Luna deep analysis — slate 20261005

Coverage receipt: COMPLETE · 9/8 sections; 2/2 games; 5/5 causes; 6/6 sources

1. Evidence receipt and gaps

Slate reviewed: 2026-10-05. The diagnosis reports 2 games reviewed, and both are represented in the game-level evidence: CHW@CLE and NYY@TB.

Source manifest — every entry accounted for:

Coverage: coverage_matrix reports 2 expected games, 2 canonical artifacts loaded, 2 continuous-learning games, 2 forensic deep-audit games, 2 games with starter comparisons, and 2 structural audits completed. Thus, the diagnosis claims complete coverage of the two-game slate. That does not make the slate clean: NYY@TB has an error-level structural finding, and the source manifest also lacks the integrity revision audit.

Internal contradictions and limitations:

• CHW@CLE’s top-level starter comparison reports Anthony Kay with an actual line of 1 ER, 1 out. The player-level matchup breakdown instead says his settled per-player pitching line is missing. Those are incompatible evidence states; the actual pitching line cannot be independently reconciled from this package.
• The NYY@TB structural error says a recency cap was violated at target_confidence=0.540 against allowed=0.532. The published pick probability is lower than calibrated confidence, but the package does not explain how the audited enhanced-prediction target and final published value relate at that point. The explicit contract violation stands; the final lower probability does not erase it.
• The diagnosis reports 14 informational structural findings in total, but provides no itemized list of those 14 findings. Their count can be confirmed, not their individual content.
• Some values are presented at different levels of aggregation. For example, the NYY@TB top-level published home-minus-away edge is −0.0792, while the probability lineage provides the underlying published pick probability for the away side, 0.539619. These are consistent after rounding (home ≈ 0.460381), not a substantive contradiction.
• Reported “total bias” uses predicted total minus actual total: CHW@CLE is +2 (9 predicted versus 7 actual), and NYY@TB is −1 (6 versus 7). I use that convention below. Team-run error is reported as the sum of the two teams’ absolute run errors: 2 and 5 respectively.

2. Executive diagnosis

Winner accuracy: 0/2 (0%). Both published selections lost: CLE over CHW and NYY over TB. The historical model’s reported walk-forward accuracy is 0.570149, but that is a historical benchmark, not evidence that this two-game slate should have been accurate. The self-assessment says this slate beat neither coin flip nor validated baseline.

Calibration quality: Both outcomes missed. The reported weighted absolute calibration error is 0.5291 over only 2 settled games; in the one represented 0.50–0.55 bucket, average confidence was 0.5291, observed hit rate was 0/2, and the bucket was labeled overconfident. This is a severe single-slate result and an established cross-slate finding (3 consecutive slates), but it is not a reliable stand-alone calibration curve. The causal register’s single_slate_calibration_instability entry is high severity: preserve its two-game evidence and its guardrail not to infer a stable curve from one slate; because it has persisted, Section 8 recommends a bounded shadow experiment rather than a production change.

Score accuracy: Both exact score forecasts missed. CHW@CLE’s team-run error was 2 and total error was 2; NYY@TB’s were 5 and 1. Pipeline-stage scores say enhanced_prediction improved aggregate run errors versus historical_ml, then embedded_savant_run_environment returned them to the worse-than-base values that persisted through publication. Across the two games, the largest reported team-run regression is at variance_applied (team-run MAE +0.1054 and margin MAE +0.2634 versus historical ML). These are two-game metrics and do not establish broad stage performance.

The run models’ historical holdout R² values are near zero: home 0.002143, away 0.030936, total 0.002415, on 2,415 rows. The causal register calls weak_run_model_signal high severity. It is a material model-validation risk, not proof that any single score miss was avoidable. It has persisted 3 consecutive slates, so the appropriate action is an isolated feature/target backtest, not a nightly score-parameter adjustment.

Starter accuracy: The available game-level comparisons show notable errors, especially Freddy Peralta’s NYY@TB line. However, one CHW@CLE starter’s actual line conflicts with the player-level feed description, so starter accuracy cannot be treated as fully reconciled. The player-level verdicts attribute this slate’s side-level breakdowns to offense on two sides and variance on two sides, with no side classified as a material pitching break. That does not make Peralta’s individual miss immaterial: he allowed 1 ER in 14 outs against a projection of 4 ER in 16 outs.

Artifact integrity: There is one error-level structural audit finding, for NYY@TB, and no critical findings. CHW@CLE’s saved artifact is reported clean. The causal register labels artifact_integrity_failure high severity and instructs that this slate be excluded from parameter learning until the artifact-generation step is fixed. Follow that action: learning eligibility is false. A malformed artifact is a file/contract defect, not proof that the model’s prediction was wrong; the actual NYY@TB pick also missed, but that does not establish the artifact caused the miss.

Causal-register reconciliation:

• artifact_integrity_failure — high, artifact integrity. Evidence: 1 error, 0 critical, 0 warnings and 14 informational findings; affected game NYY@TB. Action: exclude this slate from parameter learning until artifact generation is fixed. Guardrail: do not attribute accuracy changes to the malformed artifact.
• single_slate_calibration_instability — high, calibration. Evidence: weighted absolute calibration error 0.5291 across two games; both games affected. Action in the register: watch rolling confidence buckets before changing a calibration weight. Guardrail: two games cannot establish a stable calibration curve. Cross-slate escalation now makes a shadow test appropriate, not an unguarded production change.
• weak_run_model_signal — high, model validation. Evidence: home/away/total holdout R² of 0.0021/0.0309/0.0024; both games in the evidence. Action: treat exact run scores as low-confidence estimates and seek stronger features or a better target. Guardrail: small positive R² means only better than predicting the average; it does not make exact scores reliable.
• complete_slate_audit_coverage — info, coverage. Evidence: 2/2 games structurally audited. Action: retain per-game audit receipts. Guardrail: coverage does not cancel findings inside audited games.
• simulation_inert_on_slate — info in the causal register. Evidence: zero flip candidates and zero published flips. Action: measure simulation contribution over rolling slates, separating score realization from winner flips. Guardrail: zero flips is not itself a defect and does not mean no score-reconciliation work occurred. Cross-slate persistence has since escalated this to an actionable code/diagnostic issue; it remains inappropriate to force a parameter change.

No game had a correct winner. The diagnosis’s published probabilities and confidence edges are not equivalent to a half-scale distance from 50%; the stated system edges are modest to moderate, not near-certain calls.

3. Every game

CHW@CLE — miss

• Prediction and result: Predicted CHW 4, CLE 5; picked CLE. Actual CHW 4, CLE 3; winner CHW. Wrong winner and wrong score.
• Published probability / edge / bucket: Published home probability 0.5185, away 0.4815; published home-minus-away edge +0.037, reported as a 0.04 system edge. Conviction bucket: toss_up. Brier score: 0.268842.
• Score error: Team-run absolute error 2 (CHW exact; CLE predicted 5 versus 3 actual). Predicted total 9, actual total 7; predicted-minus-actual total bias +2.
• Projected versus actual starters (top-level comparison):
• CHW, Anthony Kay: projected 1 ER, 14 outs, ERA 1.93; top-level actual 1 ER, 1 out, ERA 27.00. ER error 0; outs error −13; ERA error +25.07.
• CLE, Gavin Williams: projected 3 ER, 17 outs, ERA 4.76; actual 2 ER, 15 outs, ERA 3.60. ER error −1; outs error −2; ERA error −1.16.
• Evidence conflict: the player-level breakdown says Anthony Kay’s settled pitching line is missing, so his top-level “actual” line cannot be reconciled against that feed here.
• Pregame edges (home minus away): Season −0.1605; recent −0.030; pitcher +0.049; composite +0.192. Season quality edge was −0.2685, with no margin adjustment because it was below the action threshold; recent quality edge was +0.04. The pitcher factors were 0.9656 for CHW runs versus the CLE pitcher and 1.0146 for CLE runs versus the CHW pitcher.
• Confidence and recency: Raw pick probability 0.551861 for the home pick; calibrated/governed pick probability 0.5256 (calibration sample 821; empirical hit rate 0.5639); score-coherence probability 0.525605; published home probability 0.5185. Recency guardrail is recorded as applied; discount factor and reason are null. No starter-experience shrinkage was applied (minimum 32 games; shrink factor 1.0).
• Score path: score_volatility_actions contains only no_change. The aggregate stage audit shows historical ML → variance applied worsened run errors; enhanced prediction recovered them; embedded Savant run environment worsened them again; subsequent stages through publication were unchanged. Per-game numeric scores at every intermediate stage are not supplied, so none are imputed.
• Player-level evidence: CLE’s lineup scored 2 fewer runs than projected. Named contributors include Angel Martínez (0-for-4 versus projected 1 hit and 1 HR; −1.4 run impact; xwOBA 0.299, below average), José Ramírez (0-for-3 versus projected 1 hit; −0.78), and Jo Adell (+0.64; xwOBA 0.329, above average). Player-level verdicts: CLE offense broke below projection; CHW side was closer to variance. The top-level diagnostic says the away team’s run forecast was exact.
• Audit and causes: Per-game structural audit PASS, 0 defects, risk score 0; 7 informational notes. The artifact is not identified as a cause. Primary driver: toss_up_variance; no secondary driver. The evidence supports a narrow CLE miss in a toss-up, plus CLE offensive underperformance; it does not establish an avoidable winner-model defect. This was not a correct-winner/wrong-score game.

NYY@TB — miss

• Prediction and result: Predicted NYY 4, TB 2; picked NYY. Actual NYY 2, TB 5; winner TB. Wrong winner and wrong score.
• Published probability / edge / bucket: Published away-pick probability 0.539619 for NYY; implied home probability 0.460381. Published home-minus-away edge −0.0792, or a 0.08 edge toward NYY; bucket: lean. Brier score: 0.291189.
• Score error: Team-run absolute error 5 (NYY 4 versus 2; TB 2 versus 5). Predicted total 6, actual total 7; predicted-minus-actual total bias −1.
• Projected versus actual starters:
• NYY, Cam Schlittler: projected 1 ER, 18 outs, ERA 1.50; actual 2 ER, 13 outs, ERA 4.15. ER error +1; outs error −5; ERA error +2.65.
• TB, Freddy Peralta: projected 4 ER, 16 outs, ERA 6.75; actual 1 ER, 14 outs, ERA 1.93. ER error −3; outs error −2; ERA error −4.82. Player-level synthesis describes this as an outlier start, with no credible venue, opponent-strength, or experience signal explaining it.
• Pregame edges (home minus away): Season −0.1757; recent −0.182; pitcher −0.020; composite −0.5461. Season quality edge was −0.2853 with no margin adjustment below the action threshold; recent quality edge was +0.028. Run factors were 1.0217 for NYY runs versus the TB pitcher and 1.0017 for TB runs versus the NYY pitcher.
• Confidence and recency: Raw pick probability 0.5808 for NYY; governor input 0.532112; calibrated/governed pick probability 0.5466 (sample 485; empirical hit rate 0.5794); score-coherence probability 0.54661; published away-pick probability 0.539619. The recency guardrail is recorded as applied, but its discount factor and reason are null. No starter-experience shrinkage was applied (minimum 32 games; factor 1.0).
• Score path: score_volatility_actions contains only no_change; a large_savant_run_environment_swing diagnostic flag is recorded. The aggregate stage path and limitations are the same as above: enhanced prediction improved run errors versus base, embedded Savant run environment reversed that gain, and published values stayed at that later stage.
• Player-level evidence: TB’s offense was under-projected by 3 runs according to contributing_factors, while NYY’s lineup was 3 runs below its projection in the matchup breakdown. Ben Rice hit 2-for-3 with 2 HR versus a projection of 1 hit and 1 HR (+1.4 run impact; xwOBA 0.378, elite). Other listed deviations include Trent Grisham (−0.8), Luis García Jr. (−0.78), Austin Wells (+0.78), and Cody Bellinger (−0.64). Peralta’s unexpectedly strong start is separately identified as variance. Player-level side verdicts classify NYY as offense below projection and TB as closer to variance.
• Audit and causes: Structural audit ERROR, risk score 40. One error: MATCHUP_RECENCY_GUARDRAIL, stating target confidence 0.540 exceeded the allowed 0.532 cap. This is an artifact/contract failure, not evidence that the model’s pick was wrong because of the file defect. Primary driver: lean_side_miss (“Small edge on the losing side”); no secondary driver. This was not a correct-winner/wrong-score game.

4. Miss reconstruction

CHW@CLE

1. Earliest wrong-direction evidence: The selected side was already CLE. The raw model assigned the home pick probability 0.551861 and the fundamental composite home edge was +0.192, despite season and recent home-minus-away edges of −0.1605 and −0.030. The evidence supplied does not expose enough of the scoring formula to say exactly which component first caused the home direction; the contradiction between the negative season/recent edges and positive composite is worth auditing, but is not itself proof of a bad input.
2. Bad input / bad weighting: No confirmed bad input is established. The available aggregate fundamentals are mixed; composite favors CLE while season and recent favor CHW. The package cannot identify whether that combination reflects an input error or weighting choice.
3. Calibration / confidence: Calibration and recency steps lowered the home pick probability from raw 0.551861 through 0.5256 to published 0.5185. They softened, but did not reverse, the wrong selection. The published edge was a toss-up-level 0.037, not a near-certain call.
4. Starter error: Gavin Williams was 1 ER and 2 outs better than projected. Kay’s game-level line appears accurate in ER but badly short in outs; the conflicting missing-line record prevents a clean starter conclusion. The player-level verdict does not classify either side as a material pitching break.
5. Offense/bullpen: CLE scored 2 fewer than projected; the player-level breakdown points to the CLE lineup, including the large Angel Martínez shortfall. CHW’s offense was not a material break in the breakdown. CLE bullpen recorded 2 ER in the player feed, but the available diagnosis does not attribute the winner miss to bullpen performance.
6. Run environment / variance: The forecast total exceeded actual by 2. The player-level data also says hitters’ slate projections were broadly high, but this two-game slate is too small to establish a run-environment correction. A one-run actual margin and a 0.037 edge are compatible with substantial outcome variance.

Conclusion: The earliest identified wrong selection is the pre-publication CLE side; later confidence controls did not repair it. The direct game-level driver is a toss-up result with CLE’s offense below forecast. No single confirmed upstream bad input is established.

NYY@TB

1. Earliest wrong-direction evidence: The selected side was NYY before publication. The available season, recent, pitcher, and composite home-minus-away edges all favored away NYY (−0.1757, −0.182, −0.020, −0.5461). Thus, the pregame evidence as recorded consistently pointed to the eventual losing side. The miss is not explained by a sign reversal in those listed fundamentals.
2. Bad input / bad weighting: No confirmed bad input or demonstrated weighting defect is isolated. The recorded inputs favored NYY, but the realized game was materially different.
3. Calibration / confidence: The raw NYY pick probability was 0.5808; calibration/governor and subsequent recorded controls lowered it, and publication was 0.539619. The system remained on NYY. A 0.0792 home-away confidence edge is moderate, not a near coin flip, but also not a high-conviction selection. The separate recency-cap contract error is real and requires correction; the evidence does not show it created the miss.
4. Starter error: Peralta allowed 3 fewer ER than projected and was the most striking starter miss. The per-game synthesis labels it an outlier and says no listed context signal credibly explains it. Schlittler allowed 1 more ER and recorded 5 fewer outs than projected.
5. Offense/bullpen: TB’s offense was under-projected by 3, while NYY’s lineup was 3 below projection. The player evidence is mixed, with Ben Rice above his projection and several NYY hitters below. The listed bullpen ER are 0 for the relevant team breakdowns, so the diagnosis does not support bullpen error as the principal explanation.
6. Run environment / score: Predicted team runs were NYY 4 and TB 2; actual was NYY 2 and TB 5. Thus, the forecast substantially missed TB’s runs and overshot NYY’s, while total runs were only one below actual. This is primarily a team allocation miss, with both Peralta’s outlier start and offense deviations contributing; it is not a simple overall-total miss.
7. Variance: Peralta’s outlier performance and individual hitter deviations are material sources of irreducible game-level variance. The package does not justify changing starter weights based on this one start.

Conclusion: Unlike CHW@CLE, the recorded pregame fundamentals consistently favored the eventual loser. The miss is most plausibly a combination of starter outlier and team-run allocation/offense miss, not a demonstrated bad input. The recency contract defect is an independent process issue.

Pipeline-stage quality across all stages and transitions

All reported stages have 2 games and 0/2 winner accuracy. A later score recovery does not erase an earlier regression.

Transitions report no per-game changed directions, improvement/regression lists, winner flips, or winner-accuracy changes. They do report aggregate run-error changes: historical ML → variance worsened; variance → enhanced improved; enhanced → embedded run environment worsened; all following transitions were unchanged. Therefore, the data establishes aggregate score-error regression and recovery, not an intermediate game-by-game winner-direction regression. The later recovery does not justify ignoring the variance-stage or embedded-environment regressions.

5. Cross-game counterfactuals

Every lever_tension entry is accounted for:

1. score.variance_multiplier / score.variance_floor (widen run distributions)
2. Helps: none listed.
3. Hurts: none listed.
4. Slate evidence: no helped/hurt games; both picks missed, but the entry’s stated targets are Brier/calibration, not a demonstrated per-game score improvement. The cross_slate_bias entry separately says the variance-floor signal is not separable from noise.
5. Net: No game-level net evidence for a production change. Calibration persistence supports only the separately specified confidence shadow experiment; it does not establish this score-variance lever as the calibration fix.

1. player.pitcher.v4 form weights / player.pitcher_prior_games
2. Helps: NYY@TB, where Peralta’s ER were off by 3.
3. Hurts: CHW@CLE, where the reported starter ER comparison was accurate to within 1 ER and reweighting could disturb it.
4. Net: Balanced one-game evidence, with Peralta explicitly characterized as an outlier. Do not change pitcher weights on this slate.

1. recent.* / season.* run-mean weighting
2. Helps: NYY@TB, whose team runs were off by 5.
3. Hurts: CHW@CLE, whose team-run error was 2.
4. Net: Balanced. The game errors do not reveal a safe direction for changing either recent or season weighting. If explored, use a small shadow-only test and compare both games and rolling metrics; do not deploy a one-game fix.

Tempting fixes to reject:

• Reweight pitcher form because Peralta beat his projection: the diagnosis calls the start an outlier, and the pitcher lever helps NYY@TB but risks CHW@CLE.
• Change run-mean weighting solely to correct TB’s 3-run underprojection: that is a one-game remedy and the lever explicitly has a countervailing CHW@CLE risk.
• Widen score variance merely because both winners were missed: the slate has only two games; the lever’s game-level helped/hurt list is empty, and widening does not repair a demonstrated mean-direction error.
• Treat a “toss-up” as a half-scale distance from 50%, or treat the NYY@TB 0.0792 edge as a near coin flip. The supplied diagnosis explicitly warns that the system edge is not that half-scale measure.

6. Data and contract defects

Enumerated audit findings

• Critical: 0.
• Error: 1, NYY@TB MATCHUP_RECENCY_GUARDRAIL. Expected the final enhanced-prediction target confidence to be at or below the soft cap. The finding gives cap 0.540, allowed 0.532, target confidence 0.540. This is an error-level contract defect.
• Warning: 0.
• Info: 14 total across the structural audit. The package provides counts, but not an itemized list of all 14. The per-game audits report 7 informational notes each; CHW’s detailed list is not supplied, and the total does not permit reconstruction of every item.

The causal register separately categorizes the canonical artifact issue as a high-severity artifact-integrity concern because there is one structural error. That does not upgrade it to an audit “critical”: structural counts explicitly show zero critical. The CHW game audit is PASS with no defects; NYY@TB is ERROR. Coverage is 2/2, but completeness of auditing is not clearance of the finding.

Prediction-model defects versus presentation/artifact bugs

• Prediction/model risks: Near-zero holdout run-model R² is a model-validation problem; slate-specific wrong sides, score allocation, and calibration outcomes are prediction performance. They are not established consequences of the file defect.
• Artifact/contract defect: The NYY@TB recency-cap violation is a recorded structural error. Exclude this slate from parameter learning as directed.
• Missing proof artifact: The integrity revision audit is absent, so the requested remediation proof cannot be checked.
• Presentation/rendering: No separate presentation bug is identified in the supplied findings. Do not infer one from the missing revision audit or the model miss.

Pick-side probability lineage audit

Both lineages record the steps from raw model through calibration/governor, score coherence/recency, and published value; each says the recency guardrail applied and unexplained_material_delta=false. The recency discount factors are null. That missing factor is a traceability gap, but the package does not demonstrate an unrecorded material bypass: both published-minus-calibrated changes are small and a named transformation is recorded. The NYY@TB cap violation is an explicit contract finding even though the final published value is below the calibrated value; the lineage does not explain the mismatch between the intermediate target confidence and final publication.

The probability-reversion audit reports two artifacts checked, zero blend flips, two blend sides surviving to publication, and no overwritten games. Thus, there is no reported probability-side overwrite explaining either miss.

7. Cross-slate persistence

Raw structural evidence

• Margin distribution: Actual margins were 1 and 3; predicted margins 1 and 2. Both predictions were at most 2, and neither actual margin was at least 4. Errors were 0 for CHW@CLE and −1 for NYY@TB. This slate alone does not show large-margin outcomes that were compressed into short predicted margins.
• Pipeline transitions: All 7 transitions report zero winner-accuracy delta and no game-level changed directions. Run-error transitions show regression from historical ML to variance, recovery at enhanced prediction, regression at embedded run environment, and no later change.
• Calibration monitoring: 2 settled games; weighted absolute calibration error 0.5291. The 0.50–0.55 bucket had confidence 0.5291, 0 hits out of 2, and was labeled overconfident. Cross-slate trends say this metric is regressing.
• High-raw cluster: Threshold 0.65; 0 games, 0 hits. This slate supplies no high-raw-pick result.
• Simulation control: simulation_flip_control, priority P2, diagnostic-only without independent support; 2 games, zero flip candidates, zero published flips. This alone is not proof of a defect, but the persistent pattern below is escalated.
• Model validation: 2,415 rows; home/away/total R² 0.002143/0.030936/0.002415; walk-forward accuracy 0.570149 and Brier 0.242204; reported winner accuracy 0.600414 and winner Brier 0.239378. The run targets have little explained holdout variance.

Every finding streak

Every escalated finding

These are established failures under the provided escalation policy, not single-slate noise:

• simulation_inert_on_slate — 11 consecutive slates. Target lever recorded as simulation.allow_score; target metrics team-run MAE and total-run MAE. But param_actionable=false and the specified direction is to confirm the module is running rather than silently no-oping. The self-assessment identifies a code-level root cause: target-realization search and win-probability distribution use the same early-exiting draws (about two by default), so the estimated win rate is not a reliable independent probability. Treat this as a code/diagnostic action, not a parameter move.
• single_slate_calibration_instability — 3 consecutive slates. Target levers: confidence.empirical_max_weight, confidence.base_cap, confidence.moderate_cap; target metrics Brier score and calibration error. The direction is to increase empirical shrinkage. This is persistent and parameter-actionable; use a bounded shadow experiment.
• weak_run_model_signal — 3 consecutive slates. Target levers list training XGBoost parameters, but the finding is explicitly param_actionable=false. The recommended action is manual, isolated investigation of features/target, not a nightly hyperparameter move. The self-assessment points to adding features such as source-labelled no-vig market probability, park/weather, bullpen rest/fatigue, and platoon-adjusted lineup quality, then retraining and comparing an isolated same-fingerprint backtest. Treat this as a feature/target and training diagnostic/code action, not a parameter adjustment.

Trend directions and top-pick streak

• margin_bias: improving (mean −1.8919).
• mean_team_run_error_on_losses: improving (mean 3.9667).
• weighted_calibration_error: regressing (mean 0.34).
• winner_accuracy: regressing (mean 0.6754).
• top_confidence_pick loss streak: 1. The 2026-10-05 top pick was a miss, following a hit on 2026-10-04. The listed 11-slate sequence includes two unknown-date/unknown-result slots, so the supplied streak value is the authoritative reported count; do not infer additional consecutive losses through unknown entries.

Cross-slate bias evidence

The two registered bias checks are on hold, not persistent enough to move their parameters:

• Starter ERA error, actual minus projected: mean −0.163 over 10 slates; same sign 5/10; z −0.42; marked not persistent. The reading is that starters beat projections and opposing offenses may be over-credited, but the register says no change.
• Lineup run delta, actual minus projected: mean −0.463 over 10 slates; same sign 6/10; z −1.43; marked not persistent. The reading is that lineups scored below projection, but the register says no change.

The recent two slates do show large negative lineup deltas (−2.0 on 2026-10-04 and −2.25 on 2026-10-03), but that does not override the supplied 10-slate noise assessment.

8. Adjustments for today

These are recommendations for bounded tests or operator action, not changes applied here. The slate is quarantined from learning, and no promotion or file changes have been made.

Ranked recommendations

1. Highest evidence strength — calibration persistence; bounded shadow experiment
2. Exact registered parameter: confidence.empirical_max_weight.
3. Direction and magnitude: Increase empirical shrinkage modestly: test +0.02 absolute weight versus control, bounded to the parameter’s valid configured range. Keep confidence.base_cap and confidence.moderate_cap fixed in this experiment so the effect can be attributed to one lever.
4. Expected metric/direction: Lower rolling weighted calibration error and improve Brier score without materially degrading winner accuracy.
5. Scope and tension: Applies to confidence outputs across games, including the two misses here. The score.variance_multiplier / score.variance_floor tension entry lists no game-level helps or hurts; no specific game-level tradeoff is established. The two-game calibration result is poor, but the 3-slate escalation, not this one slate alone, supports the shadow test.
6. Gate/rollback monitor: Run in shadow against the current baseline over the registered rolling evaluation window; do not apply to this quarantined slate. Reject or roll back if rolling Brier worsens or calibration error rises, or if winner accuracy materially deteriorates. Do not treat the two observed misses as proof the shadow is causal.

1. Persistent, non-parameter issue — simulation probability diagnostic/code action
2. Finding target: simulation.allow_score is listed, but the finding is explicitly non-parameter-actionable.
3. Exact action: Verify that the simulation module is executing and not silently no-oping; separate score-realization search from an independent, unsteered win-probability distribution that does not early-exit. The supplied code recommendation proposes 200–500 independent draws or a closed-form score-difference distribution, with batch size registered. This is a code/diagnostic repair, not a forced lever adjustment.
4. Expected metric/direction: Make the simulation’s win-rate/agreement signal meaningful; evaluate team-run MAE, total-run MAE, and winner-direction intervention separately over rolling slates. The current evidence does not quantify the expected numerical gain.
5. Scope and tension: The finding affects the simulation pipeline, not a particular listed game. The current slate has zero flip candidates and zero published flips. Do not force winner flips simply to make this slate look more active.

1. Persistent, non-parameter issue — run-model feature/target investigation
2. Finding target: training.xgb.n_estimators, training.xgb.max_depth, and training.xgb.learning_rate are listed, but the finding is explicitly param_actionable=false; the prescribed action is not a bounded nightly hyperparameter move.
3. Exact action: Isolate the shared feature/target contract and run a same-fingerprint training backtest. Test source-labelled no-vig market probability only when a real source exists, plus park/weather, bullpen rest/fatigue, and platoon-adjusted lineup quality as separately evaluable feature additions; compare holdout run-target R² and MAE and bump the model schema if semantics change. Do not combine these into an untracked production change.
4. Expected metric/direction: Improve holdout team-run and total-run MAE and explain more run variance; current R² values are near zero. No specific magnitude of improvement is established.
5. Scope and tension: Broad run-model scope; the run-weighting lever helps NYY@TB’s large team-run error but risks disturbing CHW@CLE’s smaller one. This is why the appropriate action is an isolated backtest, not a one-game weighting adjustment.

1. Single-slate driver — run-mean under/over-allocation; smallest shadow-only test
2. Registered lever from the tension entry: recent.* / season.* run-mean weighting.
3. Direction and magnitude: Test a 2% relative reduction in recent run-mean weight against the baseline in shadow, without changing season weight or publishing it. This is a small diagnostic perturbation, not a claim that the recent component is proven too high.
4. Expected metric/direction: Check team-run MAE and total-run MAE; improvement would require lower errors on rolling data, not merely a better NYY@TB reconstruction.
5. Scope and tension: NYY@TB is the listed beneficiary candidate because team runs were off by 5; CHW@CLE is the listed risk because team-run error was 2. The overall two-game player evidence points to projected lineups scoring below projection, but the cross-slate run-delta register remains on hold. Thus the net case is weak and this test must remain shadow-only.
6. Regression risk: Material. The lever’s stated help/hurt evidence is exactly one game versus one game. Reject the adjustment if the beneficiary improvement comes with worse rolling total/team-run error or consistent harm to games like CHW@CLE.

No production changes are recommended to pitcher form/prior games from this slate: NYY@TB’s Peralta result is characterized as an outlier, and the lever-tension entry says changing pitcher weights helps NYY@TB while risking CHW@CLE.

9. Bottom line

Known: Both published winners missed; both scores missed; calibration error on this two-game slate was very high; the run-model holdout R² values are near zero; NYY@TB has a concrete recency-cap structural error; simulation inertness and weak run-model signal have persisted long enough to be escalated.

Uncertain: The precise causal contribution of model weighting versus game variance; the reconciled CHW starter line; whether a particular run-mean adjustment improves future performance; and the full details of the 14 informational audit findings. The missing integrity revision audit also prevents verification of any claimed post-remediation proof.

Single safest next action: Keep this slate quarantined from parameter learning, and first have the operator investigate and repair the persistent simulation distribution/early-exit issue with an independent win-probability diagnostic. It is an 11-slate established failure, requires a code-level check rather than speculative tuning, and can be validated without attributing either game’s miss to the artifact defect.