Luna deep analysis — slate 20260911

Coverage receipt: COMPLETE · 9/8 sections; 15/15 games; 5/5 causes; 6/6 sources

1. Evidence receipt and gaps

Sources received

The manifest contains seven expected artifacts:

Source	Status	Size	Purpose
`prediction review`	Loaded	377,381 bytes	Official outcomes and scoring errors
`forensic deep audit`	Loaded	240,613 bytes	Probability, provenance, starter, and decision contracts
`continuous learning`	Loaded	2,544,895 bytes	Stage quality, calibration, lineage, score distribution, starter comparisons
`detailed report`	Loaded	352,194 bytes	Full game and simulation rendering
`ultra detailed player audit`	Loaded	32,629,527 bytes	Player-level prediction trace
`structural audit`	Loaded	1,840,289 bytes	Invariant findings
`integrity revision audit`	Missing	—	Post-remediation proof for immutable revision artifacts

The missing integrity-revision audit is the only explicit source gap. It prevents verification of post-Luna immutable-revision artifacts, but the game-level structural audits report clean saved artifacts.

Slate coverage

All 15 expected games were reviewed:

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

Coverage is complete:

Canonical artifacts loaded: 15/15
Continuous-learning games: 15/15
Forensic deep-audit games: 15/15
Starter comparisons: 15/15
Structural audits: 15/15
Overall structural audit: 15/15
Structural summary: 0 critical, 0 error, 0 warning, 110 info

Gaps and internal contradictions

Player-level actuals are unavailable for the entire slate. Every game states that the player feed had no rows for 20260911. Player adjustments can therefore be reviewed as forecast inputs, but player-level realized outcomes cannot be used to verify those adjustments.
The missing integrity-revision artifact is a contract gap. It is not evidence that the prediction artifacts were altered, but it prevents independent verification of immutable-revision remediation.
Probability-lineage contradictions exist in three games: CHW@STL, COL@DET, and LAA@WAS. Their published probabilities differ materially from calibrated/governed values without a recorded transformation.
The high_raw_cluster correctly identifies two games, but its threshold is 0.65, while the broader slate includes other probabilities above or near that level at intermediate stages. This is not necessarily an error, but the cluster definition should be checked against whether it uses raw, calibrated, or published probability.
The calibration bucket labels are internally consistent with the stated definition: negative calibration error means underconfidence and positive means overconfidence. However, the single 0.65–0.70 bucket has only one game, so its 0.6529 error is not statistically stable.

2. Executive diagnosis

Overall result

Winner accuracy: 10/15 = 66.67%
Misses: 5
Mean team-run error on losses: 5.0
Mean total-run bias on losses: +1.4 runs, meaning actual totals were generally lower than projected across the five losses, although the sign varies by game.
Model holdout winner accuracy: 59.82%
Walk-forward baseline accuracy: 55.94%
Current slate beat both, but one slate is not sufficient to establish durable improvement.
Current slate Brier score is not supplied as a single aggregate, but per-game Briers show materially poor results in the high-conviction misses, especially KC@BOS at 0.42625 and TEX@ARI at 0.322638.

Winner accuracy

Winner accuracy was acceptable and above the validated baseline, but it was not uniformly strong:

Correct winners: 10
Wrong winners: 5
Four losses were close or low-confidence situations: BAL@TOR, CLE@MIN, SEA@OAK, and arguably TEX@ARI at a 0.136 away edge.
KC@BOS was the clearest conviction failure: published home probability 0.652878, actual winner KC.

Pipeline winner accuracy was not monotonic:

historical_ml: 9/15 = 60.0%
variance_applied: 8/15 = 53.33%
enhanced_prediction: 10/15 = 66.67%
Final stages: 10/15 = 66.67%

The variance stage caused a one-win regression before later recovery. That intermediate regression must not be erased by the final score.

Calibration quality

Calibration is the largest probability-quality concern.

Weighted absolute calibration error: 0.2091 on 15 games.
0.50–0.55 bucket: confidence 0.5194 versus 0.700 observed hit rate; nominally underconfident by the supplied definition.
0.55–0.60 bucket: 0.5805 versus 0.750; also underconfident.
0.65–0.70 bucket: 0.6529 versus 0.000; overconfident, but based on one game.
High raw cluster: 1 hit out of 2. KC@BOS missed; LAD@MIA hit after a substantial recorded recency adjustment.

The correct diagnosis is not simply “lower all probabilities.” The calibration evidence is unstable on this slate, but the high-confidence miss and the trend record justify a bounded shadow test of confidence moderation.

Score accuracy

Score accuracy is weak and structurally compressed.

Holdout R²:
Away runs: 0.021038
Home runs: 0.011791
Total runs: 0.020666
The model predicted margin ≤2 in all 15 games.
Six actual games had margins ≥4, including margins of 8, 10, and 20.
Margin signed bias: −2.0, consistent with systematic underestimation of winning margins.
The margin-distribution evidence shows compression, not merely isolated score noise.

The run-model issue is a high model risk, not a critical contract failure. The data explicitly says not to infer that exact-score tuning alone will solve it.

Starter accuracy

Starter projections were mixed:

Large starter misses contributed to BAL@TOR, CIN@MIL, CLE@MIN, NYM@NYY, SD@SF, and TEX@ARI.
Some actual starter performances were substantially better than projected, particularly:
Dustin May: projected 4 ER, actual 0
Merrill Kelly: projected 5 ER, actual 1
Taj Bradley: projected 3 ER, actual 0
Some correct games had starter errors of only 1 ER, but five of the six listed games helped by starter reweighting include three correct winners and one close miss.

Cross-slate starter ERA error is not persistent: only 6 of 12 slates share its sign, with z = 0.26. No broad starter-weight change is justified from that trend alone.

Artifact integrity

Artifact integrity was strong at the game level:

Every game’s structural audit passed.
Each game states forecast-file integrity was clean.
No game had structural critical, error, or warning findings.
The missing integrity-revision audit remains a source-level contract gap.
Three probability publications have unexplained material deltas and are therefore critical evidence/contract failures, even though the saved forecast files themselves passed structural QA.

Causal issue register reconciliation

ID	Severity	Evidence	Games	Action	Guardrail
`unexplained_probability_publication_delta`	Critical	Three games have unexplained calibrated-to-published deltas ≥2 percentage points	`CHW@STL`, `COL@DET`, `LAA@WAS`	Block promotion until the missing transformation is recorded and reconciled	Do not assume governor bypass before tracing pick-side math
`single_slate_calibration_instability`	High	Weighted absolute calibration error 0.2091 over 15 games	All 15	Monitor the same buckets over the rolling window; do not immediately retune calibration weights	One slate is too small for a stable calibration curve
`weak_run_model_signal`	High	Run R² approximately 0.012–0.021 on 2,165 rows	All 15	Investigate features/target with isolated training backtest	Positive R² does not make exact scores reliable
`complete_slate_audit_coverage`	Info	15/15 structurally audited	All 15	Retain per-game audit receipts	Coverage does not cancel internal findings
`simulation_inert_on_slate`	Info on this slate, escalated cross-slate	Zero flip candidates and zero published flips	All 15	Measure simulation contribution across rolling slates; cross-slate evidence now requires code/diagnostic action	No winner flips does not prove no score reconciliation occurred

3. Every game

Notation:

P = published probability for the selected side.
edge = published system confidence edge, using the supplied home-minus-away convention.
err = team-run absolute error.
bias = total-run bias, actual total minus projected total as represented by the game record.
Starter notation is projected ER/ERA/outs → actual ER/ERA/outs.
All games had starter identity matches and a structural audit pass.
Player-level realized forensics were unavailable for every game.

1. BAL@TOR — miss; toss-up

Predicted: TOR 5, BAL 4; TOR winner
Actual: BAL 7, TOR 4; BAL winner
P=0.505 home; edge +0.010; bucket toss_up
Brier 0.255025; team-run error 4; total bias −2
Starters:
Chris Bassitt: 4.0 ER / 7.2 ERA / 15 outs → 2 / 3.6 / 15
Max Scherzer: 3.0 / 5.4 / 15 → 6 / 10.8 / 15
Edges: season +0.197, recent +0.606, pitcher −0.067, composite +0.293 home.
Controls: governor applied; raw 0.52765 → governed 0.508; recency guardrail applied; starter-experience control not applied; no volatility change.
Score path: 4.35–4.19 → 4–4 → 4–5 → 4–5 → 4–5 → 4–5.
Player adjustments: 24 material changes; largest was Scherzer projected-hit change of +3, plus bullpen and batter adjustments. No actual player feed existed to validate them.
Audit: structural PASS, 0 defects, 7 informational notes; artifact not causal.
Primary cause: toss-up variance.
Secondary cause: small lean on the losing side.
This is a correct-calibration-type uncertainty case, not a material overconfidence case.

2. CHW@STL — hit

Predicted: STL 6, CHW 4; STL winner
Actual: STL 7, CHW 3; STL winner
P=0.598363 home; edge +0.1967; bucket moderate
Brier 0.161312; team-run error 2; total bias 0
Starters:
Anthony Kay: 4 / 7.2 / 15 → 3 / 6.75 / 12
Matthew Liberatore: 2 / 3.6 / 15 → 2 / 4.15 / 13
Edges: season −0.145, recent +0.300, pitcher +0.019, composite +0.398 home.
Controls: governor applied; raw 0.5936 → governed 0.623; no recency guardrail; no volatility change.
Score path is not separately stored in this game record; final forecast was 4–6.
Player adjustments: none reported in the game-level trace; no actual player feed.
Audit: structural PASS, 0 defects, 7 informational notes.
Winner quality: strong correct call.
Important contract finding: raw/calibrated/governed probability was approximately 0.623, but published probability was 0.598363, a −2.4637 percentage-point unexplained delta. This did not change the winner, but it is a critical lineage defect.

3. CIN@MIL — correct winner, wrong score

Predicted: MIL 8, CIN 6; MIL winner
Actual: MIL 20, CIN 0; MIL winner
P=0.540 home; edge +0.08; bucket lean
Brier 0.2116; team-run error 18; total bias −6
Starters:
Andrew Abbott: 5 / 9.64 / 14 → 8 / 36.0 / 6
Dustin May: 4 / 7.71 / 14 → 0 / 0.0 / 18
Edges: season +1.526, recent +0.158, pitcher −0.038, composite +1.156 home.
Controls: governor and recency guardrail applied; supported blowout-margin expansion; margin adjustment +2.4421.
Score path: base 6–8, then 6–8 through variance, enhanced, Savant, target review, cap, and simulation.
Player adjustments: no player-level actuals; diagnostic flag for a large Savant run-environment swing.
Audit: PASS, 0 defects, 8 informational notes.
Winner quality: right winner, wrong score.
Primary score cause: simultaneous starter extremes and offensive miss; the winner was correct but score quality was poor.

4. CLE@MIN — miss; toss-up

Predicted: MIN 5, CLE 4; MIN winner
Actual: CLE 5, MIN 2; CLE winner
P=0.505263 home; edge +0.0105; bucket toss_up
Brier 0.255291; team-run error 4; total bias +2
Starters:
Parker Messick: 3 / 4.5 / 18 → 1 / 1.69 / 16
Taj Bradley: 3 / 4.76 / 17 → 0 / 0.0 / 18
Edges: season −0.089, recent +0.234, pitcher +0.054, composite +0.309 home.
Controls: governor and recency guardrail applied; raw 0.548191 → governed 0.5073; starter-experience control not applied; no volatility change.
Score path: 4.64–4.04 → 5–4 → 4–5 → 4–5 → 4–5 → 4–5.
Player adjustments: 23 material changes, including Chase DeLauter RBI +2.479 and several batter strikeout/appearance changes; no actual player feed.
Audit: PASS, 0 defects, 7 informational notes.
Primary cause: toss-up variance.
Secondary cause: small lean on losing side.
Blend evidence: an intermediate blend favored CLE at 0.484679 home probability, but later publication selected MIN at 0.505263. This is an intermediate directional movement toward the correct side that did not survive publication.

5. COL@DET — hit

Predicted: DET 5, COL 4; DET winner
Actual: DET 6, COL 2; DET winner
P=0.541169 home; edge +0.0823; bucket lean
Brier 0.210526; team-run error 3; total bias +1
Starters:
Mason Adams: 2 / 4.15 / 13 → 1 / 1.8 / 15
Framber Valdez: 2 / 3.0 / 18 → 0 / 0.0 / 18
Edges: season +1.060, recent +0.140, pitcher −0.062, composite +0.661 home.
Controls: governor applied; inexperienced-starter shrink factor 0.5125, reducing probability from 0.594505 to 0.548434; no volatility change.
Score path is not included in the per-game trace; final forecast 4–5.
Player adjustments: no actual player feed; starter-experience control was the material listed control.
Audit: PASS, 0 defects, 7 informational notes.
Winner quality: solid correct call.
Critical lineage defect: calibrated/governed probability 0.6071, published 0.541169, unexplained delta −6.5931 points. The starter-experience control is recorded, but it does not account for the full published value.

6. HOU@TB — hit; toss-up went right

Predicted: TB 5, HOU 3; TB winner
Actual: TB 3, HOU 1; TB winner
P=0.505 home; edge +0.010; bucket toss_up
Brier 0.245025; team-run error 4; total bias +4
Starters:
Miguel Ullola: 2 / 6.0 / 9 → 0 / 0.0 / 3
Drew Rasmussen: 3 / 4.5 / 18 → 1 / 1.29 / 21
Edges: season +0.482, recent −0.197, pitcher +0.058, composite +0.703 home.
Controls: governor and recency guardrail applied; starter-experience shrink to 0.504987; no volatility change.
Score path is not listed; final forecast 3–5.
Player adjustments: none reported; no player actuals.
Audit: PASS, 0 defects, 7 informational notes.
Winner quality: toss-up went our way.
Blend evidence initially favored HOU, but publication recovered to TB and was correct.

7. KC@BOS — high-conviction miss and correct-winner/wrong-score failure

Predicted: BOS 6, KC 4; BOS winner
Actual: KC 3, BOS 2; KC winner
P=0.652878 home; edge +0.3058; bucket strong
Brier 0.42625; team-run error 5; total bias +5
Starters:
Seth Lugo: 4 / 7.2 / 15 → 2 / 3.6 / 15
Sonny Gray: 3 / 4.5 / 18 → 3 / 4.5 / 18
Edges: season +1.015, recent +0.818, pitcher −0.044, composite +1.025 home.
Controls: governor applied; raw 0.674352 → governed/published 0.652878; no recency or starter-experience control; no volatility expansion.
Score path: 4.19–5.82 → 4–6 → 4–6 → 4–6 → 4–6 → 4–6.
Player adjustments: 20 material changes, including adverse KC changes to projected RBI/strikeouts and positive BOS batter changes; actual player feed absent.
Audit: PASS, 0 defects, 8 informational notes.
Primary cause: run environment over-projected.
Secondary cause: wrong conviction side.
This is the clearest calibration and score-risk event: a strong published probability backed the wrong team, while the actual game was low scoring.

8. LAA@WAS — hit

Predicted: WAS 5, LAA 4; WAS winner
Actual: WAS 4, LAA 3; WAS winner
P=0.575673 home; edge +0.1513; bucket moderate
Brier 0.180053; team-run error 2; total bias +2
Starters:
Yusei Kikuchi: 0 / 0.0 / 15 → 1 / 2.25 / 12
Cade Cavalli: 1 / 1.59 / 17 → 2 / 3.0 / 18
Edges: season +0.213, recent −0.512, pitcher +0.089, composite +0.366 home.
Controls: inexperienced-starter shrink factor 0.8917; governor applied; no volatility change.
Score path is not supplied; final forecast 4–5.
Player adjustments: no player actuals; large Savant run-environment flag.
Audit: PASS, 0 defects, 7 informational notes.
Winner quality: strong correct call.
Critical lineage defect: calibrated/governed 0.6177, published 0.575673, unexplained delta −4.2027 points.

9. LAD@MIA — hit; correct winner, wrong score

Predicted: LAD 6, MIA 4; LAD winner
Actual: LAD 6, MIA 2; LAD winner
P=0.58 for away LAD; edge −0.16 home, equivalent away edge 0.16; bucket moderate
Brier 0.1764; team-run error 2; total bias +2
Starters:
Blake Snell: 2 / 3.18 / 17 → 1 / 1.59 / 17
Ryan Gusto: 1 / 2.08 / 13 → 2 / 3.0 / 18
Edges: season −1.141, recent −0.747, pitcher −0.017, composite −1.212 home.
Controls: inexperienced-starter adjustment; recency guardrail applied; supported blowout-margin expansion.
Score path is not separately supplied; final forecast 6–4.
Player adjustments: no actual player feed; large Savant run-environment flag.
Audit: PASS, 0 defects, 7 informational notes.
Winner quality: strong correct call.
Lineage is fully recorded: calibrated 0.6905, governor input 0.58, governor output 0.6905, score coherence 0.675512, recency guardrail/publication 0.58. This is a large transformation but not an unexplained bypass.

10. NYM@NYY — hit; correct winner, wrong score

Predicted: NYY 6, NYM 4; NYY winner
Actual: NYY 6, NYM 4; NYY winner
P=0.53 home; edge +0.06; bucket lean
Brier 0.2209; team-run error 0; total bias 0
Starters:
Nolan McLean: 2 / 3.0 / 18 → 5 / 9.0 / 15
Carlos Rodón: 2 / 3.18 / 17 → 1 / 1.42 / 19
Edges: season +0.938, recent +0.083, pitcher +0.097, composite +1.111 home.
Controls: high-total-environment lift and blowout-margin expansion; recency guardrail applied.
Score path is not supplied; final forecast 4–6, which was numerically exact in total and team scores despite starter error.
Player adjustments: no actual player feed; no structural issue.
Audit: PASS, 0 defects, 8 informational notes.
Winner quality: solid correct call.

11. PHI@ATL — hit; correct winner, wrong score

Predicted: ATL 4, PHI 3; ATL winner
Actual: ATL 6, PHI 5; ATL winner
P=0.527156 home; edge +0.0543; bucket lean
Brier 0.223581; team-run error 4; total bias −4
Starters:
Aaron Nola: 4 / 6.35 / 17 → 2 / 2.84 / 19
Chris Sale: 2 / 2.84 / 19 → 3 / 3.86 / 21
Edges: season +0.253, recent −0.206, pitcher +0.133, composite +0.560 home.
Controls: recency guardrail; large Savant run-environment flag; no volatility change.
Score path not supplied; final 3–4.
Player adjustments: no actual player feed.
Audit: PASS, 0 defects, 7 informational notes.
Winner quality: solid correct call.

12. PIT@CHC — hit; correct winner, wrong score

Predicted: CHC 5, PIT 4; CHC winner
Actual: CHC 12, PIT 2; CHC winner
P=0.53 home; edge +0.06; bucket lean
Brier 0.2209; team-run error 9; total bias −5
Starters:
Wilber Dotel: 2 / 3.6 / 15 → 3 / 13.5 / 6
Shota Imanaga: 2 / 3.38 / 16 → 2 / 3.0 / 18
Edges: season +0.285, recent −0.270, pitcher +0.056, composite +0.393 home.
Controls: recency guardrail; no volatility change.
Score path not supplied; final 4–5.
Player adjustments: no actual player feed.
Audit: PASS, 0 defects, 8 informational notes.
Winner quality: right winner, wrong score.

13. SD@SF — hit; toss-up went right

Predicted: SD 5, SF 4; SD winner
Actual: SD 7, SF 5; SD winner
P=0.495 home, therefore away SD selected; edge −0.01 home; bucket toss_up
Brier 0.245025; team-run error 3; total bias −3
Starters:
Robbie Ray: 3 / 5.4 / 15 → 5 / 9.0 / 15
Anthony Molina: 4 / 7.2 / 15 → 7 / 21.0 / 9
Edges: season −0.497, recent −0.190, pitcher −0.061, composite −0.634 home.
Controls: starter-experience shrink factor 0.675; recency guardrail; margin adjustment −0.2086; no volatility change.
Score path not supplied; final 5–4.
Player adjustments: no actual player feed.
Audit: PASS, 0 defects, 7 informational notes.
Winner quality: toss-up went our way.
Blend briefly favored SF, but publication correctly selected SD.

14. SEA@OAK — miss; toss-up

Predicted: SEA 7, OAK 6; SEA winner
Actual: OAK 6, SEA 5; OAK winner
P=0.494196 home, away SEA selected; edge −0.0116 home; bucket toss_up
Brier 0.255838; team-run error 2; total bias +2
Starters:
George Kirby: 3 / 4.76 / 17 → 4 / 6.0 / 18
Jeffrey Springs: 3 / 5.79 / 14 → 2 / 3.6 / 15
Edges: season −0.514, recent +0.218, pitcher −0.007, composite −0.423 home.
Controls: recency guardrail; no volatility change.
Score path: 5.24–5.19 → 5–5 → 7–6 → 7–6 → 7–6 → 7–6.
Player adjustments: 25 material changes, including Canzone RBI +3.488 and changes involving Springs, bullpen, and multiple hitters; no actual player feed.
Audit: PASS, 0 defects, 8 informational notes.
Primary cause: toss-up variance.
Intermediate blend favored OAK, but final publication selected SEA; this is a winner-direction regression that occurred before publication.

15. TEX@ARI — miss; moderate away edge

Predicted: TEX 6, ARI 4; TEX winner
Actual: ARI 9, TEX 1; ARI winner
P=0.568012 away; away edge 0.136; bucket moderate
Brier 0.322638; team-run error 10; total bias 0
Starters:
MacKenzie Gore: 4 / 7.2 / 15 → 4 / 9.0 / 12
Merrill Kelly: 5 / 8.44 / 16 → 1 / 2.08 / 13
Edges: season +0.085 home, recent −0.493, pitcher +0.024, composite −0.384 home, supporting TEX before the final direction.
Controls: recency guardrail; no volatility change.
Score path: 4.76–3.69 → 5–4 → 5–4 → 6–4 → 6–4 → 6–4.
Player adjustments: 20 material changes, mainly TEX batter appearance/hit adjustments and ARI batter deboosts; no actual player feed.
Audit: PASS, 0 defects, 7 informational notes.
Primary cause: starter projection.
Secondary cause: wrong conviction side.
The home-away edge was 0.136, which is moderate—not a near coin flip. The incorrect TEX selection therefore deserves more scrutiny than the toss-up losses, although the starter miss was the dominant direct cause.

4. Miss reconstruction

Loss-by-loss earliest wrong-direction stage

Game	Earliest wrong-direction movement	Classification	Reconstruction
`BAL@TOR`	`historical_ml` already favored BAL: 4.35 away vs 4.19 home	Bad weighting / near-toss-up direction	Variance moved to 4–4, then enhanced prediction moved toward TOR 4–5. The model ended on the wrong side with only a 0.01 edge. Starter miss, especially Scherzer allowing 6 ER instead of 3, reinforced the loss.
`CLE@MIN`	`historical_ml` favored CLE: 4.64 away vs 4.04 home	Bad weighting / later winner-direction regression	Variance retained CLE at 5–4, but enhanced prediction flipped to MIN 4–5. The actual result was CLE 5–2. Both starters outperformed ER projections, especially Bradley, but that does not explain the winner-direction flip by itself.
`KC@BOS`	`historical_ml` favored BOS and the model remained on BOS	Run-environment miss plus overconfidence	No winner-direction regression occurred. BOS was backed with real conviction, but actual total was 5 versus projected 10. The score distribution was too narrow and the model overtrusted season/recent edges despite the pitcher edge being slightly against BOS.
`SEA@OAK`	`historical_ml` was essentially tied, 5.24 SEA vs 5.19 OAK	Bad weighting / score-stage regression	Variance tied the game; enhanced prediction moved to SEA 7–6. The actual result was OAK 6–5. The blend temporarily favored OAK, but later stages/publication selected SEA. This is an intermediate winner-direction regression even though the final result was a toss-up.
`TEX@ARI`	`historical_ml` favored TEX, 4.76–3.69	Bad input or weighting, then starter error	The model remained on TEX through all stages, increasing TEX to 6 runs after Savant adjustment. Kelly allowed 1 ER rather than 5, a 4-ER projection error; ARI scored 9. This was not a near coin flip: the away edge was 0.136.

Cause categories

Bad input: Not proven. The player feed was absent, but no direct malformed input was identified. The missing feed limits attribution.
Bad weighting: Supported in BAL@TOR, CLE@MIN, and SEA@OAK, where later stages moved toward the wrong winner direction.
Calibration: Clearly implicated in KC@BOS, where 0.652878 was too strong for the realized result; not proven as the sole cause of the toss-up losses.
Starter error: Dominant in TEX@ARI; materially contributed in BAL@TOR, and affected several score misses.
Offense/bullpen miss: Strongly present in CIN@MIL, PIT@CHC, KC@BOS, and TEX@ARI; bullpen-specific causal attribution is not proven because player actuals are absent.
Run-environment miss: Dominant in KC@BOS; important in CIN@MIL, PHI@ATL, and PIT@CHC.
Irreducible variance: Dominant classification for BAL@TOR, CLE@MIN, and SEA@OAK, all with approximately 0.01 published edges.

A 0.12 home-away confidence edge is moderate, not a near coin flip. The record supports treating TEX@ARI differently from the three 0.01 toss-ups.

Pipeline-stage quality across all stages

Stage	Winner accuracy	Team-run MAE	Total-run MAE	Margin MAE
`historical_ml`	60.0% (9/15)	2.4049	2.8753	4.1176
`variance_applied`	53.33% (8/15)	2.4000	2.9333	4.1333
`enhanced_prediction`	66.67% (10/15)	2.4333	2.7333	3.9333
`embedded_savant_run_environment`	66.67%	2.4000	2.5333	3.8667
`fundamental_edge_target_review`	66.67%	2.4000	2.5333	3.8667
`capped_target`	66.67%	2.4000	2.5333	3.8667
`simulation_result`	66.67%	2.4000	2.5333	3.8667
`published_final`	66.67%	2.4000	2.5333	3.8667

Transitions:

Historical → variance: winner accuracy fell by one game, from 9 to 8; margin MAE worsened by 0.0157.
Variance → enhanced: winner accuracy recovered by two games; HOU@TB, NYM@NYY, and SD@SF improved, while CLE@MIN regressed. Margin MAE improved by 0.20, but team-run MAE worsened by 0.0333.
Enhanced → embedded Savant: margin MAE improved by 0.0666 and total-run MAE improved by 0.20, with no winner changes.
All later transitions were inert on winner direction and numerically identical.

The later recovery does not erase the variance-stage winner regression or the enhanced-stage CLE@MIN regression. The simulation generated no independent winner intervention on this slate.

5. Cross-game counterfactuals

Lever tension 1: widen run distributions

Lever: score.variance_multiplier / score.variance_floor Target metrics: Brier score and calibration error.

Helps:
KC@BOS: actual 5 total versus projected 10; wider tails reduce overcommitment.
TEX@ARI: actual 10-run differential versus predicted 2; wider tails acknowledge score uncertainty.
Hurts:
CHW@STL: correct moderate call at 0.20; widening may soften a good edge.
Net evidence: 2 helps versus 1 hurts, but the evidence is score-oriented rather than a demonstrated winner-improvement effect. A small shadow experiment is defensible; a large production change is not.

Lever tension 2: pitcher form weights/prior games

Lever: player.pitcher.v4 form weights / player.pitcher_prior_games Target metrics: starter ER MAE and starter outs MAE.

Helps:
BAL@TOR, Scherzer off by 3 ER.
CIN@MIL, Abbott/May errors up to 4 ER.
CLE@MIN, Bradley off by 3 ER.
NYM@NYY, McLean off by 3 ER.
SD@SF, Molina off by 3 ER.
TEX@ARI, Kelly off by 4 ER.
Hurts:
CHW@STL, LAA@WAS, LAD@MIA, and PIT@CHC, where starter ER was off by only 1.
SEA@OAK, also only 1 ER error despite being a winner miss.
Net evidence: 6 helps versus 5 hurts. Current cross-slate evidence does not show a persistent starter bias, so no broad production reweight is justified.

Lever tension 3: recent/season run-mean weighting

Lever: recent.* / season.* run-mean weighting Target metrics: team-run MAE, total-run MAE, accuracy.

Helps: BAL@TOR, CIN@MIL, CLE@MIN, HOU@TB, KC@BOS, PHI@ATL, PIT@CHC, TEX@ARI.
Hurts: CHW@STL, LAA@WAS, LAD@MIA, NYM@NYY, SEA@OAK.
Net evidence: 8 helps versus 5 hurts, but the benefit is not cleanly separable from the underlying run-model weakness. Cross-slate lineup-run bias is currently nonpersistent, so this should remain a shadow experiment rather than a direct production adjustment.

Tempting one-game fixes that must not be made

Do not reverse all toss-ups after BAL@TOR, CLE@MIN, and SEA@OAK; the slate also contains correct toss-up calls in HOU@TB and SD@SF.
Do not eliminate recency controls because TEX@ARI missed; recency guardrails were recorded and helped preserve correct direction in other games.
Do not increase every high-confidence probability solely because LAD@MIA hit after a large adjustment; KC@BOS was a high-confidence loss.
Do not increase variance enough to erase the CHW@STL moderate correct call.
Do not globally change starter priors from this slate; the 6-versus-5 lever tension is nearly balanced and the cross-slate starter sign is not persistent.
Do not infer that simulation is harmless merely because final winner accuracy recovered; its intermediate and cross-slate inertness remains a diagnostic issue.

6. Data and contract defects

Critical findings

Unexplained probability publication deltas
CHW@STL: calibrated/governed 0.623, published 0.598363, delta −2.4637 pp.
COL@DET: calibrated/governed 0.6071, published 0.541169, delta −6.5931 pp.
LAA@WAS: calibrated/governed 0.6177, published 0.575673, delta −4.2027 pp.
The contract explicitly says not to call these governor bypasses without tracing pick-side math. The correct finding is an unexplained transformation, not yet a proven bypass.
Required action: block promotion until each transformation is recorded and reconciled.

Missing integrity revision audit
Manifest status is missing.
This is a source/contract defect, not a proven prediction defect.

Error findings

Structural audit reports zero error findings across all 15 games.
No malformed starter identity, forecast-file defect, or invariant failure is reported.

Warning findings

Structural audit reports zero warnings.

Informational findings

110 total informational findings in the structural audit summary.
Per-game audits generally report 7 or 8 informational notes.
All 15 games structurally audited.
Simulation produced zero flip candidates and zero published flips.
Player-level feed was absent across the slate.
The final artifacts passed their structural checks.

Model defects versus presentation/artifact defects

Prediction-model/process risks:

Near-zero run-model holdout signal.
Margin compression: all predicted margins ≤2 despite six actual margins ≥4.
High-confidence miss in KC@BOS.
Starter and offense/run-environment errors.
Simulation module’s 11-slate inert pattern.

Presentation/artifact/contract defects:

Three unexplained pick-side probability publication deltas.
Missing integrity-revision audit.
Absent player-level actual feed, which limits forensic verification but is not itself proof that the forecasts were wrong.
No evidence that the clean saved forecast artifacts caused the misses.

Complete pick-side probability lineage

The lineage was audited from raw model → calibrated → governor → score coherence/recency → published.

Fully reconciled examples:
BAL@TOR: calibrated 0.508 → published 0.505; recorded recency guardrail.
CIN@MIL: 0.6213 → 0.54; recorded score coherence and recency.
KC@BOS: 0.6529 → 0.652878; essentially unchanged.
LAD@MIA: 0.6905 → 0.58; recorded score coherence and matchup-recency guardrail.
NYM@NYY: 0.6162 → 0.53; recorded recency guardrail.
TEX@ARI: 0.58 → 0.568012; recorded recency.
Unexplained material deltas:
CHW@STL
COL@DET
LAA@WAS

A bypass should be called only for those three pending mathematical trace reconciliation. Other large deltas have recorded transformations.

7. Cross-slate persistence

The cross-slate evidence covers 11 slates for most trends and 12 slates for the separate bias analysis. A single slate is noise, but escalated findings are explicitly established failures.

Finding streaks

Finding	Consecutive-slate count	Present now?	Interpretation
`high_raw_probability_cluster`	0	No	Occurred in 2 of 11 slates but is not current; no immediate lever move
`pipeline_winner_instability`	0	No	Occurred in 3 of 11; not current and not escalated
`score_margin_compression`	0	No	Occurred in 2 of 11; not current as a registered finding, although current margin data is concerning
`simulation_inert_on_slate`	11	Yes	Established persistent failure, not single-slate noise
`single_slate_calibration_instability`	1	Yes	Current slate contributes to an incidence of 5/11, but its consecutive streak is only one; do not call the calibration curve stable
`weak_run_model_signal`	11	Yes	Established persistent failure, not single-slate noise

Escalated findings

`simulation_inert_on_slate`

Consecutive streak: 11
Incidence: 11/11
Target lever listed: simulation.allow_score
Target metrics: team-run MAE and total-run MAE
This is not parameter-actionable according to the register. The recommended direction is to confirm that simulation is running rather than silently no-oping.
The self-assessment identifies the likely root cause: the same draws are used both for target-realization search and win-distribution estimation, with an early exit after approximately two selected draws. That makes the derived win rate statistically meaningless.
Correct action: code/diagnostic repair, not a forced parameter move.

`weak_run_model_signal`

Consecutive streak: 11
Incidence: 11/11
Listed levers: training.xgb.n_estimators, training.xgb.max_depth, training.xgb.learning_rate
However, the finding is marked not parameter-actionable because the features/target contract itself is inadequate.
Target metrics: team-run MAE and total-run MAE.
Correct direction: isolated backtest with better features, including valid no-vig market probability where available, park/weather environment, bullpen fatigue, and platoon-adjusted lineup quality. Do not expect a nightly hyperparameter tweak to solve near-zero R².

Metric trends

margin_bias: regressing; mean −1.6468. This means predicted margins remain too narrow.
mean_team_run_error_on_losses: improving; mean 5.8676 across the series. Current slate is 5.0, better than the series mean, but still large.
weighted_calibration_error: regressing; mean 0.1643. Current slate 0.2091 is worse than the mean.
winner_accuracy: improving; mean 0.5751. Current slate 0.6667 is above the series mean.
Top-confidence-pick loss streak: 1. The per-slate sequence ends with a miss on 20260911 after three prior hits (20260908–20260910). This is not a long loss streak, but the current miss matters because it was a high-conviction failure.

Cross-slate bias holds

Starter ERA error: positive mean 0.118, only 6/12 same-sign slates, z=0.26; hold.
Lineup run delta: negative mean −0.142, 8/12 same-sign slates, z=−0.65; hold.
These are not escalated and should not be treated as established persistent biases.

8. Adjustments for today

Ranked by persistence, scope, expected metric, affected games, and regression risk.

1. Simulation: code/diagnostic repair, not a parameter promotion

Persistent pattern: simulation_inert_on_slate, 11 consecutive slates.
Registered target: simulation.allow_score; target metrics team-run MAE and total-run MAE.
Action: do not change the parameter blindly. Audit and repair the simulation path so the win-probability distribution is generated from an independent, unsteered batch, such as 200–500 draws or a closed-form run-difference distribution. Remove the early-exit contamination from the probability calculation.
Expected metric direction: simulation contribution becomes measurable; eventual improvement should be assessed in team-run MAE, total-run MAE, winner Brier, and winner-direction interventions.
Helps: all games where simulation could correctly identify tail or winner uncertainty, especially high-variance cases such as KC@BOS, CIN@MIL, PIT@CHC, and TEX@ARI.
Hurts/regression risk: potentially any game if independent simulation overturns a good deterministic call, especially CHW@STL, LAA@WAS, and LAD@MIA.
Gate: shadow only until independent distribution quality and rollback monitoring are demonstrated.

2. Run model: feature-contract and isolated-training repair

Persistent pattern: weak_run_model_signal, 11 consecutive slates.
Registered candidate parameters: training.xgb.n_estimators, training.xgb.max_depth, training.xgb.learning_rate; however, the finding explicitly says these are not sufficient and is not parameter-actionable.
Action: code/data work in the shared feature contract and training pipeline. Add validated park/weather, bullpen rest/fatigue, platoon-adjusted lineup quality, and source-labelled no-vig market probabilities where available; retrain with a new model schema version and run a same-fingerprint isolated backtest.
Expected metric direction: increase holdout home/away/total R² and reduce team-run MAE, total-run MAE, and margin compression.
Helps: especially CIN@MIL, KC@BOS, PIT@CHC, and TEX@ARI, where run tails were badly missed.
Hurts/regression risk: may disturb close correct calls such as CHW@STL, LAA@WAS, LAD@MIA, and NYM@NYY.
Gate: no production promotion without holdout improvement and regression monitoring.

3. Single strongest current-slate driver: bounded confidence moderation

Current driver: calibration deterioration plus the KC@BOS high-confidence miss; current weighted calibration error 0.2091, high-conviction miss rate 0.4.
Exact registered parameters: confidence.base_cap and confidence.moderate_cap.
Recommended bounded change: lower each cap by 0.02 in a shadow experiment, not a direct irreversible promotion.
Expected metric direction: lower Brier score and lower weighted calibration error, with reduced probability overstatement.
Helps: KC@BOS most directly; also moderates TEX@ARI and other games where intermediate probabilities were stronger than the evidence warranted.
Hurts: can soften the correct CHW@STL moderate call and reduce useful separation in LAD@MIA, LAA@WAS, and COL@DET.
This is preferable to an aggressive global recalibration because current bucket evidence is unstable and the 0.65–0.70 bucket contains only one game.

4. Score variance: optional secondary shadow experiment

Current evidence: margin compression is materially regressing; all 15 predicted margins were ≤2, while six actual margins were ≥4.
Exact registered parameters: score.variance_multiplier or score.variance_floor.
Recommended bounded change: increase the variance multiplier by a small amount, approximately 0.05, in shadow only; do not simultaneously make a large floor change.
Expected metric direction: improve margin MAE, tail calibration, and potentially Brier score.
Helps: KC@BOS, TEX@ARI, CIN@MIL, PIT@CHC.
Hurts: can widen already reasonable forecasts such as CHW@STL, LAA@WAS, LAD@MIA, and NYM@NYY.
This is lower priority than the code and feature-contract repairs because the run model’s near-zero R² limits what variance tuning can accomplish.

Abstentions

Abstain from changing context.opponent_strength_weight; cross-slate starter ERA bias is not persistent.
Abstain from changing score.variance_floor based only on the lineup-run negative bias; that bias is not separable from noise.
Abstain from broad pitcher-prior reweighting; the current 6-help/5-hurt tension and cross-slate evidence do not support it.
Do not force a parameter move for the two escalated findings whose registered actions are code/diagnostic actions.

9. Bottom line

Known

The slate was structurally covered: 15/15 games, all starter comparisons, all game audits.
The system selected 10 winners correctly, beating its walk-forward baseline.
It also displayed clear score weakness: near-zero run-model R², compressed margins, and large tail misses.
KC@BOS was the most important model-quality failure: a 0.6529 published probability backed the wrong team in a low-scoring game.
TEX@ARI was a meaningful moderate-edge miss, not a coin flip, driven primarily by Merrill Kelly’s four-ER projection error and an offense/run-distribution miss.
Simulation inertness and weak run-model signal are established 11-slate patterns.
Three published probabilities have unexplained material lineage deltas and must block promotion until reconciled.

Uncertain

Player-level causal attribution remains uncertain because the entire slate lacks joined player actuals.
The single-slate calibration curve is unstable despite worsening rolling calibration trends.
It is not proven that pitcher weighting, recent/season weighting, or variance changes will improve production performance; each has cross-game tradeoffs.
The missing integrity-revision audit leaves post-remediation artifact verification incomplete.

Safest next action

The safest immediate action is to block promotion and run a shadow audit of the complete probability lineage for CHW@STL, COL@DET, and LAA@WAS, while separately repairing the simulation’s independent win-distribution path. This addresses the only critical contract failure first and targets the strongest persistent process defect without making an unvalidated production parameter change.