Luna deep analysis — slate 20260924

Coverage receipt: COMPLETE · 9/8 sections; 12/12 games; 6/6 causes; 6/6 sources

1. Evidence receipt and gaps

Sources received

All seven manifest entries were checked against the supplied evidence:

The missing integrity-revision audit is a provenance gap, but it did not prevent the supplied per-game structural checks from running. It does mean immutable revision/remediation claims cannot be independently confirmed from this receipt.

Games reviewed

All 12 games in the slate were reviewed, and the coverage matrix reports complete game-level coverage:

• Expected games: 12
• Canonical artifacts loaded: 12
• Continuous-learning games: 12
• Forensic deep-audit games: 12
• Starter comparisons: 12
• Structural audits: 12
• Games structurally audited: 12

The 12 games were:

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

Internal gaps and contradictions

1. Missing integrity revision artifact. The manifest explicitly marks the integrity revision audit as missing.
2. Pipeline trace contradiction for later winner direction. The stored transitions say that after enhanced_prediction, no later stage changed direction, but at least CIN@ATL is shown with an enhanced-stage score favoring CIN while the published prediction favors ATL. The probability trace also reports a published ATL probability of 0.560. This requires reconciliation in the raw trace; the supplied evidence does not fully establish whether the later winner changed, whether score and probability paths diverged, or whether the transition summary omitted a change.
3. Stage summary versus per-game result ambiguity. historical_ml has 9/12 winner hits, variance_applied 6/12, and final stages 8/12. The aggregate transition record correctly identifies a three-game regression at variance application, but it does not list all winner states for every later stage.
4. Calibration labels are internally consistent with the stated definition. The definition is published confidence minus observed hit rate. Thus 0.5248 − 0.75 = −0.2252 is underconfident, while 0.56 − 0.50 = +0.06 is overconfident.
5. No market cross-check exists. No source-labelled moneyline was available for any game, so market agreement/disagreement cannot be used diagnostically.
6. Starter identity coverage was complete, but not every per-player feed line was complete. For example, projected starter lines were missing for Eduardo Rodriguez and David Sandlin in the settled per-player feed, despite game-level starter comparisons being present.

2. Executive diagnosis

Overall result

• Winner accuracy: 8/12, 66.67%
• Misses: 4/12
• Mean Brier score: approximately 0.240
• Published weighted absolute calibration error: 0.1701
• Historical validation winner accuracy: 0.592357
• Walk-forward accuracy: 0.57551
• Walk-forward Brier: 0.241034
• Holdout run-model \(R²\): away 0.034855, home 0.003987, total 0.003064
• Final-stage team-run MAE: 2.2083
• Final-stage total-run MAE: 3.75
• Final-stage margin MAE: 2.0833
• Final-stage winner accuracy: 8/12

The slate beat the stated walk-forward baseline and was directionally successful, but the score forecasts were weak, the probability buckets were poorly aligned, and the system had a material intermediate winner-direction regression.

Separate diagnoses

Winner accuracy

The final published pick was correct in 8 of 12 games. This is useful directional performance, but it does not imply that the numerical prediction was good.

The four losses were:

• CIN@ATL
• LAA@SEA
• NYM@TEX
• SD@LAD

Two losses were classified as wrong-side conviction calls, one as a run-environment miss, and one as toss-up variance:

• Wrong conviction side: 2
• Run environment overestimated: 1
• Genuine toss-up variance: 1

Calibration quality

Calibration was the most important winner-probability weakness on this slate.

• 0.50–0.55 bucket: average confidence 0.5248, observed hit rate 0.75, eight games; the model was underconfident in this bucket.
• 0.55–0.60 bucket: average confidence 0.56, observed hit rate 0.50, four games; the model was overconfident in this bucket.
• Weighted absolute calibration error: 0.1701 across 12 settled games.
• High-conviction miss rate: 0.5, with 2 misses among 4 high-conviction games.
• The high-raw-probability cluster had 2 hits and 1 miss; the current slate did not meet the persistent high-raw-cluster streak condition.

The correct response is not to treat all probabilities as unusable. The evidence supports bounded confidence-cap and variance shadow experiments, not a wholesale calibration rewrite based on one slate.

Score accuracy

Score accuracy was materially weaker than winner accuracy.

The model predicted every margin at two runs or less:

• Predicted margins ≥ 3: 0/12
• Actual margins ≥ 4: 3/12
• Actual margins: four one-run games, five two-run games, two four-run games, and one eight-run game.
• Signed margin errors included underestimates of 3 runs in ARI@COL, 6 in CHW@KC, and 3 in MIL@PHI.

The final-stage run model has near-zero holdout explanatory power. Exact scores should therefore remain low-confidence outputs. The model correctly identified several winners while badly missing the score magnitude.

Starter accuracy

Starter forecasts were mixed:

• Several losses were caused by outlier starter performance: Tanner Gordon, Peter Lambert, Kumar Rocker, and Paul Skenes’ unexpectedly strong outing.
• ARI@COL and CHW@KC had very large starter errors, but the winner was still correct.
• CIN@ATL had both starters only one earned run from projection, so the loss was not primarily a starter-model failure.
• LAA@SEA had both starters close enough to projection that the loss was better classified as variance or wrong-side conviction than starter error.
• NYM@TEX was affected by Kumar Rocker allowing 1 ER versus a projected 4 ER, but the stored primary diagnosis correctly emphasizes the lower-than-projected run environment.
• SD@LAD had Nick Pivetta outperforming projection and the game remained a toss-up.

The cross-slate starter ERA finding remains nonpersistent, so a strong starter-weight change is not justified solely by this slate.

Artifact integrity

Artifact integrity was strong for the game artifacts:

• All 12 structural audits passed.
• Structural audit summary: critical 0, error 0, warning 0, info 91.
• Each per-game audit had zero critical/error/warning findings.
• Forecast-file integrity was reported clean for every game.

However, the missing integrity-revision audit remains a receipt gap. It is a presentation/provenance gap, not evidence that the picks themselves were corrupted.

Causal issue register reconciliation

pipeline_winner_instability — High

• Evidence: historical_ml → variance_applied reduced winner accuracy by 0.2500, or 3 hits.
• Affected games: CHW@KC, MIA@CHC, NYM@TEX.
• Action: list exact pre- and post-stage winner states and retain both in the trace.
• Guardrail: later recovery does not make the unstable step harmless.

This was a real model/process risk. The variance stage changed three directions to ties, and the subsequent enhanced stage recovered two correctly and moved one incorrectly. The intermediate regression must remain visible even where the final result recovered.

single_slate_calibration_instability — High

• Evidence: weighted absolute calibration error 0.1701.
• Affected games: all 12.
• Action: monitor the same buckets over the rolling window before changing calibration weights.
• Guardrail: a 12-game slate cannot establish a stable calibration curve.

This is a real current-slate warning, but not sufficient by itself for a permanent calibration rewrite.

weak_run_model_signal — High

• Evidence: home/away/total holdout \(R²\) 0.0040/0.0349/0.0031.
• Affected games: all 12.
• Action: improve features/target and run an isolated training backtest.
• Guardrail: small positive \(R²\) does not make exact scores reliable.

This is both a current-slate and persistent cross-slate model risk.

complete_slate_audit_coverage — Info

• Evidence: 12/12 games structurally audited.
• Guardrail: coverage does not cancel findings inside audited games.
• Action: retain the per-game audit receipt.

Coverage is complete; the slate is not therefore “clean” in predictive terms.

probability_publication_delta_explained — Info

• Evidence: five material movements were attributed to score-coherence and/or recency transformations; no unexplained material delta was reported.
• Guardrail: a published probability above calibrated probability is not a governor bypass when a named transformation explains it.
• Action: continue exposing every probability layer.

This supports artifact lineage integrity for the five material movements.

simulation_inert_on_slate — Info on this slate, escalated persistent failure across slates

• Evidence this slate: zero flip candidates and zero published flips.
• Guardrail: zero winner flips is not automatically a defect.
• Action: measure simulation contribution over rolling slates.

Cross-slate evidence now elevates this from a single-slate observation to an established code/diagnostic issue.

3. Every game

The published probability below is the probability of the published side. A home-minus-away edge of 0.12 means a 56% home / 44% away split when the home side was favored, or 44% home / 56% away when the away side was favored. It is moderate, not near certainty and not a near coin flip.

ARI@COL — correct winner, wrong score

• Predicted: ARI 6, COL 5; ARI winner
• Actual: ARI 12, COL 8; ARI winner
• Published pick probability: ARI 0.56
• Confidence edge: −0.12 home-minus-away, moderate
• Bucket: moderate
• Brier: 0.1936
• Team-run absolute error: 9
• Total-run bias: −9; total was projected 11 versus actual 20
• Predicted margin: ARI by 1; actual margin: ARI by 4

Edges:

• Season home-minus-away: −1.1569
• Recent: −0.666
• Pitcher: −0.1533
• Composite: −1.3201

Starters:

• Eduardo Rodriguez: projected 3 ER/17 outs, actual 2 ER/18 outs; identity matched. The detailed player feed lacked a settled projected line for him in one matchup section.
• Tanner Gordon: projected 3 ER/13 outs, actual 10 ER/10 outs; ERA 27.00 versus projected 6.23, ER error +7.

Controls and score path:

• Governor reached 0.661 before score-coherence and recency controls; published home probability was 0.440, therefore ARI 0.560.
• Savant adjustment applied; large run-environment swing flag.
• Score path: historical model approximately ARI 4 / COL 5, variance stage 5/5, later score target 6/5.
• Volatility action: supported blowout-margin expansion.

Player adjustments:

• Cole Carrigg: +2.03 run impact versus projection.
• Geraldo Perdomo: +1.56.
• Pavin Smith: +1.11.
• Kyle Karros: +1.07.
• Nolan Arenado: +0.94.

Audit and causes:

• Structural audit: PASS; 0 critical/error/warning, 8 info.
• Primary cause: none assigned because the winner was correct.
• Secondary cause: none.
• Forensic diagnosis: right winner, wrong score. Tanner Gordon’s severe home/venue underperformance and under-projected offense drove the score error. The evidence supports a shadow test of context.venue_adjustment_weight, not a claim that this one start proves the adjustment is generally valid.

CHW@KC — correct winner, wrong score

• Predicted: CHW 7, KC 5; CHW winner
• Actual: CHW 9, KC 1; CHW winner
• Published pick probability: CHW 0.54462
• Confidence edge: −0.0892, lean
• Bucket: lean
• Brier: 0.207371
• Team-run absolute error: 6
• Total-run bias: +2
• Predicted margin: CHW by 2; actual margin: CHW by 8

Edges:

• Season: −0.9229
• Recent: −0.738
• Pitcher: −0.0461
• Composite: −1.1535

Starters:

• David Sandlin: projected 3 ER/15 outs, actual 1 ER/18 outs; detailed per-player feed lacked a settled projected line.
• Randy Dobnak: projected 2 ER/17 outs, actual 4 ER/4 outs; ERA 27.00 versus 3.18, ER error +2, outs error −13.

Controls and path:

• Experience control applied and shrank confidence toward 0.5; winner unchanged.
• Published probability followed the governed value after score/recency controls.
• Score path increased toward a CHW advantage and supported blowout-margin expansion.

Player adjustments:

• Carter Jensen: −1.73.
• Salvador Perez: +1.26.
• Vinnie Pasquantino: −0.80.
• Chase Meidroth: +0.78.
• Michael Massey: −0.78.

Audit and causes:

• Structural audit: PASS; 0 critical/error/warning, 8 info.
• Primary/secondary causes: none assigned because the winner was correct.
• Classification: right winner, wrong score. KC’s starter and lineup both underperformed the projected scoring environment.

This game is also one of the three documented winner-direction regressions at historical_ml → variance_applied: CHW became a tie before recovering to CHW at enhanced prediction.

CIN@ATL — wrong winner and wrong score

• Predicted: CIN 5, ATL 7; ATL winner
• Actual: CIN 7, ATL 6; CIN winner
• Published pick probability: ATL 0.560
• Confidence edge: +0.12, moderate
• Bucket: moderate
• Brier: 0.3136
• Team-run absolute error: 3
• Total-run bias: −1
• Predicted margin: ATL by 2; actual margin: CIN by 1

Edges:

• Season: +1.445
• Recent: +0.590
• Pitcher: +0.032
• Composite: +1.3726

Starters:

• Brady Singer: projected 4 ER/12 outs, actual 5 ER/9 outs; ERA error +6.
• Tyler Mahle: projected 4 ER/18 outs, actual 5 ER/15 outs; ERA error +3.
• Both starters were only one earned run off in absolute terms, so the starter projection was broadly accurate relative to this miss.

Controls and path:

• Confidence reached 0.5973 before controls; published ATL probability was 0.560.
• Savant adjustment applied; no volatility adjustment.
• Historical score path initially favored ATL, variance moved to a tie, enhanced prediction moved toward CIN, and later stored published score was again CIN 5 / ATL 7. This creates the noted transition-trace contradiction.
• Recent form favored ATL: ATL 6–4 with +6 run differential versus CIN 4–6 and −14.

Player adjustments:

• Mike Yastrzemski: largest adverse change, +1.78 projected RBI delta.
• Tyler Stephenson: +1.624 projected BB delta.
• Carlos Jorge: +1.50 projected AB delta.
• Eugenio Suárez: +1.45 favorable RBI delta.
• Michael Harris II: −1.43 AB delta.
• Elly De La Cruz and Matt Olson each had +1.39 RBI-related changes.
• Material player-adjustment count: 23.

Audit and causes:

• Structural audit: PASS; 0 critical/error/warning, 9 info.
• Primary: wrong_conviction_side.
• Secondary: none.
• The miss was not primarily caused by a gross starter projection failure. It was a moderate-confidence commitment to ATL that lost a close game.

CLE@BOS — correct winner, wrong score

• Predicted: CLE 4, BOS 2; CLE winner
• Actual: CLE 1, BOS 0; CLE winner
• Published pick probability: CLE 0.50619
• Confidence edge: −0.0124, toss-up
• Bucket: toss_up
• Brier: 0.243848
• Team-run absolute error: 5
• Total-run bias: +5
• Predicted margin: CLE by 2; actual margin: CLE by 1

Edges:

• Season: +0.1714
• Recent: −0.410
• Pitcher: +0.014
• Composite: −0.3374

Starters:

• Daniel Espino: projected 2 ER/9 outs, actual 0 ER/3 outs; ERA error −6. The audit calls this an outlier with no credible explanatory signal.
• Ranger Suarez: projected 1 ER/17 outs, actual 1 ER/15 outs; essentially accurate.

Controls and path:

• Governed probability 0.5073; published away probability 0.50619.
• Score volatility: no change.
• Both lineups scored below projection.
• The toss-up went the model’s way.

Player adjustments:

• Chase DeLauter, Adley Rutschman, Wilyer Abreu, and Mickey Gasper all had negative run impacts from underproduction.

Audit and causes:

• Structural audit: PASS; 0 critical/error/warning, 7 info.
• Primary/secondary causes: none.
• Classification: correct winner, wrong score; mostly low-run variance and an unexpectedly short, scoreless Espino appearance.

HOU@OAK — correct winner and close score

• Predicted: HOU 7, OAK 5; HOU winner
• Actual: HOU 7, OAK 5; HOU winner
• Published pick probability: HOU 0.56
• Confidence edge: −0.12, moderate
• Bucket: moderate
• Brier: 0.1936
• Team-run absolute error: 0
• Total-run bias: 0
• Predicted and actual margin: HOU by 2

Edges:

• Season: −0.8484
• Recent: −0.162
• Pitcher: +0.0496
• Composite: −0.7580

Starters:

• Peter Lambert: projected 1 ER/17 outs, actual 4 ER/8 outs; ERA error +11.91.
• Mason Barnett: projected 3 ER/11 outs, actual 4 ER/18 outs; ER error +1, outs +7.

Controls and path:

• Governed probability 0.6344 before publication; score coherence and recency produced published HOU probability 0.56.
• Blowout-margin expansion was supported.
• The score remained exactly correct despite very poor Lambert starter accuracy because HOU’s offense compensated.

Player adjustments:

• Lucas Spence: +3.40.
• Zack Gelof: +2.66 and identified as a learnable strong-hitter under-rating.
• Christian Walker: +1.40.
• Jeff McNeil: +1.40.
• Jose Altuve: −1.11.

Audit and causes:

• Structural audit: PASS; 0 critical/error/warning, 8 info.
• Primary/secondary causes: none.
• Strong correct call, but score correctness should not conceal a large pitcher error.

LAA@SEA — wrong winner and wrong score

• Predicted: LAA 4, SEA 5; SEA winner
• Actual: LAA 6, SEA 4; LAA winner
• Published pick probability: SEA 0.560
• Confidence edge: +0.12, moderate
• Bucket: moderate
• Brier: 0.3136
• Team-run absolute error: 3
• Total-run bias: −1
• Predicted margin: SEA by 1; actual margin: LAA by 2

Edges:

• Season: +0.4043
• Recent: +0.614
• Pitcher: +0.1394
• Composite: +0.8625

Starters:

• Grayson Rodriguez: projected 5 ER/15 outs, actual 4 ER/15 outs.
• Bryan Woo: projected 3 ER/17 outs, actual 4 ER/16 outs.
• Both were within one earned run; starter error was not the main cause.

Controls and path:

• Governed probability 0.6366; published SEA probability 0.56 after score-coherence and recency controls.
• Savant adjustment and supported blowout-margin expansion favored SEA.
• Model had SEA recent run differential +13 versus LAA −18, and recent scoring 5.4 versus 3.6.

Player adjustments:

• Josh Naylor: +2.57 projected RBI change.
• Lazaro Montes: +1.70 RBI change.
• Randy Arozarena: −1.70 AB change.
• Zach Neto: +1.64 deboost-related change.
• Julio Rodríguez: actual run impact +1.40 and classified as a learnable under-rating.
• Material change count: 22.

Audit and causes:

• Structural audit: PASS; 0 critical/error/warning, 7 info.
• Primary: wrong_conviction_side.
• Secondary: none.
• The evidence does not support blaming the starters; this was a moderate-confidence wrong-side commitment in a close run environment.

MIA@CHC — correct winner, wrong score

• Predicted: MIA 4, CHC 5; CHC winner
• Actual: MIA 1, CHC 2; CHC winner
• Published pick probability: CHC 0.519035
• Confidence edge: +0.0381, toss-up
• Bucket: toss_up
• Brier: 0.231327
• Team-run absolute error: 6
• Total-run bias: +6
• Predicted margin: CHC by 1; actual margin: CHC by 1

Edges:

• Season: +0.6258
• Recent: −0.098
• Pitcher: +0.0807
• Composite: +0.6772

Starters:

• Tyler Phillips: projected 1 ER/15 outs, actual 0 ER/12 outs.
• Matthew Boyd: projected 3 ER/17 outs, actual 1 ER/21 outs.
• Both outperformed their ER projections, causing the low total.

Controls and path:

• Governed probability 0.5263; published CHC probability 0.5190.
• Starter experience control was not applied.
• Variance action: no change.

Player adjustments:

• Heriberto Hernández: −1.73.
• Seiya Suzuki: +1.09 and identified as a learnable strong-hitter under-rating.
• Pete Crow-Armstrong: −0.93.
• Michael Busch: −0.80.
• Both lineups scored substantially below projection.

Audit and causes:

• Structural audit: PASS; 0 critical/error/warning, 8 info.
• Primary/secondary causes: none.
• Correct winner, wrong score. This is the second documented variance-stage winner regression: historical CHC moved to a tie at variance and recovered to CHC at enhanced prediction.

MIL@PHI — correct winner, wrong score

• Predicted: MIL 5, PHI 4; MIL winner
• Actual: MIL 5, PHI 1; MIL winner
• Published pick probability: MIL 0.53921
• Confidence edge: −0.0784, lean
• Bucket: lean
• Brier: 0.212327
• Team-run absolute error: 3
• Total-run bias: +3
• Predicted margin: MIL by 1; actual margin: MIL by 4

Edges:

• Season: −0.955
• Recent: −0.530
• Pitcher: +0.0056
• Composite: −0.8254

Starters:

• Shane Drohan: projected 1 ER/14 outs, actual 1 ER/8 outs.
• Andrew Painter: projected 5 ER/15 outs, actual 1 ER/18 outs; ER error −4.
• Painter’s strong outing materially reduced PHI’s scoring.

Controls and path:

• Governed probability 0.5543; published MIL probability 0.53921.
• No volatility change.
• Score environment was too high, especially for PHI.

Player adjustments:

• Sal Frelick: +0.97.
• Cooper Pratt: +0.80.
• Trea Turner: −0.78.
• Brice Turang: −0.64.

Audit and causes:

• Structural audit: PASS; 0 critical/error/warning, 7 info.
• Primary/secondary causes: none.
• Correct winner, wrong margin, mainly from PHI starter and offense underproduction.

NYM@TEX — wrong winner and wrong score

• Predicted: NYM 5, TEX 4; NYM winner
• Actual: NYM 1, TEX 3; TEX winner
• Published pick probability: NYM 0.520878
• Confidence edge: −0.0418, lean
• Bucket: lean
• Brier: 0.271314
• Team-run absolute error: 5
• Total-run bias: +5
• Predicted margin: NYM by 1; actual margin: TEX by 2

Edges:

• Season: +0.110
• Recent: −0.018
• Pitcher: −0.0998
• Composite: −0.3436

Starters:

• Zac Thornton: projected 2 ER/17 outs, actual 2 ER/15 outs; close.
• Kumar Rocker: projected 4 ER/14 outs, actual 1 ER/15 outs; ERA error −5.91.

Controls and path:

• Governed probability 0.5246; published NYM probability 0.520878.
• No recency guardrail was applied.
• Score path moved from approximately 4/4 to 5/4 for NYM.
• The model had just seen NYM win the prior two head-to-head games, but TEX won this game.
• Score-volatility action: no change.

Player adjustments:

• Francisco Lindor: +2.53 RBI-related change.
• Josh Jung: +1.76 hit change and actual +2.20 run impact; identified as a learnable under-rating.
• Justin Foscue: −2.50 AB change.
• Juan Soto: −1.305 deboosted BB change and actual negative run impact.
• Material change count: 24.

Audit and causes:

• Structural audit: PASS; 0 critical/error/warning, 7 info.
• Primary: run_environment_over.
• Secondary: none.
• The winner miss began with a near-even directional call and was worsened by a much lower run environment than projected. Rocker’s unexpectedly strong performance contributed, but the stored diagnosis properly identifies the broad run-environment error.

This is the third documented variance-stage instability game: historical evidence favored TEX, variance moved to a tie, and enhanced prediction moved to NYM—the wrong direction relative to the actual result.

SD@LAD — wrong winner and wrong score

• Predicted: SD 4, LAD 5; LAD winner
• Actual: SD 4, LAD 2; SD winner
• Published pick probability: LAD 0.509796
• Confidence edge: +0.0196, toss-up
• Bucket: toss_up
• Brier: 0.259892
• Team-run absolute error: 3
• Total-run bias: +3
• Predicted margin: LAD by 1; actual margin: SD by 2

Edges:

• Season: +0.6803
• Recent: −0.002
• Pitcher: −0.0449
• Composite: +0.4549

Starters:

• Nick Pivetta: projected 2 ER/15 outs, actual 0 ER/14 outs.
• Tyler Glasnow: projected 2 ER/18 outs, actual 3 ER/19 outs.
• Pivetta’s strong start and the low-scoring game were not supported by a credible predictive signal in the audit.

Controls and path:

• Starter-experience control shrank probability from 0.518595 to 0.513559; winner unchanged.
• Published LAD probability was 0.509796.
• Score path remained approximately SD 4 / LAD 5.
• Low-score probability was 0.6925, but volatility action was still no change.

Player adjustments:

• Mookie Betts: +1.58 and classified as a learnable strong-hitter under-rating.
• Dustin Harris: +2.16 hit-related change.
• Hunter Feduccia: +2.28.
• Manny Machado: −1.73.
• Material change count: 24.

Audit and causes:

• Structural audit: PASS; 0 critical/error/warning, 8 info.
• Primary: toss_up_variance.
• Secondary: none.
• This was a genuine low-confidence flip, not a strong model conviction failure.

STL@PIT — correct winner, wrong score

• Predicted: STL 4, PIT 5; PIT winner
• Actual: STL 1, PIT 2; PIT winner
• Published pick probability: PIT 0.52755
• Confidence edge: +0.0551, lean
• Bucket: lean
• Brier: 0.223209
• Team-run absolute error: 6
• Total-run bias: +6
• Predicted margin: PIT by 1; actual margin: PIT by 1

Edges:

• Season: +0.3532
• Recent: +0.462
• Pitcher: −0.0209
• Composite: +0.4438

Starters:

• Kyle Leahy: projected 2 ER/15 outs, actual 1 ER/6 outs.
• Paul Skenes: projected 4 ER/17 outs, actual 1 ER/19 outs; ER error −3.

Controls and path:

• Governed PIT probability 0.5381; published 0.52755.
• No volatility change.
• Both lineups scored three runs below projection.

Player adjustments:

• Jordan Walker: −1.07.
• Brandon Lowe: −0.93.
• Spencer Horwitz: −0.80.
• Leonardo Bernal: −0.78.

Audit and causes:

• Structural audit: PASS; 0 critical/error/warning, 7 info.
• Primary/secondary causes: none.
• Correct winner, wrong score. Skenes’ unexpectedly strong performance and broad offensive underproduction caused the six-run team-error total.

TB@NYY — correct winner, wrong score

• Predicted: TB 2, NYY 4; NYY winner
• Actual: TB 4, NYY 6; NYY winner
• Published pick probability: NYY 0.531136
• Confidence edge: +0.0623, lean
• Bucket: lean
• Brier: 0.219833
• Team-run absolute error: 4
• Total-run bias: −4
• Predicted margin: NYY by 2; actual margin: NYY by 2

Edges:

• Season: +0.185
• Recent: −0.042
• Pitcher: +0.0045
• Composite: +0.4612

Starters:

• Ian Seymour: projected 0 ER/15 outs, actual 0 ER/15 outs.
• Cam Schlittler: projected 1 ER/19 outs, actual 1 ER/9 outs.
• ER projections were exact; outs were wrong for Schlittler.

Controls and path:

• Governed probability 0.5366; published NYY probability 0.531136.
• No volatility change.
• The total was four runs higher than projected, with both offenses contributing.

Player adjustments:

• Austin Wells: +1.73.
• George Lombard Jr.: +1.71.
• Taylor Walls: +1.11.
• Ben Rice: −1.40.
• Cody Bellinger: −1.11.

Audit and causes:

• Structural audit: PASS; 0 critical/error/warning, 7 info.
• Primary/secondary causes: none.
• Correct winner, wrong score. The winner call was solid; the run total was too low.

4. Miss reconstruction

Earliest wrong-direction stages for every loss

CIN@ATL

• Historical model score: approximately CIN 3.76, ATL 5.68, initially favoring ATL—the wrong result.
• Variance stage: moved to a tie, an intermediate winner-direction regression.
• Enhanced prediction: moved toward CIN, which was directionally correct.
• Published score/probability: stored final prediction again favors ATL, creating a trace inconsistency.
• Diagnosis: bad initial weighting toward ATL, followed by documented pipeline instability; starter projections were not the principal issue.

LAA@SEA

• Historical and early score path favored SEA.
• Recent, season, pitcher, and composite edges all favored SEA.
• The eventual actual result was LAA.
• No clearly identifiable later stage moved the model toward LAA; the earliest wrong direction was the historical/form-weighted winner assessment.
• Diagnosis: bad weighting or irreducible variance; not a starter miss.

NYM@TEX

• Historical model was effectively TEX by a small amount: approximately NYM 4.32, TEX 4.35.
• Variance stage moved to a tie.
• Enhanced prediction moved to NYM, which was the wrong direction.
• Later stages retained NYM.
• Diagnosis: winner-direction regression during enhanced prediction after a near-even historical assessment; low run environment and Rocker’s unexpectedly strong start amplified the error.

SD@LAD

• Historical and enhanced stages favored LAD by a small margin.
• Actual SD won.
• Published probability was only 0.5098 for LAD after experience shrinkage.
• No stage shows a material, high-confidence directional failure.
• Diagnosis: toss-up variance, with low scoring and unexpectedly strong Pivetta performance.

Cause classification

The ARI@COL and CHW@KC stage regressions are not losses, but they demonstrate why final winner accuracy alone is insufficient: variance application lost three winner hits overall before enhanced prediction recovered two.

A 0.12 edge is moderate. It should not be described as a near coin flip, but it also does not represent a highly reliable prediction. On this slate both 0.12-confidence calls in CIN@ATL and LAA@SEA lost.

Pipeline-stage quality across all stages

Transitions:

• historical_ml → variance_applied: winner accuracy −25 percentage points, 3 fewer hits; CHW, MIA, and NYM moved to ties.
• variance_applied → enhanced_prediction: winner accuracy +16.67 points, 2 recovered hits; CHW and MIA recovered, NYM moved to NYM.
• enhanced_prediction → embedded_savant_run_environment: no winner changes; small run-metric deterioration.
• All later transitions: no winner changes and no run-metric changes.

A later recovery does not erase the variance-stage regression. The system must preserve both before- and after-stage winner states.

5. Cross-game counterfactuals

Lever tension: widen run distributions

Lever: score.variance_multiplier / score.variance_floor

• Helps:
• CIN@ATL: lost a 0.12-confidence pick; wider tails could reduce overconfidence.
• LAA@SEA: same rationale.
• Hurts: none explicitly identified in the tension register.
• Net evidence: 2 helped, 0 hurt on this slate.

This supports a bounded shadow experiment, not an unrestricted increase. The lever primarily changes spread and tail behavior; it should not be used as a substitute for fixing the near-zero run-model signal.

Lever tension: pitcher form weights and prior-game weighting

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

• Helps:
• ARI@COL: Gordon ER error 7.
• HOU@OAK: Lambert ER error 3.
• MIL@PHI: Painter ER error 4.
• NYM@TEX: Rocker ER error 3.
• STL@PIT: Skenes ER error 3.
• Hurts:
• CIN@ATL: both starter projections were reasonably accurate.
• LAA@SEA: both starters were close.
• TB@NYY: ER projections were exact.
• Net evidence: 5 helped, 3 hurt.

The net slate count favors some investigation, but cross-slate starter_era_error is explicitly nonpersistent: only 5 of 12 prior slates had the same sign, with z = 1.04. Therefore, do not make a direct permanent weight increase yet. Use a shadow test only if the system can isolate credible venue/form evidence.

Lever tension: recent and season run-mean weighting

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

• Helps:
• ARI@COL
• CHW@KC
• CLE@BOS
• MIA@CHC
• NYM@TEX
• STL@PIT
• TB@NYY
• Hurts:
• HOU@OAK, where team-run error was already zero.
• Net evidence: 7 helped, 1 hurt.

This is directionally favorable on the slate, but it must be reconciled with the persistent cross-slate finding that projected lineup runs are too high. The safer response is to lower the projected run environment or widen tails in a bounded experiment, not simply increase recent/season weight without examining feature bias.

Tempting one-game fixes that must not be made

• Do not reverse the SEA pick solely because LAA won; the starter lines were close and the game was a moderate-confidence call.
• Do not permanently increase pitcher weights because of Tanner Gordon, Randy Dobnak, Peter Lambert, or Paul Skenes; these include outlier starts and mixed directional effects.
• Do not tune directly to individual player deviations. Of 77 material player deviations, only 16, or 20.8%, were supported by the hitter’s own season contact quality.
• Do not treat CLE@BOS or SD@LAD as evidence that all low-confidence picks should be reversed; both were genuine toss-up outcomes.
• Do not use the exact score of HOU@OAK to conclude the run model is healthy; it had exact team runs despite a 3-run starter error for Lambert.
• Do not make a permanent calibration change from the 12-game bucket results alone.

6. Data and contract defects

Audit findings by severity

Across all supplied structural audits:

• Critical: 0
• Error: 0
• Warning: 0
• Info: 91

Every individual game reported structural audit PASS with zero critical, error, and warning findings.

Critical and error defects

There are no critical or error findings in the game artifacts.

The missing integrity_revision_audit is a manifest-level missing source, not a reported critical finding inside the game audits. It remains a provenance deficiency.

Warnings

There are no warning findings.

Informational findings

The principal informational findings are:

1. Eight to nine informational notes per game regarding forecast-file integrity and related diagnostics.
2. Five material calibrated-to-published probability movements with named transformations.
3. Zero simulation winner flips on this slate.
4. Complete 12/12 structural coverage.
5. Starter experience and venue-split diagnostics where applicable.
6. Large Savant run-environment swings on several games.

Prediction-model defects versus presentation/artifact bugs

Prediction-model/process defects

• Near-zero run-model holdout signal:
• Total \(R²\): 0.003064
• Home \(R²\): 0.003987
• Away \(R²\): 0.034855
• Variance stage winner-direction regression: 9 historical hits to 6 variance-stage hits.
• Persistent simulation inertness across 11 consecutive slates.
• Persistent broad run-model weakness across 11 consecutive slates.
• Current-slate calibration error of 0.1701.
• Persistent negative lineup-run delta across 9 of 12 slates, mean −0.328, z = −2.25.

Presentation/provenance defects

• Missing integrity revision audit.
• Incomplete per-player starter-line fields for Eduardo Rodriguez and David Sandlin.
• Ambiguous later-stage winner transition reporting, particularly for CIN@ATL.
• Probability and score paths are not always sufficiently explicit to reconcile final side selection from every intermediate state.

These presentation/trace defects matter because they limit forensic reproducibility, even though they do not prove that the published forecast artifact was corrupted.

Coverage reconciliation

The slate is structurally covered:

• 12 expected and 12 audited.
• 12 starter comparisons.
• 12 continuous-learning records.
• 12 forensic records.

“Complete coverage” means every game was audited; it does not mean every prediction was good. The coverage finding is therefore valid but cannot override the high-severity model findings.

Probability-lineage audit

The supplied lineage includes:

1. Raw model pick probability.
2. Calibrated pick probability and empirical sample size.
3. Governor input and output.
4. Score-coherence probability.
5. Recency guardrail.
6. Published probability.

Material examples:

• ARI@COL: raw 0.716, calibrated 0.661, score coherence 0.617535, published pick probability 0.56; named score-coherence and matchup-recency transformation.
• CHW@KC: calibrated 0.620, score coherence 0.619987, published 0.54462; named transformation.
• CIN@ATL: calibrated 0.5973, score coherence 0.585689, published 0.56.
• HOU@OAK: calibrated 0.6344, score coherence 0.634358, published 0.56.
• LAA@SEA: calibrated 0.6366, score coherence 0.605337, published 0.56.
• CLE@BOS: published probability matches the calibrated/governed value.
• NYM@TEX: no recency guardrail; published probability matches calibrated/governed value.
• SD@LAD: experience shrinkage reduced the probability before publication.

The evidence records named transformations for all five material movements and marks unexplained_material_delta: false. No probability-governor bypass is established.

There is, however, a separate probability-reversion event:

• CLE@BOS: blend favored home and changed the winner flag, with blended home probability 0.506608, trained home probability 0.521703, and published home probability 0.49381.
• The published side was away and the outcome was an away hit.
• One blend-stage overwrite occurred across 12 artifacts.

This was not a final prediction failure, but it warrants continued trace visibility.

7. Cross-slate persistence

The evidence includes 10 prior slates and an 11-slate trend window from 2026-09-14 through 2026-09-24.

Finding streaks

Escalated findings

Simulation inertness

• Streak: 11 consecutive slates
• Target lever: simulation.allow_score
• Parameter-actionable: No
• Metric harmed: team_run_mae, total_run_mae
• Direction: confirm the simulation is running and not silently no-oping.

The self-assessment gives a concrete suspected root cause: the simulation distribution is built from the same draws used to find a target-realizing play-by-play path; the loop can early-exit after roughly two selected draws. That makes the derived win rate non-probabilistic and explains repeated zero flip candidates.

Weak run-model signal

• Streak: 11 consecutive slates
• Target levers registered: training.xgb.n_estimators, training.xgb.max_depth, training.xgb.learning_rate
• Parameter-actionable: explicitly No for the current diagnosis
• Metrics harmed: team_run_mae, total_run_mae
• Direction: investigate features and target through an isolated training backtest.

The recommended action is a code/data change, not nightly hyperparameter tuning. Candidate features named in the evidence are:

• Source-labelled no-vig market probability when available
• Park/weather run environment
• Bullpen rest/fatigue
• Platoon-adjusted lineup quality

Metric trends

Every supplied metric trend is labeled improving, but the levels remain weak:

• Margin bias: improving; mean −1.4833
• Mean team-run error on losses: improving; mean 5.3237
• Weighted calibration error: improving; mean 0.2081
• Winner accuracy: improving; mean 0.5556

“Improving” does not mean solved. In particular, the current weighted calibration error is still 0.1701 and run-model \(R²\) remains near zero.

Top-confidence-pick streak

• Current top-confidence loss streak: 0
• The 11-slate sequence ends with hits on 2026-09-23 and 2026-09-24.
• The sequence still contains misses on 2026-09-14, 2026-09-19, and 2026-09-21–22, so the absence of a current streak is not evidence that top-confidence reliability is strong.

Cross-slate structural blocks

Margin distribution

Current slate:

• Actual margins ≥ 4: 3
• Predicted margins ≤ 2: 12/12
• Actual margins: [1,1,1,1,2,2,2,2,2,4,4,8]
• The model systematically compressed tails, especially in ARI@COL, CHW@KC, and MIL@PHI.

Pipeline transitions

The recurring concern is not simply final accuracy; it is intermediate winner-direction regression. Current slate evidence directly confirms the variance-stage regression.

Calibration monitoring

Current bucket behavior is mixed:

• Underconfidence at 0.50–0.55.
• Overconfidence at 0.55–0.60.
• Overall weighted error remains high.

High raw cluster

• Three current-slate games in the high raw cluster.
• Hits: 2
• Misses: 1
• Current finding streak: 0, so it is not an escalated persistent failure.

Simulation control

• 12 games observed.
• Zero simulation flip candidates.
• Zero published flips.
• Current status: diagnostic only without independent support.
• Cross-slate streak: 11, so this is an established code/diagnostic failure rather than single-slate noise.

8. Adjustments for today

Recommendations are ranked by persistence and scope.

1. Fix and independently validate simulation — highest evidence strength

• Registered target: simulation.allow_score
• Direction: code/diagnostic action; do not force a numeric parameter move
• Bounded change: separate score-target realization from winner-probability estimation. Run an independent, unsteered batch of approximately 200–500 draws, or a closed-form Skellam/Poisson-difference calculation, without early exit. Register batch size.
• Expected metric: more informative simulation contribution; improve team_run_mae, total_run_mae, and especially winner-direction diagnostics.
• Helps: all games where simulation was inert, particularly close games such as CIN@ATL, NYM@TEX, and SD@LAD.
• Hurts/regression risk: additional compute and possible score-distribution noise; no lever-tension evidence shows a specific game harmed.
• Rollback monitor: retain the current no-op path as a control and compare flip rate, winner accuracy, Brier, and run MAE over a shadow slate window.

This is the safest persistent-pattern action because the evidence identifies an architectural cause and explicitly says not to force a parameter move.

2. Isolated run-model feature/target backtest — persistent model weakness

• Registered targets: training.xgb.n_estimators, training.xgb.max_depth, training.xgb.learning_rate
• Direction: no direct nightly parameter move; code/data investigation first.
• Bounded change: add the named feature-contract candidates and retrain under a new model schema version; compare same-fingerprint holdout \(R²\) and MAE before any promotion.
• Expected metrics: increase away/home/total run \(R²\), reduce team_run_mae and total_run_mae.
• Helps: all games affected by compressed margins and poor score forecasts, especially ARI@COL, CHW@KC, MIL@PHI, NYM@TEX, STL@PIT, and TB@NYY.
• Hurts/regression risk: market or weather features may overfit or become unavailable; require source-labelled availability and rollback if holdout performance does not improve.

This is a code/data action because the evidence explicitly states that no bounded nightly knob can repair near-zero signal.

3. Bounded shadow increase in run variance — strongest single-slate actionable driver

• Registered parameter: score.variance_multiplier or, if the implementation uses a floor as the spread lever, score.variance_floor
• Direction: increase spread modestly, with the floor interpreted according to the system’s implementation.
• Suggested magnitude: one small shadow step, for example +0.05 to score.variance_multiplier, rather than a large change.
• Expected metric: reduce high-conviction miss rate and weighted calibration error; improve margin-tail coverage. Track Brier, calibration error, margin MAE, team-run MAE, and total-run MAE.
• Helps: CIN@ATL and LAA@SEA, the two 0.12-confidence losses explicitly identified by the lever tension; also helps the broader compressed-margin pattern.
• Hurts: no explicit current-slate hurt in the lever register, but HOU@OAK had exact team runs and could be disturbed; all correctly scored close games are potential regression cases.
• Gate: shadow only; rollback if Brier, total-run MAE, or winner accuracy deteriorates over the registered evaluation window.

This is not a claim that wider variance would have changed the actual winners. It is a bounded calibration/spread experiment motivated by the 3/12 large-margin outcomes and 0/12 predicted margins above two runs.

4. Hold pitcher-weight changes

• Registered lever: player.pitcher.v4 form weights / player.pitcher_prior_games
• Direction: abstain for now
• Reason: the current slate gives a 5-help versus 3-hurt tension, but cross-slate starter ERA bias is nonpersistent.
• Diagnostic action: evaluate credible venue splits and prior-game sample sizes in a shadow report.
• Expected metric if later justified: starter_er_mae, starter_outs_mae, and indirectly total_run_mae.
• Risk: disturbing accurate starter projections in CIN@ATL, LAA@SEA, and TB@NYY.

5. Do not change confidence calibration weights permanently

• Registered levers: confidence.empirical_max_weight, confidence.base_cap, confidence.moderate_cap
• Direction: no permanent move from this slate.
• Reason: current calibration error is high, but the 12-game bucket sample has conflicting underconfidence and overconfidence behavior, and the causal register explicitly requires a rolling-window watch.
• Shadow option: if the system’s existing auto-action mechanism permits it, test a small cap reduction paired with the variance experiment; evaluate Brier and calibration error separately rather than using winner accuracy alone.

9. Bottom line

Known

• The slate produced 8 correct winners out of 12.
• Winner accuracy exceeded the stated walk-forward baseline, but that does not make the prediction system well calibrated or numerically accurate.
• Exact and team-run scores were materially weaker than winners; predicted margins were too compressed.
• The run model has a persistent near-zero holdout signal across 11 slates.
• The variance stage materially degraded winner direction from 9/12 to 6/12 before later recovery.
• Published probability movements have recorded transformations; no material probability bypass is established.
• All 12 games received structural QA with zero critical, error, or warning findings.
• Two persistent escalated failures are established: simulation inertness and weak run-model signal.
• The current slate’s largest practical model risk is overconfident spread/run behavior, especially on moderate-confidence calls.

Uncertain

• Whether changing pitcher venue/form weighting would improve future performance; cross-slate evidence is not persistent enough.
• Whether the current 0.1701 calibration error represents a stable calibration curve problem or slate noise.
• How much a real independent simulation distribution would change winner probabilities, because the current simulation has not supplied an independent signal.
• The exact later-stage winner path for CIN@ATL, because the stored transition summary and per-game published result do not fully reconcile.
• Whether the missing integrity-revision audit contains any additional provenance information.

Single safest next action

Run the bounded simulation code/diagnostic repair in shadow: separate independent winner-probability draws from target-realization draws, remove the early-exit dependence, register the batch size, and compare its winner/Brier/run-metric contribution against the current control before changing model parameters.