# Luna deep analysis — slate 20260904

> Coverage receipt: **COMPLETE** · 9/8 sections; 16/16 games; 4/4 causes; 5/6 sources

# 1. Evidence receipt and gaps

## Source manifest

Six of seven declared artifacts were loaded:

| Source | Status | Purpose |
|---|---:|---|
| `prediction_review_20260904.json` — 399,800 bytes | Loaded | Official outcomes and scoring errors |
| `PREDICTION_SYSTEM_DEEP_AUDIT_20260904.json` — 255,680 bytes | Loaded | Probability, provenance, starter, and decision contracts |
| `continuous_learning_daily_20260904.json` — 2,555,562 bytes | Loaded | Pipeline quality, calibration, probability lineage, score distribution, and starter comparisons |
| `detailed_20260904.html` — 380,084 bytes | Loaded | Full-game and simulation rendering |
| `20260904_ultra_detailed_player_audit.html` — 33,148,291 bytes | Loaded | Player-level prediction trace |
| `audit_report_20260904.html` — 1,949,987 bytes | Loaded | Invariant findings |
| `audit_report_20260904_integrity_revision.html` | **Missing** | Post-Luna remediation proof for immutable revision artifacts |

The missing integrity-revision artifact is a receipt gap, but the supplied data does not attribute any game-level prediction error to an artifact defect. Every game’s structural audit passed.

## Slate coverage

All 16 listed games were reviewed:

ARI@HOU, ATL@PHI, BOS@BAL, CHC@MIA, DET@CLE Game 1, DET@CLE Game 2, LAA@PIT, MIL@CIN, MIN@CHW, NYY@SD, OAK@SEA, SF@NYM, STL@COL, TB@TEX, TOR@KC, WAS@LAD.

Coverage claims:

- Expected games: 16.
- Games reviewed: 16.
- Structural audits: 16/16.
- Canonical artifacts loaded: 16.
- Continuous-learning games: 16.
- Forensic deep-audit games: 16.
- Starter comparisons: 15/16.
- Structural audit summary: 0 critical, 0 error, 117 info, 0 warning.

The second DET@CLE entry has no `starter_lines`, explaining the 15/16 starter-comparison coverage. Two entries share the same matchup label, but the second has a distinct game ID (`20260904_DET@CLE_2`), so this is an identifiable doubleheader rather than an unexplained duplicate.

## Other evidence gaps

- Player-level actuals were unavailable for BOS@BAL, MIN@CHW, NYY@SD, OAK@SEA, TB@TEX, and TOR@KC. The diagnosis explicitly says “no per-player actuals for this game yet.”
- No source-labelled moneyline was available for any game; market agreement was therefore not testable.
- Six games do not include a detailed `score_evolution` array in the supplied diagnosis: ARI@HOU, ATL@PHI, CHC@MIA, DET@CLE Game 1, DET@CLE Game 2, LAA@PIT, MIL@CIN, SF@NYM, STL@COL, and WAS@LAD. Their final scores and stage-level aggregate metrics are available, but not their individual score paths.
- The probability lineage contains two critical unexplained publication deltas: ARI@HOU and SF@NYM.
- The causal register lists DET@CLE twice in several affected-game arrays. This appears to reflect the doubleheader or repeated aggregation, but the register does not distinguish the game IDs.

# 2. Executive diagnosis

## Overall result

- Winner accuracy: **10/16, 62.5%**.
- Walk-forward baseline accuracy: **55.9%**.
- Holdout/model validation winner accuracy: **58.4%**.
- Slate winner result: better than the supplied baseline, but not strong enough to establish a durable improvement by itself.
- Losses: BOS@BAL, MIN@CHW, NYY@SD, OAK@SEA, TB@TEX, TOR@KC.
- Mean team-run absolute error on losses: **6.0**.
- Mean total-run bias on losses: **+2.0**, meaning the model generally overprojected total scoring across the losses.
- Weighted absolute calibration error: **0.1316**.
- High-conviction misses: 3/9, a **33.3% miss rate**.
- Exact score accuracy is not reported as a percentage. Four correct-winner games were explicitly marked `right_winner_wrong_score`: ARI@HOU, DET@CLE Game 1, LAA@PIT, and SF@NYM.

A correct winner does not imply a good full prediction. For example, LAA@PIT selected the correct side but missed the total by 11 runs; ARI@HOU selected HOU but projected 6–4 instead of 3–1.

## Separate diagnoses

### Winner accuracy

The directional model was moderately successful at 62.5%, with the stronger directional evidence generally working:

- CHC@MIA: composite home edge −1.2062, CHC won.
- LAA@PIT: composite home edge +1.0202, PIT won.
- MIL@CIN: composite home edge −1.9382, MIL won.
- WAS@LAD: composite home edge +0.4900, LAD won.

The major directional failures were less about tiny random margins in some cases and more about low-scoring or starter-driven reversals:

- NYY@SD: a strong 0.29 away confidence edge lost.
- TOR@KC: a moderate 0.11 home edge lost by seven runs.
- OAK@SEA: a lean 0.10 home edge lost after Logan Gilbert allowed seven earned runs.
- MIN@CHW: near-even composite evidence, but the model still selected MIN.

### Calibration quality

Calibration was weak despite acceptable winner accuracy:

- Weighted absolute calibration error: 0.1316.
- The 0.50–0.55 bucket had average confidence 0.5396 and observed hit rate 0.5714.
- The 0.55–0.60 bucket had average confidence 0.5626 and observed hit rate 0.6667.
- The 0.60–0.65 bucket contained one game at 0.6434 that lost.
- The 0.65–0.70 bucket contained one game at 0.662 that hit.
- The 0.70–0.80 bucket contained one game at 0.7245 that hit.

The bucket labels are internally consistent with the stated definition—published confidence minus observed hit rate—but the one-game high buckets are not statistically reliable. The high-raw cluster was 3/3 on this slate, while its empirical historical hit rates were only approximately 0.548–0.550. That is evidence against trusting raw high confidence, not evidence that the current cluster is correctly calibrated.

### Score accuracy

The run model is the clearest model limitation:

- Historical holdout R²:
  - Away runs: 0.026352.
  - Home runs: 0.018723.
  - Total runs: 0.026024.
- Walk-forward Brier: 0.244044.
- Historical ML stage total-run MAE: 4.2626.
- Published final total-run MAE: 4.1875.
- Published final team-run MAE: 2.4062.
- Margin MAE: 2.0625.
- Margin distribution: 13/16 predicted margins were ≤2, while 3 actual margins were ≥4.
- Mean signed predicted-margin error was +0.562 in the spread analysis, while the cross-slate margin-bias series is improving toward zero.

A small positive R² means the model explains slightly more than a trivial average target. It does not make exact scores reliable. The slate had both under-dispersed and over-dispersed outcomes: BOS@BAL, MIN@CHW, and ARI@HOU were large total overprojections, while SF@NYM and TB@TEX were material underprojections.

### Starter accuracy

Starter projections were a material contributor:

- Three losses were labelled primarily `starter_projection`: MIN@CHW, OAK@SEA, TOR@KC.
- Several correct-winner games also had very large starter ER errors, including ARI@HOU, LAA@PIT, and MIL@CIN.
- The lever-tension evidence says pitcher-weight changes helped 10 games and hurt 3, but this is an observational slate count, not proof that reweighting will improve future performance.
- Starter identity was false in both DET@CLE Game 1 comparisons, so those comparisons should not be used as clean evidence for starter-model tuning.

### Artifact integrity

Artifact integrity was not the cause of the six misses:

- Every game had a structural audit status of PASS.
- There were no critical, error, or warning findings in the per-game audits.
- Forecast-file integrity was reported clean in every game.
- However, the complete probability contract is not clean: ARI@HOU and SF@NYM have unexplained calibrated-to-published deltas of at least two percentage points. This is a **critical evidence/contract failure**, not merely a model-quality issue.
- The missing integrity-revision artifact is an additional receipt gap.

## Causal issue register reconciliation

### Critical: `unexplained_probability_publication_delta`

- Severity: **Critical**.
- Evidence: two games had unexplained calibrated-to-published deltas ≥2 percentage points.
- Affected games: ARI@HOU and SF@NYM.
- Action: block promotion until the missing probability transformation is recorded and reconciled.
- Guardrail: do not assume the probability governor was bypassed before tracing the pick-side math.

The supplied lineage does show governor values and score-coherence values, but it does not explain the final publication delta in those two games. This is enough to block promotion, not enough to conclude a governor bypass.

### High: `weak_run_model_signal`

- Severity: **High**.
- Evidence: home/away/total R² of 0.0187/0.0264/0.0260.
- Affected games: all 16, with DET@CLE listed twice by the register.
- Action: treat exact scores as low-confidence; investigate stronger features or a better target before tuning score tails.
- Guardrail: small positive R² does not establish reliable exact-score forecasts.

This is a material model risk, not a contract failure. The proposed action is isolated feature/target training and backtesting, not blind nightly tuning.

### Info: `complete_slate_audit_coverage`

- Severity: **Info**.
- Evidence: 16/16 games structurally audited.
- Action: retain a per-game audit receipt.
- Guardrail: complete audit coverage does not cancel findings within audited games.

### Info: `simulation_inert_on_slate`

- Severity: **Info on this slate; escalated across slates**.
- Evidence: zero flip candidates and zero published flips on 16 games.
- Action: measure simulation contribution across rolling slates and separate score realization from winner flipping.
- Guardrail: zero winner flips is not automatically a defect and does not prove that simulation did no score reconciliation.

Because this finding has persisted for eight consecutive slates, it is now an established operational pattern requiring code/diagnostic action.

# 3. Every game

Notation: starter lines are projected ER/outs versus actual ER/outs. “Total error” is the absolute total-run error, equivalent to the absolute value of the listed total bias.

## ARI@HOU — correct winner, wrong score

- Prediction: HOU 6, ARI 4; HOU winner.
- Actual: HOU 3, ARI 1; HOU winner.
- Published HOU probability: **0.5825**; home-minus-away edge **+0.165**.
- Conviction: moderate.
- Brier: **0.174304**.
- Team-run error: **6**; total error/bias: **6 over**.
- Starters:
  - Merrill Kelly: 3 ER/15 outs projected; 3 ER/18 actual.
  - Cristian Javier: 4 ER/13 projected; 1 ER/15 actual.
- Edges: season −0.029, recent +0.444, pitcher +0.038, composite +0.539 home.
- Controls: raw 0.6153 → governed 0.6219; no recency guardrail; later published 0.5825. Score-volatility action: no change. Savant adjustment applied.
- Score path: not supplied beyond final published score.
- Player adjustments: none supplied.
- Audit: PASS, 0 critical/error/warning, 7 info; artifact clean.
- Causes: no directional miss; primary/secondary causes are not assigned. The score miss was largely the HOU starter allowing three fewer ER than projected and an overestimated total environment.
- Audit caveat: the calibrated-to-published probability delta is unexplained and critical for lineage, even though the winner was correct.

## ATL@PHI — correct winner and strong score call

- Prediction: ATL 4, PHI 3; ATL winner.
- Actual: ATL 5, PHI 2; ATL winner.
- Published ATL probability: **0.5599**; home-minus-away edge **−0.1199**.
- Conviction: moderate.
- Brier: **0.193658**.
- Team-run error: **2**; total error: **0**.
- Starters:
  - Chris Sale: 2 ER/19 projected; 2 ER/18 actual.
  - Cristopher Sánchez: 3 ER/19 projected; 3 ER/18 actual.
- Edges: season −0.446, recent −0.318, pitcher −0.120, composite −0.768 home.
- Controls: raw 0.6000 → governed 0.5815; recency guardrail applied; final published probability 0.5599. No volatility change; Savant adjustment applied.
- Score path: not supplied beyond final score.
- Player adjustments: none supplied.
- Audit: PASS, 0 critical/error/warning, 8 info.
- Causes: none material. This was a strong correct call; starter and total-run projections were accurate.

## BOS@BAL — directional miss and major score overprojection

- Prediction: BAL 6, BOS 4; BAL winner.
- Actual: BOS 1, BAL 0; BOS winner.
- Published BAL probability: **0.5300**; home-minus-away edge **+0.060**.
- Conviction: lean.
- Brier: **0.280900**.
- Team-run error: **9**; total error: **9 over**.
- Starters:
  - Ranger Suárez: 3 ER/15 projected; 0 ER/21 actual.
  - Shane Baz: 4 ER/17 projected; 1 ER/19 actual.
- Edges: season −0.256, recent +0.642, pitcher +0.009, composite +0.400 home.
- Controls: raw/governor input 0.5300; calibrated/governed 0.6241; recency guardrail applied; published probability 0.5300. Volatility action: supported blowout-margin expansion. Savant adjustment applied.
- Score path: historical 4.63–3.89 → variance 5–4 → enhanced 4–6 → Savant 4–6 → final 4–6. The winner changed from the historical away lean to the home side at enhanced prediction.
- Player adjustments: 21 material changes. Largest listed effects included Encarnacion-Strand, Rutschman, Roman Anthony, Alonso, Eli White, Abreu, Story, and Taveras.
- Audit: PASS, 0 critical/error/warning, 7 info. Player actuals unavailable.
- Primary cause: `run_environment_over`.
- Secondary cause: none formally assigned.
- Diagnosis: both starters materially outperformed projections, producing a nine-run overprojection. The recent/home evidence pulled the model toward BAL, but the actual game was a low-scoring BOS shutout.

## CHC@MIA — strong correct call

- Prediction: CHC 6, MIA 3; CHC winner.
- Actual: CHC 6, MIA 1; CHC winner.
- Published CHC probability: **0.7245**; home-minus-away edge **−0.449**.
- Conviction: strong.
- Brier: **0.075890**.
- Team-run error: **2**; total error: **2 over**.
- Starters:
  - Shota Imanaga: 0 ER/16 projected; 1 ER/18 actual.
  - Janson Junk: 4 ER/15 projected; 1 ER/15 actual.
- Edges: season −0.747, recent −0.408, pitcher −0.070, composite −1.206 home.
- Controls: raw 0.7034 → governed 0.7104; no recency guardrail; score-coherence publication 0.7245. Volatility actions: high-total-environment lift and supported blowout expansion.
- Score path: not supplied beyond final score.
- Player adjustments: none supplied.
- Audit: PASS, 0 critical/error/warning, 7 info.
- Causes: none material. This was a strong correct call, though the high-confidence probability remains historically aggressive given the cluster’s empirical hit rate near 0.55.

## DET@CLE Game 1 — correct winner, wrong score

- Prediction: CLE 4, DET 3; CLE winner.
- Actual: CLE 7, DET 6; CLE winner.
- Published CLE probability: **0.5251**; home-minus-away edge **+0.0502**.
- Conviction: lean.
- Brier: **0.225537**.
- Team-run error: **6**; total error: **6 under**.
- Starters:
  - Keider Montero: 2 ER/18 projected; 1 ER/17 actual; identity mismatch.
  - Logan Allen: 0 ER/9 projected; 3 ER/18 actual; identity mismatch.
- Edges: season +0.034, recent +0.830, pitcher +0.011, composite +0.431 home.
- Controls: raw 0.5220 → governed 0.5083; recency guardrail applied; published 0.5251. No volatility change.
- Score path: not supplied beyond final score.
- Player adjustments: none supplied.
- Audit: PASS, 0 critical/error/warning, 7 info.
- Causes: no formal primary/secondary cause. The correct side was retained, but CLE’s starter allowed three more ER than projected and both offenses produced more than forecast.
- Data caveat: starter identity mismatch reduces the evidentiary value of this starter comparison.

## DET@CLE Game 2 — correct winner and exact score

- Prediction: CLE 4, DET 3; CLE winner.
- Actual: CLE 4, DET 3; CLE winner.
- Published CLE probability: **0.5544**; home-minus-away edge **+0.1088**.
- Conviction: moderate.
- Brier: **0.198558**.
- Team-run error: **0**; total error: **0**.
- Starter lines: not supplied.
- Edges: season +0.034, recent +0.830, pitcher +0.059, composite +0.535 home.
- Controls: raw 0.5644 → governed 0.5662; recency guardrail applied; published 0.5544. No volatility change.
- Score path: not supplied beyond final score.
- Player adjustments: none supplied.
- Audit: PASS, 0 critical/error/warning, 7 info.
- Causes: none. Exact score and winner were correct.
- Coverage caveat: this is the one game without a starter comparison.

## LAA@PIT — correct winner, wrong score

- Prediction: PIT 7, LAA 5; PIT winner.
- Actual: PIT 1, LAA 0; PIT winner.
- Published PIT probability: **0.6620**; home-minus-away edge **+0.3241**.
- Conviction: strong.
- Brier: **0.114226**.
- Team-run error: **11**; total error: **11 over**.
- Starters:
  - Ryan Johnson: 7 ER/15 projected; 1 ER/18 actual.
  - Jared Jones: 5 ER/15 projected; 0 ER/18 actual.
- Edges: season +0.840, recent +0.648, pitcher +0.017, composite +1.020 home.
- Controls: raw 0.6441 → governed 0.6591; no recency guardrail; published 0.6620. No volatility change despite a large Savant swing flag.
- Score path: not supplied beyond final score.
- Player adjustments: none supplied.
- Audit: PASS, 0 critical/error/warning, 8 info.
- Causes: no directional cause; the score failure was a joint starter overprojection, especially Ryan Johnson by six ER and Jared Jones by five ER.
- Interpretation: a good winner forecast can still be a poor score forecast. This is the clearest example.

## MIL@CIN — correct winner and acceptable margin

- Prediction: MIL 8, CIN 5; MIL winner.
- Actual: MIL 10, CIN 7; MIL winner.
- Published MIL probability: **0.5400**; home-minus-away edge **−0.080**.
- Conviction: lean.
- Brier: **0.211600**.
- Team-run error: **4**; total error: **4 under**.
- Starters:
  - Shane Drohan: 1 ER/16 projected; 6 ER/12 actual.
  - Rhett Lowder: 5 ER/15 projected; 6 ER/12 actual.
- Edges: season −1.580, recent −0.654, pitcher −0.166, composite −1.938 home.
- Controls: raw 0.5400 → governed 0.6526; recency guardrail applied; final probability was pulled back to 0.5400 after score coherence/guardrail processing. Supported blowout expansion.
- Score path: not supplied beyond final score.
- Player adjustments: none supplied.
- Audit: PASS, 0 critical/error/warning, 7 info.
- Causes: none directional. The MIL side was correctly identified despite a five-ER starter underprediction for Drohan.

## MIN@CHW — directional miss

- Prediction: MIN 6, CHW 5; MIN winner.
- Actual: MIN 1, CHW 4; CHW winner.
- Published MIN probability: **0.5381**; home-minus-away edge **−0.0762**.
- Conviction: lean.
- Brier: **0.289555**.
- Team-run error: **6**; total error: **6 over**.
- Starters:
  - Zebby Matthews: 4 ER/18 projected; 4 ER/15 actual.
  - Erick Fedde: 5 ER/14 projected; 0 ER/15 actual.
- Edges: season +0.359, recent −0.046, pitcher −0.032, composite −0.065 home.
- Controls: raw 0.5317 → governed 0.5211; recency guardrail applied; published 0.5381. No volatility change.
- Score path: historical 4.81–4.01 → variance 5–4 → enhanced 6–5 → Savant 6–5 → final 6–5. The model increased MIN’s scoring after the enhanced stage.
- Player adjustments: 24 material changes, including Royce Lewis, Benintendi, Sam Antonacci, Keaschall, Murakami, Montgomery, Vargas, and Jeffers.
- Audit: PASS; player actuals unavailable.
- Primary cause: `starter_projection`, chiefly Fedde allowing five fewer ER than projected.
- Secondary cause: `run_environment_over`.
- Earliest wrong-direction evidence: the weak composite edge was already near neutral, but the enhanced stage moved the score to MIN 6–5; the Savant stage retained that direction rather than correcting it.

## NYY@SD — high-conviction directional miss

- Prediction: NYY 6, SD 3; NYY winner.
- Actual: NYY 2, SD 3; SD winner.
- Published NYY probability: **0.6434**; home-minus-away edge **−0.2869**.
- Conviction: strong.
- Brier: **0.414021**, worst on the slate.
- Team-run error: **4**; total error: **4 over**.
- Starters:
  - Max Fried: 0 ER/17 projected; 1 ER/17 actual.
  - Walker Buehler: 4 ER/16 projected; 1 ER/15 actual.
- Edges: season −0.727, recent −0.544, pitcher −0.007, composite −1.091 home.
- Controls: raw 0.6290 → governed 0.6219; no recency guardrail; published probability 0.6434. No volatility change despite a large Savant swing flag.
- Score path: historical 4.46–2.99 → variance 4–3 → enhanced 5–3 → Savant 6–3 → final 6–3. The Savant stage added one NYY run and preserved a strong away margin.
- Player adjustments: 22 material changes, including McMahon, Bellinger, France, Spencer Jones, Wells, Jazz Chisholm Jr., Tatis Jr., and Merrill.
- Audit: PASS; player actuals unavailable.
- Primary cause: `wrong_conviction_side`.
- Secondary cause: `lean_side_miss` in the register-style characterization, although the game was actually classified strong.
- Additional factor: Buehler allowed three fewer ER than projected; the run environment was four runs lower than forecast.
- Earliest wrong-direction evidence: the historical model already favored NYY, and enhanced/Savant stages increased the away score. No later stage recovered toward SD. A 0.29 edge is strong relative to this slate, but it was not supported by the outcome.

## OAK@SEA — lean directional miss

- Prediction: SEA 6, OAK 5; SEA winner.
- Actual: OAK 7, SEA 6; OAK winner.
- Published SEA probability: **0.5485**; home-minus-away edge **+0.097**.
- Conviction: lean.
- Brier: **0.300850**.
- Team-run error: **2**; total error: **2 under**.
- Starters:
  - Kade Morris: 5 ER/9 projected; 5 ER/8 actual.
  - Logan Gilbert: 1 ER/17 projected; 7 ER/14 actual.
- Edges: season +0.625, recent +0.158, pitcher +0.034, composite +0.639 home.
- Controls: raw 0.5600 → governed 0.5499; recency guardrail applied; published probability 0.5485. No volatility change.
- Score path: historical 3.16–6.15 → variance 3–6 → enhanced 5–6 → Savant 5–6 → final 5–6. The model never changed away from SEA.
- Player adjustments: 25 material changes, including Heim, Morris, Canzone, Alvarado, Julio Rodríguez, Muncy, Cortes, and McNeil.
- Audit: PASS; player actuals unavailable.
- Primary cause: `starter_projection`.
- Secondary cause: none.
- The decisive miss was Gilbert allowing six more ER than projected. This is a starter-tail failure more than a broad team-run failure.

## SF@NYM — correct winner, wrong score

- Prediction: NYM 5, SF 4; NYM winner.
- Actual: NYM 10, SF 6; NYM winner.
- Published NYM probability: **0.5468**; home-minus-away edge **+0.0937**.
- Conviction: lean.
- Brier: **0.205362**.
- Team-run error: **7**; total error: **7 under**.
- Starters:
  - Matt Wilkinson: 3 ER/12 projected; 4 ER/9 actual.
  - Nolan McLean: 2 ER/18 projected; 2 ER/18 actual.
- Edges: season +0.016, recent −0.185, pitcher −0.075, composite −0.024 home.
- Controls: raw 0.5559 → governed 0.5258; no recency guardrail; final published probability 0.5468. No volatility change despite a large Savant swing flag.
- Score path: not supplied beyond final score.
- Player adjustments: none supplied.
- Audit: PASS, 0 critical/error/warning, 8 info.
- Causes: no directional miss. NYM offense was underprojected by five runs; Wilkinson was one ER worse than projected.
- Critical lineage issue: the published-minus-calibrated delta is unexplained.

## STL@COL — correct winner and reasonable score

- Prediction: STL 6, COL 5; STL winner.
- Actual: STL 7, COL 6; STL winner.
- Published STL probability: **0.5485**; home-minus-away edge **−0.097**.
- Conviction: lean.
- Brier: **0.203854**.
- Team-run error: **2**; total error: **2 under**.
- Starters:
  - Andre Pallante: 2 ER/18 projected; 1 ER/15 actual.
  - Ryan Feltner: 6 ER/14 projected; 4 ER/15 actual.
- Edges: season −0.715, recent −0.113, pitcher +0.034, composite −0.527 home.
- Controls: raw 0.5600 → governed 0.5499; recency guardrail applied; published 0.5485. No volatility change.
- Score path: not supplied beyond final score.
- Player adjustments: none supplied.
- Audit: PASS, 0 critical/error/warning, 7 info.
- Causes: none. The STL direction and close margin were sound.

## TB@TEX — moderate directional miss

- Prediction: TEX 4, TB 3; TEX winner.
- Actual: TB 7, TEX 6; TB winner.
- Published TEX probability: **0.5599**; home-minus-away edge **+0.1199**.
- Conviction: moderate.
- Brier: **0.313526**.
- Team-run error: **6**; total error: **6 under**.
- Starters:
  - Nick Martinez: 2 ER/18 projected; 3 ER/19 actual.
  - Trevor Williams: 1 ER/14 projected; 0 ER/3 actual.
- Edges: season −0.497, recent 0.000, pitcher −0.041, composite −0.183 home.
- Controls: raw 0.5614 → governed 0.5815; recency guardrail applied; published 0.5599. No volatility change.
- Score path: historical 3.43–4.70 → variance 3–5 → enhanced 3–4 → Savant 3–4 → final 3–4. The model remained on TEX despite the negative composite home edge.
- Player adjustments: 23 material changes, including Hicks, Martinez, Mesa Jr., Simpson, Nimmo, Pederson, Walls, and Langford.
- Audit: PASS; player actuals unavailable.
- Primary cause: `run_environment_under`.
- Secondary cause: `wrong_conviction_side`.
- The published edge was moderate, not a near coin flip: +0.1199 is a meaningful but limited signal. It should not be treated as confidence-free ambiguity.

## TOR@KC — moderate directional miss and major score miss

- Prediction: KC 7, TOR 5; KC winner.
- Actual: TOR 9, KC 2; TOR winner.
- Published KC probability: **0.5571**; home-minus-away edge **+0.1141**.
- Conviction: moderate.
- Brier: **0.310311**.
- Team-run error: **9**; total error: **1 over**.
- Starters:
  - Jameson Taillon: 4 ER/14 projected; 1 ER/6 actual.
  - Daniel Lynch IV: 1 ER/15 projected; 5 ER/15 actual.
- Edges: season −0.234, recent +0.165, pitcher −0.039, composite +0.186 home.
- Controls: raw 0.5600 → governed 0.5499; recency guardrail applied; published 0.5571. No volatility change.
- Score path: historical 3.61–5.80 → variance 4–6 → enhanced 5–6 → Savant 5–7 → final 5–7. The Savant stage increased KC’s projected lead.
- Player adjustments: 25 material changes, including Carter Jensen, Springer, Kirk, Salvador Perez, Isbel, Loftin, Straw, and Caglianone.
- Audit: PASS; player actuals unavailable.
- Primary cause: `starter_projection`.
- Secondary cause: `wrong_conviction_side`.
- Taillon allowed three fewer ER than projected, while Lynch allowed four more. TOR also scored four more runs than projected and KC scored five fewer.
- The 0.1141 home edge is moderate, not a near coin flip.

## WAS@LAD — strong correct call

- Prediction: LAD 5, WAS 4; LAD winner.
- Actual: LAD 5, WAS 3; LAD winner.
- Published LAD probability: **0.5615**; home-minus-away edge **+0.123**.
- Conviction: moderate.
- Brier: **0.192284**.
- Team-run error: **1**; total error: **1 over**.
- Starters:
  - Jackson Kent: 4 ER/13 projected; 3 ER/15 actual.
  - Blake Snell: 2 ER/16 projected; 0 ER/18 actual.
- Edges: season +0.629, recent +0.042, pitcher −0.021, composite +0.490 home.
- Controls: raw 0.5541 → governed 0.5603; no recency guardrail; published 0.5615. No volatility change.
- Score path: not supplied beyond final score.
- Player adjustments: none supplied.
- Audit: PASS, 0 critical/error/warning, 7 info.
- Causes: none. This was a strong winner and score call.

# 4. Miss reconstruction

## Earliest wrong-direction stage for every loss

### BOS@BAL

- Historical ML: 4.63 BOS–3.89 BAL, still away-favored.
- Variance stage: 5–4, still away-favored.
- Enhanced prediction: 4–6, moved to BAL.
- This is the earliest clear winner-direction regression.
- Evidence moved wrong because recent form and matchup adjustments outweighed the negative season edge and near-neutral pitcher edge.
- Main failure type: bad weighting / run-environment overprojection.
- Starter miss then amplified the score error: both starters outperformed projections.
- Irreducible variance remains possible, but the nine-run total overprojection is a systematic run-model concern.

### MIN@CHW

- Historical ML: MIN 4.81–CHW 4.01.
- Variance stage: MIN 5–CHW 4.
- Enhanced stage: MIN 6–CHW 5.
- The wrong direction was present before the final stages; Savant and later stages did not recover.
- Main failure types: starter error, bad weighting, and run-environment overprojection.
- Fedde’s projected five ER versus actual zero was decisive.
- The composite edge was only −0.0653 home, so the model had weak directional separation and should not have produced a strong margin.

### NYY@SD

- Historical ML already favored NYY 4.46–2.99.
- Variance stage retained NYY 4–3.
- Enhanced stage increased NYY to 5–3.
- Savant stage increased NYY to 6–3.
- There was no later recovery; the winner-direction error was established early and reinforced.
- Main failure types: bad input/feature weighting, overconfidence, starter miss, and irreducible variance.
- The model’s season/recent/composite evidence strongly favored NYY, but the actual game was decided by a one-run margin.
- Buehler’s actual one ER versus projected four reduced the predicted NYY advantage.

### OAK@SEA

- Historical ML favored SEA 6.15–3.16.
- Every later score stage retained SEA as winner.
- No stage introduced a new winner-direction regression; the error originated in the initial evidence and was not corrected.
- Main failure type: starter projection error.
- Gilbert’s actual seven ER versus projected one was the key reversal.
- This is not primarily a probability-publication issue; published probability closely matched the calibrated/governed value.

### TB@TEX

- Historical ML favored TEX 4.70–3.43.
- Variance stage retained TEX.
- Enhanced stage narrowed the score to TB 3–TEX 4 but did not flip.
- All later stages retained TEX.
- Main failure types: bad weighting and run-environment underdispersion, with a smaller starter contribution.
- The pregame composite edge actually favored TB by −0.1833 home, so the model’s home pick was not strongly supported by its own composite evidence.
- Actual scoring was six runs above the projected total.

### TOR@KC

- Historical ML favored KC 5.80–3.61.
- Variance and enhanced stages retained KC.
- Savant moved KC to 7 runs and reinforced the wrong side.
- Main failure types: starter projection, bad weighting, and directional overcommitment.
- Lynch’s actual five ER versus projected one and Taillon’s one versus projected four reversed the expected matchup.
- The final margin error was nine runs.
- The moderate +0.1141 home edge should not be interpreted as a near coin flip, but it was also not strong enough to justify ignoring the contradictory season and pitcher signals.

## Required distinction among miss types

- **Bad input:** possible in the underlying player, lineup, park, weather, or pitcher features, but not directly proven. No source-labelled market data was available, and several player actuals were absent.
- **Bad weighting:** demonstrated in cases where recent/Savant adjustments moved a weak or contradictory matchup into a stronger wrong-side prediction, especially BOS@BAL, MIN@CHW, and TOR@KC.
- **Calibration:** demonstrated by 0.1316 weighted calibration error and 33.3% high-conviction miss rate, especially NYY@SD.
- **Starter error:** directly evidenced in MIN@CHW, OAK@SEA, and TOR@KC, and materially present in BOS@BAL.
- **Offense/bullpen miss:** present in BOS@BAL, SF@NYM, and TOR@KC, but bullpen-specific causal separation is not supplied.
- **Run-environment miss:** overdispersion failure in BOS@BAL, MIN@CHW, and NYY@SD; underdispersion in TB@TEX and missed scoring in SF@NYM.
- **Irreducible variance:** remains plausible in close games such as NYY@SD and TB@TEX, but cannot explain away the persistent run-model and simulation evidence.

## Pipeline-stage quality across all stages and transitions

| Stage | Winner accuracy | Team-run MAE | Total-run MAE | Margin MAE |
|---|---:|---:|---:|---:|
| Historical ML | 10/16, 0.625 | 2.4463 | 4.2626 | 2.2201 |
| Variance applied | 8/16, 0.500 | 2.4688 | 4.3125 | 2.1875 |
| Enhanced prediction | 10/16, 0.625 | 2.3438 | 4.0625 | 2.0625 |
| Embedded Savant run environment | 10/16, 0.625 | 2.4062 | 4.1875 | 2.0625 |
| Fundamental edge review | 10/16, 0.625 | 2.4062 | 4.1875 | 2.0625 |
| Capped target | 10/16, 0.625 | 2.4062 | 4.1875 | 2.0625 |
| Simulation result | 10/16, 0.625 | 2.4062 | 4.1875 | 2.0625 |
| Published final | 10/16, 0.625 | 2.4062 | 4.1875 | 2.0625 |

Important transitions:

1. **Historical ML → variance applied**
   - Winner accuracy fell by 0.125, from 10 to 8 hits.
   - Changed direction: DET@CLE tie, MIL@CIN tie, STL@COL tie.
   - Regressed games: MIL@CIN and STL@COL.
   - Margin MAE improved slightly, but team-run MAE worsened by 0.0225.
   - A later recovery does not erase this intermediate winner-direction regression.

2. **Variance applied → enhanced prediction**
   - Winner accuracy recovered from 8 to 10.
   - Improved games: DET@CLE, MIL@CIN, STL@COL.
   - Regressed game: BOS@BAL.
   - Margin and team-run MAE improved by 0.125.
   - This recovery does not prove the enhanced stage is universally beneficial; it traded one BOS@BAL directional failure against three recoveries.

3. **Enhanced → embedded Savant**
   - No winner changes.
   - Team-run MAE worsened from 2.3438 to 2.4062.
   - No score-stage improvement in aggregate.

4. **Embedded Savant → fundamental review → capped target → simulation → published**
   - No winner changes.
   - No later score-stage improvement after embedded Savant.
   - Simulation made no measured winner intervention.

The pipeline therefore shows a useful enhanced-stage aggregate recovery after variance damage, but also a measurable regression at the Savant transition and no evidence that simulation contributed independent winner information.

# 5. Cross-game counterfactuals

## Lever tension 1: widen run distributions

Lever: `score.variance_multiplier / score.variance_floor`.

- Helps:
  - NYY@SD: lost a 0.29-confidence pick; wider tails could reduce overconfidence.
  - TB@TEX: lost a 0.12-confidence pick; wider tails could better represent high scoring.
  - TOR@KC: lost a 0.11-confidence pick; wider tails could reduce narrow-score certainty.
- Hurts:
  - CHC@MIA: strong correct call at 0.45 edge; widening could soften a good separation.
  - LAA@PIT: strong correct call at 0.32 edge; widening could reduce useful confidence.
- Net evidence: 3 helps versus 2 hurts, but the lever targets Brier/calibration and score distribution, not direct winner accuracy. A small bounded change or shadow experiment is more defensible than a broad widening.

## Lever tension 2: pitcher form weights and prior games

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

- Helps:
  - ARI@HOU, BOS@BAL, CHC@MIA, DET@CLE Game 1, LAA@PIT, MIL@CIN, MIN@CHW, NYY@SD, OAK@SEA, TOR@KC.
- Hurts:
  - ATL@PHI, SF@NYM, TB@TEX.
- Net evidence: 10 helps versus 3 hurts.
- Caution: several “help” cases are correct-winner games with poor score predictions, so this lever may improve starter ER MAE without necessarily improving winner accuracy. DET@CLE Game 1 also has identity mismatches.

## Lever tension 3: recent/season run-mean weighting

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

- Helps:
  - ARI@HOU, BOS@BAL, DET@CLE Game 1, LAA@PIT, MIL@CIN, MIN@CHW, NYY@SD, SF@NYM, TB@TEX, TOR@KC.
- Hurts:
  - ATL@PHI, CHC@MIA, DET@CLE Game 2, OAK@SEA, STL@COL, WAS@LAD.
- Net evidence: 10 helps versus 6 hurts.
- Net support is weak-to-moderate because “helps” includes several losses where the lever might improve runs while not fixing the winner.

## One-game fixes that must not be made

- Do not globally reverse or suppress NYY@SD based on one high-conviction loss; that would damage CHC@MIA and LAA@PIT-style strong correct calls.
- Do not force every moderate edge toward the road team because TOR@KC and TB@TEX lost; ATL@PHI, MIL@CIN, and WAS@LAD show moderate directional calls can work.
- Do not broadly increase variance solely because of BOS@BAL; CHC@MIA and LAA@PIT were correct strong calls.
- Do not reduce Savant influence solely because of the NYY@SD or TOR@KC losses; the aggregate transition was neutral on winners and improved over historical ML on margin MAE, though it worsened team-run MAE relative to enhanced prediction.
- Do not tune exact-score tails before addressing the near-zero run-model signal and the inert simulation contract.

# 6. Data and contract defects

## Critical findings

### Unexplained probability publication deltas

Affected games:

- ARI@HOU:
  - Calibrated/governed: 0.6219.
  - Published: 0.582503.
  - Published minus calibrated: −0.039397.
  - Score-coherence probability: 0.603129.
  - Transformation recorded as: “no recorded transformation explains the publication delta.”
- SF@NYM:
  - Calibrated/governed: 0.5258.
  - Published: 0.546831.
  - Published minus calibrated: +0.021031.
  - Score-coherence probability: 0.555095.
  - Same missing-transformation finding.

These are material lineage gaps. The data does not prove the governor was bypassed; it proves that the complete transformation is not recorded.

## Error findings

No error findings were reported in the supplied structural-audit summary or per-game audits.

## Warning findings

No warning findings were reported.

## Informational findings

- 117 total structural-audit info findings.
- All 16 games structurally audited.
- Simulation produced zero flip candidates and zero published flips.
- Various games had informational notes, but no per-game info item was supplied as a prediction-contract failure.
- Six games lacked player-level actuals, limiting causal attribution.

## Prediction-model defects versus artifact bugs

### Prediction-model/process defects

1. Near-zero run-model signal:
   - Total R² 0.026024.
   - Home R² 0.018723.
   - Away R² 0.026352.
2. High calibration error:
   - Weighted absolute calibration error 0.1316.
3. Score-margin compression:
   - 13/16 predicted margins ≤2, while three actual margins were ≥4.
4. Starter projection instability:
   - Three losses primarily caused by starter projections.
5. Simulation inertness:
   - Zero winner interventions on this slate and an eight-slate persistence streak.

### Presentation/artifact/contract defects

1. Missing integrity-revision audit artifact.
2. Unexplained final probability transformations for ARI@HOU and SF@NYM.
3. Duplicate matchup labeling in the causal register for DET@CLE without explicit game-ID distinction.
4. Incomplete starter comparison coverage for DET@CLE Game 2.
5. Missing player-level actuals for six games.

The game artifacts themselves were structurally clean. The defects are chiefly lineage, receipt, coverage, and process-contract issues rather than corrupted score files.

## Coverage reconciliation

The slate is not fully “clean” merely because 16/16 structural audits passed:

- Structural coverage is complete.
- Starter coverage is 15/16.
- Probability lineage is materially incomplete in 2/16.
- Integrity-revision receipt is missing.
- Player-level causal coverage is incomplete in 6/16.
- Simulation diagnostics are present but demonstrate an eight-slate inert pattern.

## Complete pick-side probability lineage

The normal lineage is:

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

Examples of complete recorded transformations:

- CHC@MIA:
  - Raw 0.7374.
  - Governor input 0.703355.
  - Governed 0.7104.
  - Score coherence/published 0.724519.
  - Transformation explicitly recorded.
- MIL@CIN:
  - Raw 0.6628.
  - Governor input 0.5400.
  - Governed 0.6526.
  - Score coherence 0.658802.
  - Published 0.5400 after matchup-recency guardrail.
  - Transformation explicitly recorded.
- OAK@SEA:
  - Raw 0.5510.
  - Governor input 0.5600.
  - Governed 0.5499.
  - Published 0.548498.
  - Transformation says published probability matches calibrated/governed value.

A material bypass is not declared for any game because the supplied evidence does not show an unrecorded material transformation except the two critical cases. For ARI@HOU and SF@NYM, the correct finding is **unreconciled transformation**, not proven governor bypass.

# 7. Cross-slate persistence

Seven prior slates were loaded, producing an eight-slate window from 2026-08-28 through 2026-09-04.

## Finding streaks

| Finding | Consecutive-slate count | Present now? | Interpretation |
|---|---:|---:|---|
| High raw probability cluster | 0 | No | Historical incidence in 2/8 slates; not current |
| Pipeline winner instability | 0 | No | Incidence in 3/8; not current |
| Score-margin compression | 0 | No | Incidence in 3/8; not current |
| Simulation inert on slate | 8 | Yes | Escalated established pattern |
| Single-slate calibration instability | 0 | No | Incidence in 5/8, but no current streak |
| Weak run-model signal | 8 | Yes | Escalated established pattern |

The two eight-slate findings must be treated as established failures, not single-slate noise.

## Escalated finding: simulation inertness

- Streak: 8 consecutive slates.
- Target lever: `simulation.allow_score`.
- Target metrics: team-run MAE and total-run MAE.
- Parameter actionability: false.
- Direction: confirm the simulation module is running and not silently no-oping.
- Evidence: zero winner flips and zero flip candidates on the current 16-game slate; self-assessment identifies the same-draw target-realization/early-exit design as the root cause.
- Correct action: code/diagnostic repair, not a forced parameter adjustment.

The supplied self-assessment says the simulation reuses draws collected while searching for a target-realizing play-by-play and may exit after roughly two selected draws. Such a sample cannot support a probability or flip/agreement signal.

## Escalated finding: weak run-model signal

- Streak: 8 consecutive slates.
- Target levers listed: `training.xgb.n_estimators`, `training.xgb.max_depth`, `training.xgb.learning_rate`.
- Target metrics: team-run MAE and total-run MAE.
- Parameter actionability: false in the escalation record.
- Direction: investigate features and target via isolated training backtest.
- Evidence: current holdout R² remains near zero across away, home, and total runs.
- Correct action: feature/target contract work, retraining, and same-fingerprint holdout comparison—not blind nightly hyperparameter tuning.

## Top-confidence-pick loss streak

- Reported loss streak: **0**.
- Per-slate results: hit, hit, miss, hit, miss, miss, miss, hit.
- The current slate’s top-confidence pick, CHC@MIA at 0.7245 away probability, was a hit.
- There is no current top-confidence loss streak to address.

## Metric trends

- Margin bias: **improving**. Series moved from −1.333 to −0.688; mean −2.1959.
- Mean team-run error on losses: **flat**. Mean 6.1906; current slate 6.0.
- Weighted calibration error: **improving**. Series ended at 0.1316 from 0.2179, though still materially high.
- Winner accuracy: **improving**. Current slate 0.625; eight-slate series ends at 0.625, mean 0.5694.
- Top-confidence loss streak: 0, therefore no regressing streak exists.

Improving trends do not neutralize the escalated eight-slate run-model and simulation defects.

# 8. Adjustments for today

Recommendations are ranked by persistence first, then scope, expected metric impact, and regression risk.

## 1. Repair and independently validate simulation — highest evidence strength

- Pattern: simulation inertness, eight consecutive slates.
- Registered lever: `simulation.allow_score`, but this is a diagnostic/code target rather than a safe parameter move.
- Change: **do not force a numeric parameter adjustment; enable and validate an independent unsteered simulation batch** for win-probability estimation. The batch should not early-exit while searching for a target-realizing score.
- Expected metrics: improve the validity of winner-probability contribution; potentially improve team-run MAE and total-run MAE. Direction expected: lower MAE and nonzero, interpretable winner-intervention rate.
- Helps: games with large score-path disagreement or high volatility, especially NYY@SD, TB@TEX, TOR@KC, and the large-swing cases.
- Hurts/risk: could soften strong correct calls such as CHC@MIA and LAA@PIT if the independent distribution is noisier; rollback if rolling Brier or winner accuracy worsens.
- Required action: code/diagnostic repair in the simulation builder and draw loop, followed by a shadow comparison. Do not promote until the missing distribution contribution is measurable.

## 2. Run an isolated run-model feature/target backtest — persistent pattern

- Pattern: weak run-model signal, eight consecutive slates.
- Registered targets: `training.xgb.n_estimators`, `training.xgb.max_depth`, `training.xgb.learning_rate`.
- Recommended bounded change: **no direct nightly hyperparameter move**. Perform the registered isolated training experiment after adding or validating park/weather environment, bullpen rest/fatigue, platoon-adjusted lineup quality, and properly source-labelled no-vig market information where available.
- Expected metrics: increase away/home/total holdout R² and reduce team-run MAE and total-run MAE.
- Helps: broad run-error games such as BOS@BAL, LAA@PIT, MIN@CHW, SF@NYM, and TOR@KC.
- Hurts/risk: could overfit and damage low-error games such as ATL@PHI, CHC@MIA, DET@CLE Game 2, STL@COL, and WAS@LAD. Require same-fingerprint holdout comparison and rollback on degraded total-run MAE or calibration.
- This is a code/data/training action, not a forced parameter adjustment.

## 3. Apply a small shadow change to pitcher prior weighting — strongest current single-slate driver

- Pattern: starter projection was the primary cause in three of six losses: MIN@CHW, OAK@SEA, TOR@KC.
- Registered parameter: `player.pitcher_prior_games` within `player.pitcher.v4` form weighting.
- Change: **small increase in prior-history weight / modestly increase prior games**, preferably in shadow mode. The exact numeric magnitude is not supplied in the evidence, so the change should be the smallest registered configuration increment available, not an invented large shift.
- Expected metric: lower `starter_er_mae` and `starter_outs_mae`; directional improvement expected secondarily.
- Helps: the 10 games listed by the lever-tension record, including the three primary starter-projection losses.
- Hurts/risk: ATL@PHI, SF@NYM, and TB@TEX had accurate starter projections; excessive reweighting could disturb them. Roll back if starter MAE improves but winner Brier or total-run MAE degrades.
- This is a bounded shadow experiment, not a full promotion.

## 4. Small confidence/variance calibration shadow

- Pattern: current weighted calibration error 0.1316 and high-conviction miss rate 33.3%.
- Registered parameters: `confidence.base_cap`, `confidence.moderate_cap`, `score.variance_multiplier`, `score.variance_floor`.
- Change: **lower confidence caps slightly and widen run variance slightly**, in shadow mode only.
- Expected metrics: lower Brier score, lower calibration error, and reduce overconfident tail behavior; score-margin spread should become less compressed.
- Helps: NYY@SD, TB@TEX, and TOR@KC are the direct tension beneficiaries.
- Hurts/risk: CHC@MIA and LAA@PIT were strong correct calls that could be softened; monitor high-confidence hit rate and calibration by bucket.
- Because the high-raw cluster is not present on this slate and its current result was 3/3, this recommendation is weaker than the persistent code/model actions.

## Changes explicitly not recommended today

- Do not force `score.variance_multiplier` to a large value based on NYY@SD alone.
- Do not alter winner-side selection rules based solely on the moderate TB@TEX or TOR@KC losses.
- Do not promote a probability-calibration change until ARI@HOU and SF@NYM lineage is reconciled.
- Do not claim simulation is fixed merely because a future slate produces a winner flip; verify independent draws and rolling metrics.

# 9. Bottom line

## Known

- The slate went 10–6, or 62.5%, above the supplied walk-forward baseline.
- Score quality was materially weaker than winner accuracy: 13/16 predicted margins were ≤2, while three actual margins were at least four runs.
- Six losses had mean team-run error of six runs.
- Starter projections were the primary labelled cause in three losses.
- Run-model holdout R² is near zero across all run targets and has persisted as an escalated eight-slate problem.
- Simulation inertness has persisted for eight slates and is a code/diagnostic concern, not a parameter-tuning opportunity.
- Every structural game audit passed, but probability lineage is not complete for ARI@HOU and SF@NYM.
- The integrity-revision artifact is missing.

## Uncertain

- Whether recent/Savant weighting is causally responsible for the individual wrong-side calls; lever-tension counts are observational.
- Whether the best pitcher adjustment is more prior history, less recent-form weight, or improved pitcher features.
- Whether widening variance improves future Brier/calibration without weakening strong correct calls.
- The extent to which bullpen, weather, park, lineup, and other absent or imperfect inputs contributed to the run misses.
- Exact score accuracy cannot be summarized as a rate from the supplied fields because no slate-wide exact-score metric is provided.

## Single safest next action

**Block promotion and repair the probability/simulation contracts first: reconcile the ARI@HOU and SF@NYM pick-side probability transformations, then run an independent non-early-exit simulation in shadow mode while preserving the current production output.**

That action addresses the only critical contract defect, tests the eight-slate simulation failure without risking the next slate, and creates the evidence needed before changing model weights or confidence caps.