Luna deep analysis — slate 20261007

Coverage receipt: COMPLETE · 9/8 sections; 4/4 games; 6/6 causes; 7/7 sources

1. Evidence receipt and gaps

All 7 source_manifest entries are recorded as loaded. The supplied evidence identifies each by path, purpose, and byte count:

• prediction review — /app/output/reports/prediction_review_20261007.json, 32,361 bytes; official outcomes and scoring errors.
• forensic deep audit — /app/output/reports/PREDICTION_SYSTEM_DEEP_AUDIT_20261007.json, 75,958 bytes; probability, provenance, starter, and decision contracts.
• continuous learning — /app/output/learning/continuous_learning_daily_20261007.json, 681,969 bytes; pipeline-stage quality, calibration, probability lineage, score distribution, and starter comparisons.
• detailed report — /app/output/reports/detailed_20261007.html, 108,248 bytes; game and simulation rendering.
• ultra detailed player audit — /app/output/reports/20261007_ultra_detailed_player_audit.html, 6,962,831 bytes; player-level prediction trace.
• structural audit — /app/output/reports/audit_report_20261007.html, 473,259 bytes; invariant findings.
• integrity revision audit — /app/output/reports/audit_report_20261007_integrity_revision.html, 1,646,215 bytes; audit of 14 post-start revision artifacts.

The manifest says the files were loaded; this analysis cannot independently verify their contents against the paths. All 4 games listed in the diagnosis are reviewed, matching games_reviewed: 4, expected_games: 4, and the coverage matrix’s 4/4 counts for structural audits, deep audits, continuous-learning games, canonical artifacts, and starter comparisons.

Gaps and internal inconsistencies to retain in the record:

• TB@NYY has no joined player-level actuals: the slate feed has 25 rows but none match that game ID. Its player-level causes therefore cannot be assessed from these data.
• The MIL@SD probability lineage says predicted_side: home and published pick probability 0.505, while the game record predicts MIL (away), the diagnostic says home probability 0.495, and the displayed home-minus-away edge is −0.01. These point to an away pick. This is an internal probability-side/label inconsistency.
• Several probability records contain distinct raw_model_pick_probability and governor_input_probability values. The trace records the governor inputs and named transformations, but the meaning and exact ordering of all “raw” fields are not fully explicit. This is a lineage-clarity issue; it does not, by itself, establish an unrecorded calibrated-to-published bypass.
• probability_reversions reports one blend flip at the blend stage, four sides surviving to publication, no overwrites, and an empty games list. The flip cannot be identified from that object.
• The aggregate structural-audit result is 0 critical, 0 error, 0 warning, 4 info; each game says PASS with 0 critical/error/warning and one informational note. Those summaries do not supply the details of the four aggregate informational items.

2. Executive diagnosis

Winner accuracy

The slate went 3/4 (75%), with one miss, TB@NYY. This exceeds the supplied walk-forward accuracy of 0.570149 and historical winner accuracy of 0.600414, but four games do not establish improved performance. The cross-slate winner-accuracy trend is explicitly flat, with mean 0.7381 over the reported series.

A correct winner is not evidence that the score or confidence was good. Three games had the correct winner but substantial run errors.

Calibration quality

The four-game confidence bucket reports average confidence 0.5228, hit rate 0.75, and signed bucket calibration error −0.2272, labeled underconfident under the stated convention (published confidence minus observed hit rate). Its weighted absolute calibration error is 0.2272. That is a large single-slate mismatch, but the bucket contains only four games and the causal register explicitly says not to change calibration on this slate alone.

The reported published pick-side confidence values average approximately 0.528; the bucket’s 0.5228 matches the average of the displayed home probabilities (including probabilities below 0.5), not the average pick-side confidence. The signed-error definition is stated, but the confidence convention underlying that bucket is not fully clear. The conclusion is still poor slate-level alignment, not a reliable direction for a global curve.

Score accuracy

Published team-run MAE is 2.375 and total-run MAE is 3.75 across four games. Team-run absolute errors are 5, 5, 7, and 2. Predicted margins are compressed: 1, 2, 1, and 1 runs, versus actual margins of 6, 3, 2, and 1. The holdout run-model R² values are near zero: home 0.002143, away 0.030936, total 0.002415. Exact score outputs therefore have weak demonstrated signal.

Starter accuracy

All four games have starter comparisons in starter_lines, with identity matches true. The starter outcomes were mixed: CLE’s Daniel Espino allowed one earned run in two outs against a 3-ER projection, while TB’s Max Fried allowed four ER against a 2-ER projection. These comparisons do not show a uniform slate-wide starter miss, and the conspicuous Espino result is identified as an outlier/variance event. The diagnosis does not support tuning starter weights from this slate alone.

Artifact integrity

No supplied game structural audit has critical, error, or warning findings; all four are PASS, and the game diagnostic says the forecast file was clean. Coverage is complete for the four games. These facts argue against a saved-score artifact causing the TB miss. They do not resolve the internal MIL@SD probability-side contradiction or the missing TB player join. The latter are evidence/contract-quality issues to address in reporting and trace validation.

Causal issue register: reconciliation of every entry

• High — pipeline_winner_instability: historical_ml → variance_applied reduced winner accuracy by 0.2500 (one hit); affected game MIL@SD. Keep and expose the exact before/after game picks. Do not treat a later recovery as harmless.
• High — single_slate_calibration_instability: weighted absolute calibration error 0.2272 over four games; affects all four. Watch the same buckets over the rolling window before changing calibration. Do not infer a stable calibration curve from this slate.
• High — weak_run_model_signal: holdout home/away/total R² 0.0021/0.0309/0.0024; affects all four. Treat exact scores as low-confidence and investigate features/targets; do not infer that positive but tiny R² makes exact scores reliable.
• Info — complete_slate_audit_coverage: 4/4 structurally audited; all four. Keep the audit receipt. Coverage does not cancel in-game findings.
• Info — probability_publication_delta_explained: one material movement, with recorded score-coherence and/or recency transformations; TB@NYY is listed. Keep every probability layer visible. A published value above a calibrated value is not a bypass when the movement is recorded.
• Info — simulation_inert_on_slate: zero flip candidates and zero published flips; all four. Measure simulation contribution over rolling slates. Zero flips alone is not a defect and does not mean score reconciliation did no work.

No critical or error item appears in the causal register. The separate internal probability-label inconsistency is an evidence/contract issue discussed in section 6.

3. Every game

“Home-minus-away edge” below is the supplied signed system edge, not a probability distance from 50%. A 0.12 signed home-away edge is moderate, not a near coin flip. Runs and starter lines are the supplied projections and outcomes; no missing player results are inferred.

CLE@CHW

• Predicted / actual: CLE 5, CHW 4, predicted CLE; actual CLE 9, CHW 3, actual CLE. Correct winner, wrong score.
• Published confidence / edge / bucket: Published pick probability 0.535129 for away CLE; published home probability 0.465 in the game diagnostic; home-minus-away edge −0.0703, displayed system edge about 0.07; lean.
• Errors: Brier 0.216105; team-run absolute error 5 (CLE off by 4, CHW by 1); total-run bias −3 (9 projected versus 12 actual).
• Starter projected → actual: Daniel Espino, CLE: 3 ER, 12 outs, 6.75 ERA projected; 1 ER, 2 outs, 13.50 ERA actual. Sean Newcomb, CHW: 2 ER, 10 outs, 5.40 ERA projected; 0 ER, 6 outs, 0.00 ERA actual. Both identities match. Espino’s short outing makes the ERA rate an outlier; it does not mean CLE’s offense failed.
• Pregame edges: Season +0.1605, recent +0.030, pitcher −0.008, composite −0.1018 (home-minus-away). Season-quality adjustment was 0.0 because the quality edge was below threshold; pitcher run factors were 0.9914 for CLE runs off the home pitcher and 0.9834 for CHW runs off the away pitcher.
• Confidence/recency controls: Raw pick probability in the trace 0.555841; governor applied, governed probability 0.5486, calibration sample 464, empirical hit rate 0.5819; recency guardrail applied, no numeric discount factor supplied; starter-experience shrinkage not applied. Lineage separately lists raw model pick probability 0.5846.
• Score path and adjustments: Final forecast was 5–4 CLE; volatility action no_change; Savant adjustment applied. The supplied stage table does not identify a CLE-specific winner flip.
• Player evidence: Jo Adell 2-for-4, 3 RBI, run impact +1.42; Patrick Bailey 2-for-4, +0.80; Kyle Teel 1-for-4, 1 RBI, +0.78. These are among the deviations the audit labels supported by above-average contact quality. Other listed contributors include Murakami (−1.25) and Grichuk (−0.93); do not treat those individual results as repeatable parameters.
• Audit and causes: Structural audit PASS, no listed findings; forecast-file integrity reported clean. The game was a correct-winner/wrong-score result, principally a run/score miss. The starter feed identifies Espino’s poor rate line as variance; the official starter comparison also shows he recorded only two outs. Primary/secondary driver fields are null, so no formal driver label is assigned.

LAD@ATL

• Predicted / actual: LAD 6, ATL 4, predicted LAD; actual LAD 4, ATL 1, actual LAD. Correct winner, wrong score.
• Published confidence / edge / bucket: Published pick probability 0.507177 for away LAD; published home probability about 0.493; home-minus-away edge −0.0144, displayed edge about 0.01; toss_up.
• Errors: Brier 0.242875; team-run absolute error 5 (LAD by 2, ATL by 3); total-run bias +5 (10 projected versus 5 actual).
• Starter projected → actual: Tyler Glasnow, LAD: 2 ER, 16 outs, 3.38 ERA projected; 1 ER, 14 outs, 1.93 ERA actual. Tyler Mahle, ATL: 3 ER, 15 outs, 5.40 ERA projected; 1 ER, 19 outs, 1.42 ERA actual. Both starters performed better, by ERA, than projected.
• Pregame edges: Season −0.6096, recent −0.314, pitcher +0.1387, composite −0.4686 (home-minus-away). Season-quality adjustment moved the score toward LAD by −0.256; pitcher run factors were 0.9252 (LAD) and 1.0639 (ATL).
• Confidence/recency controls: Trace raw pick probability/governor input 0.535236; governed 0.5084, calibration sample 612, empirical hit rate 0.5474; recency guardrail applied, numeric factor absent; starter-experience shrinkage not applied. Lineage’s separate raw-model field is 0.5097.
• Score path and adjustments: Published score 6–4 LAD; large_savant_run_environment_swing diagnostic flag; Savant adjustment applied; volatility action no_change. The final run total remained five too high.
• Player evidence: No player-level breakdown supplied for this game.
• Audit and causes: Structural audit PASS; forecast-file integrity reported clean. The winner was right, but both starters beat their ERA projections and the total was over-projected by five. The supplied record does not assign a formal primary/secondary cause.

MIL@SD

• Predicted / actual: MIL 6, SD 5, predicted MIL; actual MIL 3, SD 1, actual MIL. Correct winner, wrong score.
• Published confidence / edge / bucket: Game diagnostic gives home probability 0.495, away pick probability 0.505, home-minus-away edge −0.01; toss_up. Probability lineage instead says predicted side home, published pick probability 0.505. This is the noted side-label contradiction.
• Errors: Brier 0.245025; team-run absolute error 7 (MIL by 3, SD by 4); total-run bias +7 (11 projected versus 4 actual).
• Starter projected → actual: Robert Gasser, MIL: 2 ER, 11 outs, 4.91 ERA projected; 1 ER, 12 outs, 2.25 ERA actual. Walker Buehler, SD: 3 ER, 11 outs, 7.36 ERA projected; 1 ER, 9 outs, 3.00 ERA actual. Both performed better by ERA than projected.
• Pregame edges: Season −0.723, recent +0.334, pitcher −0.0447, composite −0.2325 (home-minus-away). Season-quality margin adjustment −0.3037; run factors 1.0173 for MIL and 0.9726 for SD.
• Confidence/recency controls: Trace raw pick/governor input 0.524296; governed 0.505, sample 567, empirical hit rate 0.5485; no recency guardrail; starter-experience shrinkage not applied. Lineage separately lists raw-model probability 0.5.
• Score path and adjustments: Published score 6–5 MIL; Savant adjustment applied; volatility action no_change. Pipeline stages show this is the one game with a winner-direction regression at variance application and later recovery; details appear in section 4.
• Player evidence: No player-level breakdown supplied.
• Audit and causes: Structural audit PASS; forecast-file integrity reported clean. The three-run total over-projection is consistent with the starter comparisons and the supplied run factors, but there are no player actuals here to allocate the offense/bullpen portions. Formal primary/secondary cause fields are null.

TB@NYY

• Predicted / actual: TB 3, NYY 4, predicted NYY; actual TB 4, NYY 3, actual TB. Only winner miss.
• Published confidence / edge / bucket: Published home NYY probability 0.544016, home-minus-away edge +0.088 (displayed edge about 0.09); lean. This is a meaningful moderate home edge, not a near-coin-flip edge.
• Errors: Brier 0.295953; team-run absolute error 2 (each side off by 1); total-run bias 0 (7 projected, 7 actual). The total was right while the run allocation and winner were wrong.
• Starter projected → actual: Nick Martinez, TB: 2 ER, 13 outs, 4.15 ERA projected; 2 ER, 11 outs, 4.91 ERA actual. Max Fried, NYY: 2 ER, 16 outs, 3.38 ERA projected; 4 ER, 17 outs, 6.35 ERA actual. Both identities match; Fried’s ER were two above projection, while Martinez matched projected ER.
• Pregame edges: Season +0.1757, recent +0.182, pitcher −0.0142, composite +0.2909 (home-minus-away). The net pregame signals favored NYY despite the near-neutral pitcher edge. Season-quality adjustment was 0.0; run factors were 0.9797 (TB) and 0.9655 (NYY).
• Confidence/recency controls: Trace raw pick/governor input 0.521431; governed/calibrated 0.5665, sample 393, empirical hit rate 0.5674; recency guardrail applied, no numeric factor supplied; starter-experience shrinkage not applied. Lineage separately gives raw-model probability 0.6185. Score-coherence probability is 0.566481 and published probability 0.544016; the recorded transformation is score coherence followed by matchup-recency guardrail.
• Score path and adjustments: Published score 3–4 NYY; large_savant_run_environment_swing flag; Savant adjustment applied; volatility action no_change. Prediction put one run on the wrong side.
• Player evidence: Not available because none of the 25 feed rows joined to this game ID. No player result should be inferred.
• Audit and causes: Structural audit PASS, forecast-file integrity reported clean. Recorded primary driver is mixed (“Several errors of similar size”), no secondary driver. Fried’s additional ER are a plausible starter contribution, but the data do not isolate offense, bullpen, or individual-player causes. The score total itself was not the problem; pregame side selection and run allocation were.

4. Miss reconstruction

TB@NYY: earliest wrong-direction evidence and attribution

The earliest explicitly supplied pregame evidence already favored the result that lost: season edge +0.1757, recent edge +0.182, and composite home edge +0.2909 all favored NYY; pitcher edge was only −0.0142 for the home side. The confidence trace then reports governed probability 0.5665 for the home pick, and the published home probability remains 0.544016. The system therefore did not merely miss on a nominal 0.50 selection: its published lean was for NYY at a 0.088 home-away edge.

The probability fields are not fully ordered transparently: the trace governor input is 0.521431, but the governed output is 0.5665; the separate lineage raw-model field is 0.6185. The named calibration/governor, score-coherence, and recency stages are recorded, so the supplied evidence does not justify calling the calibrated-to-published movement a bypass. It does justify checking the mapping and semantics of the multiple raw fields.

Attribution by category:

• Bad input: No specific faulty game input is established. Starter identities match. A player-actual join is missing, so player-level diagnostics are incomplete—not evidence that the model’s pregame inputs were wrong.
• Bad weighting: Season and recent evidence both pointed home, while the pitcher edge was nearly neutral and away-leaning. The miss is consistent with a side-weighting error, but one game does not establish which edge should have been weighted differently.
• Calibration: This was a lean with published home probability 0.544, not a high-conviction pick. It lost, but the four-game bucket was underconfident in aggregate because three of four picks hit. Calibration is a persistent cross-slate concern, not a proven TB-specific cause.
• Starter error: Fried allowed 4 ER versus 2 projected. Martinez allowed the projected 2 ER. This supports a home-starter contribution; it does not explain the entire loss by itself.
• Offense/bullpen miss: Player actuals are absent. The team-score split was off by one run on each side, but no offensive or bullpen attribution is available.
• Run-environment miss: Total runs were exactly right (7), so a total-level environment miss is not demonstrated for TB. The allocation across teams was wrong.
• Irreducible variance: Still plausible, especially for the starter/side outcome, but the supplied evidence cannot quantify it.

Pipeline-stage quality across all stages and transitions

All stages are measured against official final scores; positive deltas versus historical_ml are regressions.

The historical → variance transition regressed winner accuracy by one hit (−0.25) on MIL@SD; its team-run MAE improved only 0.0148. The next stage changed MIL from tie to SD without restoring the correct direction. The embedded-Savant-to-fundamental-review transition later flipped MIL from SD to MIL and restored the slate to 3/4. That recovery does not erase the intermediate regression. From fundamental review onward, the score and winner did not change through publication; simulation produced no winner intervention.

Team-run MAE ended 0.4852 worse than historical ML, despite margin MAE improving 0.4245. Total-run MAE finished essentially unchanged from the historical baseline (3.75 versus 3.7594). This is not a consistent score-quality improvement.

5. Cross-game counterfactuals

Every supplied lever_tension entry:

1. score.variance_multiplier / score.variance_floor (widen run distributions), targets Brier/calibration
2. Helps: none identified.
3. Hurts: none identified.
4. Evidence: 0 helped vs 0 hurt. There were no high-conviction games and no high-raw cluster. This slate does not demonstrate that widening distributions would improve particular games. A bounded shadow test is defensible only because of the persistent calibration pattern, not because this entry shows a per-game win.

1. player.pitcher.v4 form weights / player.pitcher_prior_games, targets starter ER/outs MAE
2. Helps: none identified.
3. Hurts: none identified.
4. Evidence: 0 helped vs 0 hurt. CLE’s Espino outlier and TB’s Fried miss pull in different directions; neither establishes a systematic starter-weight change.

1. recent.* / season.* run-mean weighting
2. Helps: CLE@CHW (team runs off by 5), LAD@ATL (off by 5), MIL@SD (off by 7).
3. Hurts: TB@NYY (team runs already close, off by 2; reweighting risks disturbing it).
4. Evidence: 3 helped vs 1 hurt on run error. This supports testing the run-mean weighting against run-error metrics, but not a claimed winner-accuracy improvement or an unqualified live change; the data do not specify a direction or which sub-weight to alter.

Tempting one-game fixes to reject:

• Do not change pitcher weights solely to prevent a repeat of Espino’s two-out outing or Fried’s four ER. The observed starter results are mixed, and the cross-game pitcher tension shows no demonstrated benefit.
• Do not increase TB’s home-confidence adjustment because its total was correct; the home side lost and the player feed is missing.
• Do not use CLE’s Adell or Teel game as a one-game hitter-weight calibration. The audit says their deviations have supporting contact quality, but one slate and a 15-hitter sample do not establish a stable adjustment.
• Do not interpret the “3 helped/1 hurt” run-mean tension as evidence to change winner selection; its direct evidence is team-run error.

6. Data and contract defects

Audit findings and severity counts

• Per-game structural audits: all four games have status PASS, risk score 0, and no listed findings. Each reports 0 critical, 0 error, 0 warning, 108 good, and 1 info.
• Aggregate structural audit summary: 0 critical, 0 error, 0 info, 0 warning in its separate summary object, whereas the per-game summaries sum to four informational notes. This is a reporting-summary discrepancy, not proof of an invariant failure.
• Coverage matrix: all reported structural/deep/learning/starter/canonical counts are 4/4 and all_games_structurally_audited: true. It is appropriate to say game-level structural coverage is complete, but not that every diagnostic is complete: TB player-level actuals do not join.
• Source-integrity statements: all four game diagnostics say forecast-file integrity is clean; the integrity revision audit says 14 revision artifacts were audited, but no finding detail is supplied.
• Causal register findings: the three high and three info items are reproduced in section 2; none is a structural critical/error finding.

Specific contract and data issues

1. MIL probability-side label inconsistency: lineage says home side/pick 0.505; game score, displayed edge, and diagnostic home probability imply away MIL is the pick at 0.505. Treat as an evidence/contract error requiring label validation. The contradictory label should not be silently used in calibration summaries.
2. Probability-field semantics: lineage’s separate raw-model probability differs from the governor input/trace raw pick value for all four games. The named stages explain the calibrated/governed and calibrated-to-published changes at a high level, but the raw-field definitions and transition ordering should be explicit.
3. probability_reversions incomplete event identification: one blend-stage flip is reported, but games is empty. No game can be named from this field, so it is not possible to check that flip against the per-game trace.
4. TB player join gap: no per-player actuals joined for TB@NYY. This is a forensic coverage gap for player causality, not a demonstrated score artifact defect.
5. Calibration bucket convention: the reported average 0.5228 matches the mean of the displayed home probabilities, while the prediction outcomes are pick-side outcomes. Because the bucket’s confidence convention is not fully described and MIL is mislabeled, its summary should be interpreted cautiously.

Complete pick-side probability lineage

Values below preserve the supplied lineage. “Governor input” corresponds to the trace’s raw pick probability; the separate raw-model field is also shown so the distinction is visible.

All four lineage entries say unexplained_material_delta: false. Calibrated-to-published movements are recorded as score-coherence and/or recency transformations; the TB material movement is therefore not called a bypass. The audit register specifically says TB’s published value being above a calibrated one is not a governor bypass when the recorded transformations explain it. The raw-model/governor-input distinction and the absent numeric recency factors remain trace-clarity issues, not established unrecorded bypasses.

7. Cross-slate persistence

Cross-slate findings and escalation status

The report includes 10 prior slates and an 11-slate trend window. Finding streaks:

For every escalated finding:

• simulation_inert_on_slate — 11 consecutive, 11 incidences; targets simulation.allow_score; metrics team-run MAE and total-run MAE. Direction: ensure the simulation module is actually producing an independent, informative distribution rather than a silent no-op. The self-assessment identifies the shared target-search draws and early exit as the root cause; fix/diagnose the simulation path, not a forced parameter change.
• single_slate_calibration_instability — 5 consecutive, 8 incidences; targets confidence.empirical_max_weight, confidence.base_cap, confidence.moderate_cap; metrics Brier and calibration error. Direction: increase empirical shrinkage via a bounded shadow experiment, then judge on rolling Brier/calibration.
• weak_run_model_signal — 5 consecutive, 9 incidences; lists training.xgb.n_estimators, training.xgb.max_depth, and training.xgb.learning_rate, but is marked not parameter-actionable. The self-assessment says no nightly knob can supply missing run signal; investigate features/target with an isolated training backtest. Do not turn those listed training controls into an unsupported tuning change.

The self-assessment separately records run_model_near_zero_signal, with total/home/away R² 0.002415/0.002143/0.030936, as an operator action. This agrees with the escalated weak-run-signal finding rather than adding a reason for a nightly parameter adjustment.

Structural cross-slate evidence

• Margin distribution: actual margins were 1, 2, 3, 6; predicted margins were 1, 1, 1, 2. All four predicted margins were ≤2; one actual margin was ≥4. CLE@CHW had signed margin error −5 in that block; other listed errors were −1, −1, 0. This is evidence of slate-level compression, but score_margin_compression is not current and has a zero consecutive streak.
• Calibration monitoring: four settled games; confidence bucket average 0.5228, observed hit rate 0.75, weighted absolute calibration error 0.2272. This is current-slate evidence within a five-slate escalation.
• High-raw cluster: threshold 0.65, but 0 games, 0 hits. There is no high-raw-cluster performance evidence on this slate.
• Simulation control: four observed games, 0 flip candidates, 0 published flips, priority P2, diagnostic_only_without_independent_support. This aligns with the 11-slate escalation.
• Pipeline transitions: the historical-to-variance MIL regression is current. Later recovery does not remove that intermediate-direction defect.
• Validation: 2,415 historical-temporal rows; winner accuracy 0.600414, walk-forward accuracy 0.570149, Brier 0.239378, walk-forward Brier 0.242204. These are broader baselines, not evidence that run scores are reliable.

Metric trends and top-confidence-pick streak

The trend directions are explicitly:

• Margin bias: regressing, mean −1.619.
• Mean team-run error on losses: improving, mean 2.86.
• Weighted calibration error: improving, mean 0.3854; despite the direction, its current value is 0.2272 and the finding has escalated.
• Winner accuracy: flat, mean 0.7381.

The top-confidence-pick loss streak is 1: the 20261005 top pick missed, 20261006 hit, and 20261007 missed. The current slate’s top-confidence pick is therefore a loss but does not extend a consecutive loss streak from 20261005 because a hit intervened. Two dates in the supplied sequence are blank/unknown and should not be treated as results.

8. Adjustments for today

These are bounded proposals for shadow evaluation, not live changes. Persistence is ranked ahead of single-slate evidence. No promotion is recommended from this diagnosis alone.

1. Persistent calibration instability — shadow confidence.empirical_max_weight upward by +0.02 absolute.
2. Reason/rank: five consecutive slates, 8 incidences; strongest parameter-actionable persistent pattern. The registered direction is increased empirical shrinkage.
3. Expected metrics: lower rolling Brier and lower absolute calibration error; evaluate by confidence bucket and overall, not this four-game result alone.
4. Games/tension: no named confidence lever tension lists game-level help/harm. TB@NYY is the current miss with a lean home probability of 0.544 and should be monitored for overconfidence regressions; three of four picks hit, so do not frame this as evidence to indiscriminately suppress all confidence.
5. Risk/gate: small shadow-only test with rolling calibration and Brier rollback monitoring. Do not infer that the change will fix TB specifically.

1. Persistent simulation inertness — code/diagnostic action, not a parameter move.
2. Registered target: simulation.allow_score is listed, but the finding is marked param_actionable: false; the self-assessment gives a code-level root cause.
3. Action: inspect the simulation path so score-target realization and win-probability estimation use an independent, unsteered distribution that does not early-exit; record its contribution before/after. The supplied action describes an independent batch or closed-form run-difference distribution and a registered batch-size control.
4. Expected metrics: first establish a non-vacuous distribution and measure team-run/total-run MAE contribution across rolling slates; do not promise a winner-accuracy gain from this slate.
5. Games/tension: zero flip candidates across the four games; no per-game help or harm is identified in lever tension.
6. Risk/rank: persistent 11-slate evidence warrants priority diagnostic/code work. Do not force simulation.allow_score as a tuning change in place of fixing the no-op behavior.

1. Persistent weak run signal — isolated feature/target backtest, not a nightly parameter change.
2. Registered levers listed: training.xgb.n_estimators, training.xgb.max_depth, training.xgb.learning_rate; finding is explicitly not parameter-actionable. The supplied action calls for a manual isolated training backtest.
3. Action: test the shared feature contract and target with the supplied candidate feature families—source-labelled no-vig market probabilities when available, park/weather, bullpen rest/fatigue, and platoon-adjusted lineup quality—then compare same-fingerprint holdout R²; bump the model schema version if semantics change.
4. Expected metrics: improve holdout run-target R² and out-of-sample team-/total-run error without degrading winner or calibration measures.
5. Games/tension: CLE, LAD, and MIL have the largest team-run errors and are listed as helped by run-mean reweighting; TB is the close-score harm case. The persistent code/feature diagnosis applies slate-wide; no single listed feature is proven by these four games.
6. Risk/rank: high scope and operator effort, but the persistent near-zero R² is stronger evidence than any one-game score adjustment. No XGBoost knob-only shadow is recommended absent an isolated feature/target comparison.

1. Strongest single-slate structural driver — shadow score.variance_multiplier upward by 2% relative to its current value.
2. Reason/rank: every predicted margin was at most two runs while one actual margin was six; this is a concrete current-slate compression signal. It is weaker than the persistent findings, and the associated cross-slate streak is currently zero.
3. Expected metrics: test whether score-distribution calibration/Brier improves; do not expect a direct improvement in run-mean MAE, since widening variance does not correct a biased mean.
4. Games/tension: the variance lever’s tension record identifies no helped and no hurt games, so no game-specific benefit is claimed. Monitor the close TB@NYY score and the three correct-winner/wrong-score games for regressions.
5. Risk/gate: shadow only; roll back if rolling Brier/calibration or score-distribution checks deteriorate. The supplied evidence does not support a live change or a larger adjustment.

No recent.* / season.* run-mean change is recommended today: its slate tension is 3 helped versus 1 hurt, but the entry does not specify the direction or sub-weight, and its clearest evidence concerns run errors rather than winners. No pitcher-weight change is recommended from this mixed slate.

9. Bottom line

Known: the slate finished 3/4, with three correct winners but materially inaccurate scores; TB@NYY was the only winner miss, despite a moderate 0.088 published home edge. The pipeline temporarily moved MIL in the wrong winner direction at variance application before later recovering. Run-model holdout signal is near zero, calibration instability and simulation inertness are persistent escalations, and the saved forecast artifacts passed the reported game-level structural checks.

Uncertain: the precise TB offensive/bullpen/player contribution, the causal source of the wrong team-run allocation, the interpretation of every raw probability field, and whether a variance change would improve outcomes. TB player actuals are missing; MIL probability-side lineage is internally inconsistent.

Safest next action for the very next slate: preserve the full probability and stage trace, add a contract check that rejects side/probability-label contradictions such as MIL@SD, and keep any calibration change in a bounded shadow experiment with rolling Brier/calibration monitoring. This addresses a concrete evidence-integrity risk without pretending that one slate identifies a winning model adjustment.