs under, CHW@MIN was 5 runs over, and SEA@TOR was 2 runs over.

Margin underprediction was persistent on this slate:

• Six actual margins were at least four runs.
• Ten predicted margins were two runs or less.
• Mean signed margin bias was −2.0003 in the cross-slate trend series, marked regressing.

The score pipeline’s best margin MAE was the historical baseline at 4.4881. Enhanced prediction worsened it to 4.7143, with team-run MAE worsening from 2.4184 to 2.5000, although total-run MAE improved from 2.1848 to 1.8571.

Starter accuracy

Starter projections were a material source of score error:

• BOS@NYY: Ranger Suarez +3 ER error; Will Warren −4.
• CHW@MIN: Zebby Matthews −3.
• CIN@CHC: Shota Imanaga +3; Chase Burns −2.
• KC@CLE: Seth Lugo +7, although the winner was still correct.
• SEA@TOR: Logan Gilbert +2; Max Scherzer −4.
• PHI@LAA, BAL@OAK, and others had mixed but sometimes directionally useful starter errors.

Starter identity matched in every listed comparison. The evidence therefore points to projection error rather than an identity-matching failure.

Artifact integrity

• Structural coverage was complete.
• Structural audit totals: 0 critical, 0 error, 1 warning, 100 info.
• The one explicit warning was the Chase Burns pitch-count discrepancy in CIN@CHC.
• A separate causal register marks an artifact-integrity warning as high severity, requiring reconciliation of the named presentation/stat contract.
• The missing integrity revision audit is a documentation and proof gap.
• The CHW@MIN probability-lineage issue is critical, because it is a contract/provenance failure, not merely a modeling error.

Causal issue register reconciliation

Critical: unexplained probability-publication delta

• Severity: critical.
• Affected game: CHW@MIN.
• Evidence: one game with an unexplained calibrated-to-published delta ≥2 percentage points.
• Facts: calibrated/governed probability 0.7117; published pick probability 0.686674; published-minus-calibrated −0.025026.
• Action: block promotion until the missing transformation is recorded and reconciled.
• Guardrail: do not assume governor bypass without tracing pick-side semantics.

High: canonical artifact warning

• Severity: high.
• Affected games: none assigned by the register.
• Evidence: 0 critical, 0 error, 1 warning, 100 info.
• Action: reconcile the named presentation/stat contract before treating the artifact as clean evidence.
• Guardrail: a warning is not proof that model direction was wrong.

High: pipeline winner instability

• Severity: high.
• Affected games: BAL@OAK and PIT@STL are explicitly identified as regressed games at the historical_ml → variance_applied boundary.
• Evidence: winner accuracy fell from 7/14 to 5/14, a −0.1429 transition delta; later enhanced prediction recovered both.
• Action: identify exact boundary flips and preserve pre/post winners in the trace.
• Guardrail: later recovery does not make the unstable intermediate transition harmless.

High: weak run-model signal

• Severity: high.
• Affected games: all 14.
• Evidence: home/away/total R² of 0.0429/0.0253/0.0224.
• Action: investigate features and target through isolated training backtests.
• Guardrail: positive R² only means better than the mean baseline; it does not establish reliable exact-score forecasts.

Medium: high-raw probability cluster

• Severity: medium.
• Affected games: CHW@MIN, CIN@CHC, PHI@LAA.
• Evidence: only 1/3 high-raw picks won.
• Action: monitor across multiple slates and preserve lineage.
• Guardrail: do not infer a new high-confidence shrinkage rule from this slate alone.

Info: complete slate audit coverage

• Severity: info.
• Evidence: 14/14 games structurally audited.
• Action: retain a per-game receipt.
• Guardrail: complete coverage does not erase findings within audited games.

Info: simulation inert on slate

• Severity: info this slate, but escalated across slates.
• Evidence: 0 flip candidates and 0 published flips.
• Action: measure simulation contribution over rolling slates and distinguish score realization from winner flipping.
• Guardrail: zero flips is not itself proof of a defect or proof that simulation did no score work.

3. Every game

Published probability below is expressed for the published predicted side. The supplied home-minus-away confidence edge is retained separately.

Correct-winner/wrong-score games

The following five games got the winner right but not the exact score:

• BAL@OAK
• COL@ATL
• KC@CLE
• LAD@DET
• SD@TB

PIT@STL and PHI@LAA got both winner and exact score correct. The five correct-winner/wrong-score cases demonstrate why winner accuracy must not be treated as full prediction quality.

4. Miss reconstruction

Earliest wrong-direction movement by loss

BOS@NYY

• Earliest stage: no clear strong winner-direction regression is demonstrated. The raw probability was only 0.5085 for BOS, with season and recent edges near neutral and composite edge −0.1816.
• Likely issue class: starter projection and run-environment miss, not an obvious directional weighting failure.
• Suarez was projected for 3 ER and allowed 6; Warren was projected for 5 ER and allowed 1.
• The total was under-projected by 8 runs.
• The published BOS edge was only 0.06, a lean—not a high-conviction call.

CHW@MIN

• Earliest stage: the model was already strongly committed to CHW before the final publication layer. Raw pick probability was 0.7341; the fundamental composite home edge was −1.3982, with season −0.8087 and recent −0.726.
• Issue class: bad or incomplete directional evidence, bad weighting, or matchup-specific structural miss; exact cause is not identifiable from the evidence.
• The later score-volatility action expanded the blowout tail but did not correct the winner.
• Run environment was over-projected by 5, and Matthews allowed three fewer ER than projected.
• The publication lineage is also defective: 0.7117 calibrated versus 0.686674 published with no recorded transformation.
• This is the slate’s clearest combined model-direction and contract failure.

CIN@CHC

• Earliest stage: fundamental/model direction was already toward CHC: season edge +1.4567, recent +0.702, composite +1.4353, and raw home probability 0.7597.
• Issue class: likely bad weighting or overreliance on season/recent team-strength evidence, compounded by starter error.
• Imanaga was projected for 3 ER and allowed 6; Burns allowed 2 versus a projection of 4.
• A later guardrail materially reduced the published home probability to 0.56, but the pick remained CHC.
• The 0.12 edge was moderate, not a near coin flip. The model still made a directional commitment despite the relatively small published edge, while its upstream calibrated probability was much higher.
• The pitch-count warning is an input/contract concern but is not proven to have caused the winner miss.

HOU@NYM

• Earliest stage: the evidence was nearly neutral: raw 0.5311 for the published home side, composite edge +0.1017, and published edge +0.0833.
• Issue class: lean-side miss and irreducible game variance are more defensible than a clear model regression.
• Thornton was projected for 0 ER and allowed 2; Pecko was projected for 4 and allowed 2.
• Team-run error was 4, but total bias was 0.
• No confidence or score-volatility control produced a strong directional change.

MIA@WAS

• Earliest stage: the published side was supported by only modest fundamental evidence: season +0.0638, recent +0.074, pitcher +0.0222, composite +0.2824.
• Issue class: over-weighting a small home-side composite edge or ordinary variance; not a demonstrated starter failure.
• The model committed to WAS at a 0.15 home-away confidence edge, which is moderate.
• MIA scored three more runs than projected; both starter ER projections were close.
• The lineage field says no transformation explains a 1.6362-point calibrated-to-published difference, but the causal register does not classify it as the ≥2-point critical failure.

SEA@TOR

• Earliest stage: direction was weak and conflicted: season and recent evidence favored TOR, while pitcher evidence favored SEA; composite favored SEA by only −0.2614.
• Issue class: starter projection plus lean-side variance.
• Gilbert allowed 3 ER versus 1 projected; Scherzer allowed 0 versus 4 projected.
• SEA offense was over-projected by 5, TOR offense under-projected by 3.
• The published SEA edge was only 0.06.

TEX@MIL

• Earliest stage: direction was already materially toward MIL from season and recent evidence: +1.522 and +1.114, composite +1.2408.
• Issue class: wrong weighting or incomplete treatment of away-team strength; starter error was small.
• TEX offense was under-projected by 3; Rocker and May were each within 1 ER of projection.
• The model’s supported blowout expansion did not prevent choosing the wrong side.
• This loss is stronger evidence for directional team-strength weighting than for starter reweighting.

Classification of loss causes

• Bad input / contract: CIN pitch-count discrepancy; CHW publication-lineage discrepancy. Neither is proven to be the direct cause of a winner loss.
• Bad weighting / directional model: CHW, CIN, MIA, and TEX are the strongest candidates, though causal certainty is limited.
• Calibration: high-raw cluster losses in CHW and CIN; overconfidence is clearly present, but calibration did not necessarily determine the wrong side.
• Starter error: BOS, SEA, and part of CIN. KC had the largest starter error despite a correct winner.
• Offense/bullpen miss: BOS, CHW, CIN, MIA, SEA, and TEX show substantial team-offense residuals. Bullpen-specific outcomes are not supplied and must not be invented.
• Run-environment miss: BOS under-dispersed by 8 total runs; CHW over-dispersed by 5. The run model is weak globally.
• Irreducible variance: HOU, MIA, and the low-edge BOS/SEA games retain substantial variance after accounting for known projection errors.

Pipeline-stage audit across all stages

At historical_ml → variance_applied, direction changed in five games:

• BAL: BAL → tie
• BOS: BOS → tie
• CIN: CHC → tie
• MIA: WAS → tie
• PIT: PIT → tie

Two were explicitly classified as regressions: BAL@OAK and PIT@STL. The enhanced stage restored both. That recovery must not erase the intermediate regression.

From enhanced_prediction onward, no stage changed winner direction. Simulation produced zero winner flips, so its winner-direction contribution was inert on this slate, although its score contribution cannot be ruled out from the supplied aggregate trace.

5. Cross-game counterfactuals

Lever: score.variance_multiplier / score.variance_floor

Proposed effect: widen run distributions.

Helps:

• CHW@MIN: 0.37-confidence loss; wider tails could reduce overconfidence.
• CIN@CHC: 0.12-edge loss; wider tails could better represent uncertainty.
• MIA@WAS: 0.15-edge loss.
• TEX@MIL: 0.15-edge loss.

Hurts:

• PHI@LAA: strong correct call at 0.33; widening would soften a well-calibrated edge and potentially worsen Brier.

Net evidence: 4 helps versus 1 hurts, supporting a bounded shadow experiment for calibration/tail behavior, but not a large production change. The lever addresses uncertainty and score tails; it does not directly fix wrong team-strength direction.

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

Helps:

• BOS@NYY: both starters materially wrong.
• CHW@MIN: Matthews ER projection off by 3.
• CIN@CHC: Imanaga off by 3 and Burns by 2.
• KC@CLE: Lugo off by 7.
• PHI@LAA: Kikuchi off by 3.
• SEA@TOR: both starters materially wrong.

Hurts or risks harm:

• MIA@WAS: starter projections were close.
• SD@TB: both starters exactly matched ER projections.
• TEX@MIL: starters were within 1 ER.

Net evidence: 6 helps versus 3 hurts. This supports investigation and a small shadow change, but not aggressive reweighting because the lever would disturb several games where starter projections were accurate.

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

Helps:

• BOS@NYY
• CHW@MIN
• CIN@CHC
• COL@ATL
• HOU@NYM
• KC@CLE
• LAD@DET
• MIA@WAS
• SEA@TOR
• TEX@MIL

Hurts or risks harm:

• BAL@OAK
• PHI@LAA
• PIT@STL
• SD@TB

Net evidence: 10 helps versus 4 hurts, but “helps” here means lower score error after observing outcomes, not proof that increasing or decreasing the weights would have helped prospectively. The direction of the parameter change is not registered in the evidence. Do not force a one-game directional adjustment from this count alone.

Tempting one-game fixes that must not be made

• Do not simply fade CHW because it lost at 0.37 confidence; PHI@LAA was a strong correct call and the high-raw cluster is only three games.
• Do not remove score-volatility expansion because CHW and CIN lost; the lever was intended to widen tails, and the issue is not equivalent to winner selection.
• Do not broadly increase pitcher recency/form weight because of KC’s 7-ER Lugo miss; SD and MIA had accurate starter projections.
• Do not reduce season weighting solely because TEX@MIL lost; COL@ATL was directionally correct using similar season-strength evidence.
• Do not treat zero simulation flips as proof that simulation should be disabled; first verify execution and score contribution.
• Do not call the slate clean merely because structural critical/error counts were zero.

6. Data and contract defects

Critical findings

1. Unexplained CHW@MIN probability publication delta
2. Calibrated/governed probability: 0.7117.
3. Published pick probability: 0.686674.
4. Delta: −0.025026.
5. No recorded transformation explains it.
6. This is a contract/provenance failure and blocks promotion until traced.

Error findings

• Structural audit summary reports zero error findings.
• This does not eliminate model-quality errors such as poor score R² or starter misses; those are prediction defects, not structural audit errors.

Warning findings

1. CIN@CHC pitch-count mismatch
2. Chase Burns: summary H=3, BF=21.
3. Pitching box score: H=2, BF=20.
4. Source/category: simulation.
5. Severity: warning.
6. Canonical artifact warning
7. Aggregate audit totals: 0 critical, 0 error, 1 warning, 100 info.
8. The named presentation/stat contract is not included, so reconciliation remains incomplete.

Info findings

• Structural audit reports 100 info findings in aggregate.
• The supplied evidence does not enumerate each informational message individually. Therefore, they can be counted but not independently itemized without inventing details.
• Complete coverage is informationally positive but not a clean bill of health.

Presentation/artifact versus model defects

Prediction-model defects

• Weak validated run-model signal: R² of 0.0253–0.0429.
• Persistent margin underprediction.
• Starter projection errors, especially BOS, SEA, CIN, and KC.
• Overconfidence in all reported confidence buckets.
• Wrong directional calls in CHW, CIN, MIA, and TEX.

Presentation/artifact defects

• Missing integrity revision audit.
• CHW calibrated-to-published lineage gap.
• MIA transformation field also incomplete, though below the causal register’s material-delta threshold.
• CIN pitch-count contract mismatch.
• Intermediate stage winner instability requiring better trace preservation.

Complete pick-side probability lineage

The expected lineage is:

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

Examples:

• BAL@OAK: raw 0.568 / governor input 0.58 / governed 0.5662 / score coherence 0.575584 / published 0.551397; transformation recorded as score coherence plus matchup-recency guardrail.
• CIN@CHC: raw 0.7597 / governor input 0.56 / governed 0.7444 / score coherence 0.749693 / published 0.56; transformation recorded as score coherence plus matchup-recency guardrail. The large input-to-governor discrepancy is recorded, but its semantic origin deserves review.
• PHI@LAA: raw 0.67 / governed 0.6609 / score coherence and published 0.665459; transformation recorded.
• CHW@MIN: raw 0.7341 / governor input 0.704517 / governed 0.7117 / score coherence 0.733343 / published 0.686674; the final publication delta is not recorded. This is the material bypass-like gap, but the evidence does not establish an actual governor bypass.
• MIA@WAS: governed 0.5913 / score coherence 0.588162 / published 0.574938; transformation field says unexplained, but the delta is approximately 1.64 points.

A bypass should be called only where a material delta lacks a recorded transformation. On the supplied evidence, that criterion is met for CHW@MIN, not conclusively for MIA@WAS.

7. Cross-slate persistence

The trend window covers three slates: 2026-08-28 through 2026-08-30, with two prior slates loaded.

Escalated findings

These are established failures, not single-slate noise.

simulation_inert_on_slate

• Consecutive-slate count: 3.
• Present now: yes.
• Metric group: run model.
• Target lever listed: simulation.allow_score.
• However, param_actionable=false; the registered action is diagnostic/code verification, not a parameter promotion.
• Metric harmed: team_run_mae, total_run_mae.
• Direction: confirm the simulation module is running and not silently no-oping.
• This slate had zero flip candidates and zero published flips, consistent with the persistent pattern.

weak_run_model_signal

• Consecutive-slate count: 3.
• Present now: yes.
• Metric group: run model.
• Listed target levers: training.xgb.n_estimators, training.xgb.max_depth, training.xgb.learning_rate.
• param_actionable=false; the prescribed action is an isolated training/backtest investigation, not immediate parameter movement.
• Metric harmed: team_run_mae, total_run_mae.
• Direction: investigate feature and target signal using a manual isolated backtest.

Other finding streaks

• high_raw_probability_cluster: consecutive count 1; present now yes. It affects Brier and calibration error. This remains a single-slate signal despite its current presence.
• pipeline_winner_instability: consecutive count 1; present now yes. It affects accuracy and Brier. It requires diagnostic trace repair, not an automatic parameter move.
• single_slate_calibration_instability: consecutive count 0; present now no. The record says incidence in window 2, but it is not currently persistent under the streak definition.

Metric trends

• Margin bias: regressing. Mean −2.0003; series −1.333, −1.882, −2.786. The model is increasingly underestimating winning margins.
• Mean team-run error on losses: regressing. Mean 5.0; series 5.0, 3.0, 7.0.
• Weighted calibration error: improving. Mean 0.2278; series 0.2179, 0.3961, 0.0693. The current slate improved materially, although its bucket-level overconfidence remains real.
• Winner accuracy: regressing. Mean 0.636; series 0.4667, 0.9412, 0.5.

Top-confidence-pick streak

• 2026-08-28: hit.
• 2026-08-29: hit.
• 2026-08-30: miss.
• Current loss streak: 1.

Structural raw blocks

• Margin distribution: six actual margins were ≥4, while ten predictions were ≤2; this slate exhibits substantial tail underestimation.
• Pipeline transitions: variance application caused a −2-hit winner regression, later recovered by enhanced prediction; transition instability is real on this slate.
• Calibration monitoring: all three buckets overconfident; weighted absolute error 0.0693.
• High-raw cluster: 1/3 hits, with CHW and CIN losses.
• Simulation control: 14 observed games, zero candidates, zero published flips; diagnostic-only status, now persistent over three slates through the escalation register.

8. Adjustments for today

Ranked by persistence, scope, expected benefit, and regression risk.

1. Run simulation execution diagnostic — highest persistence

• Pattern: simulation_inert_on_slate, persistent for 3 slates.
• Registered lever: simulation.allow_score.
• Recommended change: no parameter move; perform the registered code/diagnostic action to confirm that simulation.allow_score is actually read and that score simulation is executing.
• Expected metric: team-run MAE and total-run MAE should show measurable simulation contribution; direction is improvement/lower error.
• Helps: potentially the broad run-model problem across all 14 games, especially large-tail cases such as BOS@NYY, CHW@MIN, KC@CLE, and SEA@TOR.
• Hurts/risk: no direct game-level harm should be introduced if this is diagnostic-only; risk is confusing score realization with winner flips.
• Gate: retain pre/post score and winner traces; roll back or quarantine any change that alters winner direction without improving rolling run metrics.

2. Isolated run-model backtest — second-highest persistence

• Pattern: weak_run_model_signal, persistent for 3 slates.
• Registered levers: training.xgb.n_estimators, training.xgb.max_depth, training.xgb.learning_rate.
• Recommended change: do not promote a parameter move because the finding is registered param_actionable=false. Run an isolated backtest across these parameters and inspect feature/target signal.
• Expected metric: lower team-run MAE and total-run MAE; no deterioration in winner accuracy or calibration.
• Helps: broad score errors, particularly BOS, CHW, KC, SEA, and the overall margin tail.
• Hurts/risk: may disturb accurate score calls at PHI@LAA, PIT@STL, BAL@OAK, and SD@TB.
• Gate: shadow only; require improvement across rolling slates, not one-game score gains.

3. Bounded confidence-cap shadow experiment — strongest current single-slate driver

• Pattern: high-raw probability cluster, present now for 1 slate; three high-raw games, one hit.
• Registered levers: confidence.empirical_max_weight, confidence.base_cap, confidence.moderate_cap.
• Recommended bounded change: shadow a small 0.02 downward adjustment to confidence.moderate_cap only, without changing raw model direction.
• Expected metric: lower Brier score and lower calibration error through reduced overconfidence; winner accuracy is not expected to improve directly.
• Helps: CHW@MIN, CIN@CHC, MIA@WAS, and TEX@MIL by reducing commitment strength.
• Hurts: PHI@LAA, where a strong correct call could be unnecessarily softened.
• Risk: current evidence is only one slate; do not promote absent rolling improvement.
• Gate: compare published probabilities and Brier/calibration metrics against unchanged control; rollback if PHI-like strong calls degrade without reducing high-confidence loss severity.

4. Probability-lineage contract repair — mandatory, not a tuning experiment

• Pattern: CHW@MIN unexplained material publication delta.
• Registered parameter: none appropriate.
• Action: trace pick-side semantics and record every calibrated → governor → score-coherence/recency → published transformation; include the missing integrity revision proof.
• Expected metric: zero unexplained material deltas and complete auditability.
• Helps: CHW@MIN directly and the integrity of every future slate.
• Hurts: no modeling lever should be changed as part of this repair.
• Promotion rule: block promotion until reconciled.

9. Bottom line

Known:

• The slate finished 7–7, or 50% winner accuracy.
• Calibration was overconfident in every bucket; weighted absolute calibration error was 0.0693.
• Exact-score/run modeling remains weak, with R² values between 0.0224 and 0.0429.
• Five games had correct winners but wrong scores.
• Starter projection error materially affected BOS@NYY, SEA@TOR, CIN@CHC, and KC@CLE.
• The run pipeline had an intermediate winner-direction regression at variance application and no winner-changing simulation intervention.
• Simulation inertness and weak run-model signal are persistent over three slates and must be treated as established failures.
• CHW@MIN has a critical unexplained probability-publication delta.
• Structural coverage is complete, but the integrity revision audit is missing and one pitch-count warning remains.

Uncertain:

• Whether CHW@MIN’s publication delta represents a governor bypass, a pick-side semantic mismatch, or an undocumented legitimate transformation.
• Whether the high-raw cluster reflects a persistent calibration defect or short-term variance.
• Whether changes to recent/season run weighting would improve future directional accuracy; the cross-game count is suggestive but not causal.
• Whether simulation is entirely inactive or merely produced no winner flips; score-level contribution is not isolated here.
• Bullpen-specific causes cannot be assessed because bullpen event details are absent.

Single safest next action: block promotion and repair the probability-lineage/integrity contract—especially CHW@MIN—while running a diagnostic-only verification that simulation.allow_score executes. This improves artifact reliability without making an unproven one-slate model adjustment.