Luna deep analysis — slate 20260905

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

1. Evidence receipt and gaps

Sources received

All seven source_manifest entries were reviewed:

1. Prediction review — loaded; 376,581 bytes; official outcomes and scoring errors.
2. Forensic deep audit — loaded; 241,224 bytes; probability, provenance, starter, and decision contracts.
3. Continuous learning — loaded; 2,499,160 bytes; stage quality, calibration, probability lineage, score distribution, and starter comparisons.
4. Detailed report — loaded; 355,886 bytes; full game and simulation rendering.
5. Ultra-detailed player audit — loaded; 30,881,727 bytes; player-level prediction trace.
6. Structural audit — loaded; 1,848,251 bytes; invariant findings.
7. Integrity revision audit — missing; path /app/output/reports/audit_report_20260905_integrity_revision.html; intended to provide post-remediation proof for immutable revision artifacts.

Slate coverage

The evidence reports:

• Expected games: 15
• Games reviewed: 15
• Canonical artifacts loaded: 15
• Continuous-learning games: 15
• Forensic deep-audit games: 15
• Starter comparisons: 15
• Structural audits completed: 15
• All games structurally audited: true
• Slate result: 8 hits, 7 misses
• Learning eligibility: eligible, because all game artifacts passed critical/error structural gates.

The 15 reviewed games were:

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

Gaps and contradictions

• The integrity revision audit is missing. This prevents independent confirmation of the immutable revision-artifact remediation claim.
• The player feed had no rows for the entire slate. Consequently, player-level actual-result joins are unavailable for every game. Player-level adjustments are present, but actual player-performance attribution is not independently available.
• No source-labelled moneyline was available. Market agreement was therefore not testable.
• Probability lineage is complete in row coverage, but not contractually complete: three games have material calibrated-to-published deltas without recorded transformations.
• SF@NYM is marked unexplained_material_delta: false despite its transformation text saying “no recorded transformation explains the publication delta.” Its delta is 1.8578 percentage points, below the causal register’s material threshold of 2 points, so this is a documentation weakness rather than one of the three registered critical failures.
• The “high raw probability” finding is only a one-slate finding: 1 hit in 3 games. It is not sufficient by itself to justify a permanent shrinkage rule.
• The simulation produced zero winner flips on this slate, and that behavior has persisted for nine slates. It is therefore not merely a one-slate observation.

2. Executive diagnosis

Overall result

The slate achieved:

• Winner accuracy: 8/15 = 53.33%
• Walk-forward validated baseline: 55.85%
• Historical model winner accuracy: 59.57% in the broader validation block
• Slate directional cover rate: 8/15 = 53.33%
• Weighted absolute calibration error: 0.100
• High-conviction misses: 2/5 = 40%
• Mean team-run error on losses: 5.286 runs
• Mean total-run bias on losses: +1.286 runs
• Model margin MAE: 3.333 runs
• Model margin signed bias: +1.333 runs
• Exact score quality: weak and materially less reliable than winner quality.

A correct winner does not imply a good prediction. For example:

• TOR@KC correctly selected Toronto but predicted 7–5 versus the actual 4–3; this is explicitly classified as right_winner_wrong_score.
• BOS@BAL correctly selected Boston but overestimated the total by four runs.
• WAS@LAD correctly selected Los Angeles but underpredicted the total by four runs.

Winner accuracy

Winner performance was weak:

• 8 hits and 7 misses.
• It underperformed the 55.85% walk-forward baseline.
• Cross-slate winner accuracy is marked regressing, with mean 56.54% across the nine-slate window and the latest slate at 53.33%.
• The variance stage materially damaged winner direction: historical_ml → variance_applied reduced accuracy from 8/15 to 5/15, a -0.2000 change.
• Enhanced prediction later recovered three winner hits, but recovery does not erase the intermediate regression.

Calibration quality

Calibration was mixed rather than uniformly defective:

• Weighted absolute calibration error: 0.100, improving across the nine-slate window.
• The 0.50–0.55 bucket was overconfident: average confidence 0.5302 versus a 0.500 hit rate.
• The 0.55–0.60 bucket was underconfident: 0.5659 versus 0.6667 observed.
• The 0.60–0.65 bucket had one game, confidence 0.6357, and lost.
• The 0.70–0.80 bucket had one game and won.
• The three-game high-raw cluster went 1–2: CHC@MIA hit, LAA@PIT and MIL@CIN missed.
• The data supports a bounded calibration shadow experiment, but not an aggressive permanent shrinkage rule from this slate alone.

Score accuracy

Score prediction is the clearest persistent model weakness:

• Holdout run-model R² was only:
• Away runs: 0.020313
• Home runs: 0.010614
• Total runs: 0.017602
• The model predicted margins of 2 runs or fewer in 13 of 15 games, while six actual margins were at least four runs.
• The pipeline’s embedded_savant_run_environment stage had:
• Team-run MAE 2.000
• Total-run MAE 3.200
• Margin MAE 3.333
• Winner accuracy 8/15
• Relative to historical_ml, this was a regression of:
• Team-run MAE: +0.1271
• Total-run MAE: not separately delta-listed at that stage, but total MAE was 3.2
• Margin MAE: +0.2750
• The persistent nine-slate finding is not “variance is slightly mis-tuned”; it is that the run model carries almost no validated predictive signal.

Starter accuracy

Starter projections were a mixed contributor:

• Some misses were driven by severe starter errors:
• LAA@PIT: Yusei Kikuchi projected for 6 ER, allowed 0.
• OAK@SEA: Jeffrey Springs projected for 5 ER, allowed 0.
• SF@NYM: Zac Thornton projected for 1 ER, allowed 6.
• Other games had starter errors of only one run, including CHC@MIA, MIL@CIN, MIN@CHW, STL@COL, and TOR@KC.
• The lever tension was exactly balanced: five games helped versus five games hurt. This does not support a broad starter-weight change without a more targeted experiment.

Artifact integrity

Structural artifact integrity was strong for the loaded game artifacts:

• Structural audit summary: 0 critical, 0 error, 0 warning, 111 info.
• Every game’s structural audit passed.
• The saved forecast artifact was explicitly reported as not being the reason for the individual misses.
• However, the missing integrity revision audit remains a receipt gap.
• The three critical probability-lineage cases remain critical contract failures even though the structural artifact audits passed.

Causal issue register reconciliation

Critical: unexplained probability publication delta

• Severity: critical.
• Evidence: three games have unexplained calibrated-to-published deltas of at least two percentage points.
• Affected games: ATL@PHI, LAA@PIT, WAS@LAD.
• Action: block promotion until the missing transformation is recorded and reconciled.
• Guardrail: do not assume the probability governor was bypassed until pick-side math is traced.

The evidence confirms the deltas:

• ATL@PHI: calibrated 0.5050, published 0.537979; delta +3.2979 pp.
• LAA@PIT: calibrated 0.6591, published 0.635729; delta -2.3371 pp.
• WAS@LAD: calibrated 0.6219, published 0.573431; delta -4.8469 pp.

High: intermediate winner instability

• Severity: high.
• Evidence: historical_ml → variance_applied reduced winner accuracy by 0.2000, or three hits.
• Affected games: registered as NYY@SD, SF@NYM, TB@TEX; the transition also changed MIL@CIN to a tie.
• Action: record exact before-stage and after-stage picks for every affected game.
• Guardrail: a later recovery does not make the unstable intermediate stage harmless.

High: weak run-model signal

• Severity: high.
• Evidence: away/home/total R² of 0.020313/0.010614/0.017602.
• Affected games: all 15.
• Action: investigate features and target through an isolated training backtest; do not attempt to solve this with nightly variance tuning alone.
• Guardrail: positive but tiny R² does not make exact scores reliable.

Medium: high raw probability cluster

• Severity: medium.
• Evidence: only 1/3 raw-probability-at-least-0.65 games won.
• Affected games: CHC@MIA, LAA@PIT, MIL@CIN.
• Action: monitor across slates and preserve raw-to-calibrated-to-published lineage.
• Guardrail: do not create a new per-bucket shrinkage rule from this slate alone.

Informational: complete slate audit coverage

• Severity: info.
• Evidence: 15/15 games structurally audited.
• Action: retain per-game audit receipts.
• Guardrail: complete coverage does not cancel findings within the audited games.

Informational: simulation inert on slate

• Severity: info for this individual slate, but escalated cross-slate.
• Evidence: zero flip candidates and zero published flips.
• Action: measure simulation contribution across rolling slates, separating score realization from winner flipping.
• Guardrail: zero winner flips does not prove zero score-reconciliation work.

3. Every game

The listed probability is the published probability of the selected side. The confidence edge is the absolute published-side edge from 50%; the raw signed home-minus-away edge is also shown where useful.

ARI@HOU — correct winner, wrong score

• Prediction: ARI 5, HOU 4; selected ARI.
• Actual: ARI 4, HOU 3; selected winner correct.
• Published probability / edge / bucket: ARI 0.517148; edge 0.03; toss_up.
• Brier: 0.233146.
• Errors: team-run absolute error 2; total bias +2; predicted margin -1, actual margin -1, margin error 0.
• Starters: Brandon Pfaadt projected 4 ER/19 outs, actual 3 ER/19 outs; Ethan Pecko projected 4 ER/12 outs, actual 1 ER/13 outs.
• Edges: season home-minus-away -0.033; recent +0.706; pitcher +0.014; composite +0.010.
• Controls: governor applied; raw probability 0.530; governed 0.5083; starter-experience shrinkage applied because one starter had only three career games; winner unchanged; score volatility no_change.
• Score path: historical 4.1–5.4, variance baseline 4–5, enhanced/final 5–4.
• Player adjustments: no player actual feed; no joined player-level result attribution.
• Audit: structural PASS, zero critical/error/warning; forecast artifact clean.
• Causes: no miss cause; this was a toss-up that went correctly. The score error was primarily run-model noise and Pecko outperforming his projected ER.

ATL@PHI — miss

• Prediction: ATL 5, PHI 4; selected ATL.
• Actual: ATL 2, PHI 4; winner was PHI.
• Published probability / edge / bucket: ATL probability 0.462021, edge 0.08; lean.
• Brier: 0.289421.
• Errors: team-run error 3; total bias +3; predicted margin ATL - PHI = +1, actual margin = -2, signed margin error -1 in the margin distribution representation.
• Starters: Martín Pérez projected 2 ER/17 outs, actual 4 ER/11 outs; Zack Wheeler projected 3 ER/16 outs, actual 1 ER/18 outs.
• Edges: season -0.545; recent -0.510; pitcher +0.002; composite -0.603, all favoring ATL except the negligible pitcher edge.
• Controls: raw 0.533406; governed 0.505; no recency guardrail; score volatility no_change; large Savant run-environment swing flag.
• Score path: historical 4.1–5.4, baseline 4–5, enhanced 5–4, final 5–4. The enhanced stage flipped the selection from PHI to ATL.
• Player adjustments: 24 material changes; notable adjustments included Drake Baldwin predicted-AB reduction, Realmuto predicted-hit increase, Edmundo Sosa predicted-AB reduction, Dominic Smith RBI reduction, and Schwarber strikeout reduction.
• Audit: structural PASS; probability publication is a registered critical lineage defect. No player actuals were available.
• Primary cause: mixed, with starter error on both sides and ATL offense overprojection.
• Secondary cause: bad winner weighting / confidence direction; the evidence edges favored ATL strongly enough in aggregate, but actual starter outcomes reversed the game.

BOS@BAL — correct winner, wrong score

• Prediction: BOS 5, BAL 4; selected BOS.
• Actual: BOS 5, BAL 0; selected winner correct.
• Published probability / edge / bucket: BOS 0.532306; edge 0.06; lean.
• Brier: 0.218738.
• Errors: team-run error 4; total bias +4; predicted margin -1 versus actual -5.
• Starters: Sonny Gray projected 3 ER/18 outs, actual 0 ER/21 outs; Chris Bassitt projected 3 ER/15 outs, actual 3 ER/17 outs.
• Edges: season -0.355; recent +0.254; pitcher +0.005; composite -0.300.
• Controls: raw 0.53732; governed 0.5051; recency guardrail applied; score volatility no_change.
• Score path: historical 5–4, baseline 5–4, enhanced/final 5–4.
• Player adjustments: no joined player actuals; player-level attribution unavailable.
• Audit: PASS; artifact clean.
• Causes: no winner miss. Score miss came mainly from underestimating Boston’s pitching and overestimating Baltimore’s offense.

CHC@MIA — correct winner and strong score

• Prediction: CHC 6, MIA 3; selected CHC.
• Actual: CHC 6, MIA 5; winner correct.
• Published probability / edge / bucket: CHC 0.73953; edge 0.48; strong.
• Brier: 0.067845.
• Errors: team-run error 2; total bias -2; predicted margin -3 versus actual -1.
• Starters: Javier Assad projected 2 ER/15 outs, actual 3 ER/14 outs; Ryan Gusto projected 6 ER/13 outs, actual 5 ER/14 outs.
• Edges: season -0.799; recent -0.770; pitcher +0.086; composite -0.984.
• Controls: raw 0.7723; calibrated/governed 0.7404; no recency guardrail; high-total-environment lift and supported blowout-margin expansion.
• Score path: historical 4–5, baseline 4–5, enhanced/final 6–3.
• Player adjustments: no player actuals; no actual player attribution.
• Audit: PASS; artifact clean.
• Causes: none for winner. This was the strongest call. It is one of the games that would be softened by indiscriminate variance widening or confidence shrinkage.

DET@CLE — miss

• Prediction: DET 3, CLE 4; selected CLE.
• Actual: DET 6, CLE 0; winner was DET.
• Published probability / edge / bucket: CLE 0.5544; edge 0.11; moderate.
• Brier: 0.307378.
• Errors: team-run error 7; total bias +1; predicted margin CLE +1 versus actual DET +6.
• Starters: Framber Valdez projected 2 ER/16 outs, actual 0 ER/21 outs; Parker Messick projected 3 ER/18 outs, actual 3 ER/18 outs.
• Edges: season +0.026; recent +0.706; pitcher -0.050; composite +0.262, favoring CLE overall.
• Controls: raw 0.58; governed 0.5663; recency guardrail applied; score volatility no_change.
• Score path: historical 4–5, baseline 4–5, enhanced/final 3–4.
• Player adjustments: 22 material changes; notable changes included Petey Halpin strikeout increase, Angel Martínez hit increase, Jo Adell hit increase, Parker Messick strikeout increase, Cade Smith strikeout reduction, and Javier Báez walk increase.
• Audit: PASS; artifact clean.
• Primary cause: wrong conviction side; the model backed CLE with a real but only moderate 0.11 edge.
• Secondary cause: small-edge miss and severe run realization error. This is a key lever-tension game for wider tails.

LAA@PIT — miss

• Prediction: LAA 4, PIT 6; selected PIT.
• Actual: LAA 6, PIT 1; winner was LAA.
• Published probability / edge / bucket: PIT 0.635729; edge 0.27; strong.
• Brier: 0.404151.
• Errors: team-run error 7; total bias +3; predicted margin PIT +2 versus actual LAA +5.
• Starters: Yusei Kikuchi projected 6 ER/14 outs, actual 0 ER/18 outs; Braxton Ashcraft projected 3 ER/19 outs, actual 3 ER/15 outs.
• Edges: season +0.909; recent +1.162; pitcher +0.153; composite +1.488, all favoring PIT.
• Controls: raw 0.67; governed/calibrated 0.6591; starter-experience shrinkage reduced the probability to published 0.635729; no recency guardrail; volatility no_change; large Savant swing flag.
• Score path: historical 4.5–6.1, baseline 4–6, enhanced/final 4–6.
• Player adjustments: 22 material changes; Kikuchi’s predicted strikeouts were materially adjusted, with notable changes to Oneil Cruz, Bryan Reynolds, Rafael Flores Jr., Zach Neto, Wade Meckler, and Adam Frazier.
• Audit: PASS structurally, but critical probability-lineage defect: calibrated 0.6591 to published 0.635729 has no recorded material transformation.
• Primary cause: starter projection error, especially Kikuchi’s six-ER projection versus zero actual ER.
• Secondary cause: wrong high-conviction side. This is a high-raw-cluster loss.

MIL@CIN — miss

• Prediction: MIL 8, CIN 5; selected MIL.
• Actual: MIL 3, CIN 5; winner was CIN.
• Published probability / edge / bucket: MIL 0.540; edge 0.08; lean.
• Brier: 0.291600.
• Errors: team-run error 5; total bias +5; predicted margin MIL +3 versus actual CIN +2.
• Starters: Dustin May projected 3 ER/15 outs, actual 2 ER/11 outs; Andrew Abbott projected 2 ER/15 outs, actual 1 ER/12 outs.
• Edges: season -1.666; recent -0.830; pitcher -0.062; composite -1.801, favoring MIL.
• Controls: raw 0.7085; calibrated 0.6861; governor input 0.54; recency guardrail applied; it reduced the published pick-side probability to 0.54; supported blowout-margin expansion.
• Score path: historical 4.4–3.8, baseline 4–4, enhanced/final 8–5. The variance/enhanced path strongly expanded Milwaukee’s score.
• Player adjustments: 27 material changes; notable effects included Andrew Vaughn AB reduction, DL Hall projected-IP change, Luis Lara/Carlos Jorge AB increases, and Jackson Chourio RBI increase.
• Audit: PASS; the large probability transformation is recorded as the matchup-recency guardrail, so it is not a lineage bypass.
• Primary cause: run-environment overprojection.
• Secondary cause: excessive score expansion converted a nominally favorable MIL edge into an overconfident published direction. This is the second high-raw-cluster loss.

MIN@CHW — correct winner, wrong score

• Prediction: MIN 6, CHW 5; selected MIN.
• Actual: MIN 6, CHW 4; winner correct.
• Published probability / edge / bucket: MIN 0.530; edge 0.06; lean.
• Brier: 0.220900.
• Errors: team-run error 1; total bias +1; predicted margin -1 versus actual -2.
• Starters: Taj Bradley projected 5 ER/17 outs, actual 4 ER/15 outs; Anthony Kay projected 4 ER/16 outs, actual 5 ER/13 outs.
• Edges: season +0.462; recent +0.298; pitcher -0.040; composite +0.061.
• Controls: raw 0.52225; governed 0.5083; recency guardrail applied; volatility no_change.
• Score path: historical 4–5, baseline 4–5, enhanced/final 6–5.
• Player adjustments: no actual player feed; no player-result attribution.
• Audit: PASS.
• Causes: no winner miss. Score quality was comparatively good. Starter errors were modest and partly offsetting.

NYY@SD — correct winner, wrong score

• Prediction: NYY 4, SD 3; selected NYY.
• Actual: NYY 5, SD 1; winner correct.
• Published probability / edge / bucket: NYY 0.548506; edge 0.10; lean.
• Brier: 0.203847.
• Errors: team-run error 3; total bias +1; predicted margin -1 versus actual -4.
• Starters: Carlos Rodón projected 0 ER/17 outs, actual 1 ER/15 outs; Robbie Ray projected 3 ER/17 outs, actual 5 ER/7 outs.
• Edges: season -0.736; recent -0.278; pitcher -0.030; composite -0.720.
• Controls: raw 0.5509; governed 0.5499; recency guardrail applied; published probability matched calibrated/governed value; volatility no_change.
• Score path: historical approximately 4–4, baseline 4–4, variance stage changed NYY to a tie, enhanced restored NYY, final 4–3.
• Player adjustments: no player actual feed.
• Audit: PASS.
• Cause relevance: this game was one of the registered winner-direction regressions at historical_ml → variance_applied, although the later enhanced stage recovered the correct winner. This illustrates why later recovery does not erase the unstable transition.

OAK@SEA — miss

• Prediction: OAK 6, SEA 8; selected SEA.
• Actual: OAK 6, SEA 2; winner was OAK.
• Published probability / edge / bucket: SEA 0.530; edge 0.06; lean.
• Brier: 0.280900.
• Errors: team-run error 6; total bias +6; predicted margin SEA +2 versus actual OAK +4.
• Starters: Jeffrey Springs projected 5 ER/12 outs, actual 0 ER/21 outs; George Kirby projected 6 ER/16 outs, actual 3 ER/20 outs.
• Edges: season +0.478; recent -0.342; pitcher -0.052; composite +0.424.
• Controls: raw 0.53; governed 0.5083; recency guardrail applied; volatility no_change; large Savant swing flag.
• Score path: historical 4–5, baseline 4–5, enhanced 6–7, final 6–8.
• Player adjustments: 23 material changes; notable changes involved Cal Raleigh strikeouts, Henry Bolte RBI, Randy Arozarena RBI, Jonah Heim hits, Taylor Ward strikeouts, Jeff McNeil doubles, and Dominic Canzone doubles.
• Audit: PASS.
• Primary cause: starter projection; Springs allowed zero rather than the projected five ER.
• Secondary cause: run environment overprojection, especially Seattle’s offense. This was a substantial score miss despite only a lean winner edge.

SF@NYM — miss

• Prediction: SF 4, NYM 5; selected NYM.
• Actual: SF 9, NYM 5; winner was SF.
• Published probability / edge / bucket: NYM 0.523578; edge 0.05; lean.
• Brier: 0.274134.
• Errors: team-run error 5; total bias -5; predicted margin NYM +1 versus actual SF +4.
• Starters: Anthony Molina projected 2 ER/9 outs, actual 0 ER/15 outs; Zac Thornton projected 1 ER/17 outs, actual 6 ER/15 outs.
• Edges: season +0.183; recent +0.206; pitcher +0.016; composite +0.368, favoring NYM.
• Controls: raw 0.531642; governed 0.505; starter-experience shrinkage reduced the intermediate probability; no recency guardrail; high-total-environment lift.
• Score path: historical 4.4–4.3, baseline 4–4, enhanced/final 4–5.
• Player adjustments: 26 material changes; notable adjustments involved Andrew Knizner, Jonah Cox, Anthony Molina’s projected innings, Zac Thornton strikeouts, Daniel Duarte, Juan Soto, Jared Young, and Bryce Eldridge.
• Audit: structural PASS. The lineage text says no recorded transformation, but the 1.8578-point delta is below the registered material threshold, so it is a documentation issue rather than a critical registered breach.
• Primary cause: starter projection, particularly Thornton’s 6 actual ER versus 1 projected.
• Secondary cause: run environment underprojection; SF scored five more runs than forecast. This game was also an intermediate winner-direction regression at the variance stage before later recovery to NYM, which was still wrong relative to the final outcome.

STL@COL — miss

• Prediction: STL 7, COL 6; selected STL.
• Actual: STL 7, COL 10; winner was COL.
• Published probability / edge / bucket: STL 0.509201; edge 0.02; toss_up.
• Brier: 0.259286.
• Errors: team-run error 4; total bias -4; predicted margin STL +1 versus actual COL +3.
• Starters: Matthew Liberatore projected 4 ER/16 outs, actual 5 ER/13 outs; Mason Adams projected 4 ER/13 outs, actual 3 ER/14 outs.
• Edges: season -0.736; recent -0.199; pitcher +0.057; composite -0.509, favoring STL.
• Controls: raw 0.53551; governed 0.5083; starter-experience guardrail applied because one starter had only two career games; volatility no_change.
• Score path: historical 6–5, baseline 6–5, enhanced/final 7–6.
• Player adjustments: 22 material changes; notable effects included Bryan Torres and Leonardo Bernal reductions, Joshua Baez RBI increase, Troy Johnston and Connor Norby AB increases, and Jordan Beck adjustment.
• Audit: PASS.
• Primary cause: toss-up variance.
• Secondary cause: small-edge miss and COL offense underprojection. This was not a high-conviction failure.

TB@TEX — correct winner, wrong score

• Prediction: TB 6, TEX 5; selected TB.
• Actual: TB 6, TEX 3; winner correct.
• Published probability / edge / bucket: TB 0.533459; edge 0.07; lean.
• Brier: 0.217661.
• Errors: team-run error 2; total bias +2; predicted margin -1 versus actual -3.
• Starters: Drew Rasmussen projected 3 ER/18 outs, actual 1 ER/21 outs; Jacob deGrom projected 4 ER/15 outs, actual 2 ER/18 outs.
• Edges: season -0.589; recent -0.298; pitcher -0.049; composite -0.685.
• Controls: raw 0.49725; governed 0.5083; recency guardrail applied; volatility no_change.
• Score path: historical approximately 4–4, baseline 4–4, enhanced/final 6–5.
• Player adjustments: no player actual feed.
• Audit: PASS.
• Cause relevance: registered as an intermediate winner-direction regression at historical_ml → variance_applied, where TB moved to a tie, then recovered at enhanced prediction. The final winner was correct.

TOR@KC — correct winner, wrong score

• Prediction: TOR 7, KC 5; selected TOR.
• Actual: TOR 4, KC 3; winner correct.
• Published probability / edge / bucket: TOR 0.570; edge 0.14; moderate.
• Brier: 0.184900.
• Errors: team-run error 5; total bias +5; predicted margin TOR +2 versus actual +1.
• Starters: Max Scherzer projected 4 ER/15 outs, actual 3 ER/17 outs; Seth Lugo projected 2 ER/15 outs, actual 3 ER/15 outs.
• Edges: season -0.491; recent -0.488; pitcher -0.090; composite -0.951.
• Controls: raw 0.57; governed/calibrated 0.6202; recency guardrail applied; supported blowout-margin expansion.
• Score path: historical 5–4, baseline 5–4, enhanced/final 7–5.
• Player adjustments: no player actual feed.
• Audit: PASS.
• Cause: correct winner but exact-score failure. This is explicitly a right_winner_wrong_score game. The score expansion was not supported by the final run total.

WAS@LAD — correct winner, wrong score

• Prediction: WAS 3, LAD 4; selected LAD.
• Actual: WAS 5, LAD 6; winner correct.
• Published probability / edge / bucket: LAD 0.573431; edge 0.15; moderate.
• Brier: 0.181961.
• Errors: team-run error 4; total bias -4; predicted margin LAD +1 versus actual +1.
• Starters: Cade Cavalli projected 4 ER/17 outs, actual 2 ER/18 outs; Tyler Glasnow projected 3 ER/17 outs, actual 4 ER/18 outs.
• Edges: season +0.741; recent +0.306; pitcher +0.041; composite +0.753.
• Controls: raw 0.61585; governed/calibrated 0.6219; starter-experience shrinkage applied; final published probability 0.573431; volatility no_change.
• Score path: historical approximately 3–4, baseline 3–4, enhanced/final 3–4.
• Player adjustments: no actual player feed.
• Audit: structural PASS, but critical probability-lineage defect: calibrated 0.6219 to published 0.573431 has no recorded material transformation.
• Causes: no winner miss. Exact-score miss was run-environment underprojection. This is another correct-winner/wrong-score game.

4. Miss reconstruction

Earliest wrong-direction stage for each loss

ATL@PHI

• Earliest wrong-direction movement: enhanced_prediction, which changed the score from PHI 5–ATL 4 at the variance stage to ATL 5–PHI 4.
• Classification: bad weighting / winner-direction regression, compounded by starter error.
• Evidence: composite edge was -0.603 toward ATL, but actual Pérez allowed 4 ER versus 2 projected and Wheeler allowed 1 versus 3 projected.
• Not primarily: artifact failure; structural audit passed.
• Additional issue: publication delta was unexplained and critical.

DET@CLE

• Earliest wrong-direction stage: the historical/variance/enhanced path already favored CLE; no later stage corrected it.
• Classification: bad weighting plus irreducible/realized offense variance.
• Evidence: composite edge +0.262, recent edge +0.706, but DET scored 6 and CLE 0. Valdez allowed zero ER rather than two; Messick matched his ER projection.
• Run-model failure: seven team runs of error.

LAA@PIT

• Earliest wrong-direction stage: historical model and all subsequent stages favored PIT.
• Classification: starter error, then overconfidence.
• Evidence: Kikuchi projected 6 ER and allowed 0. The model’s composite home edge was +1.488, so weighting consistently reinforced PIT.
• Contract issue: calibrated-to-published probability change unexplained.
• Important: the starter-experience guardrail reduced confidence but did not change the winner.

MIL@CIN

• Earliest wrong-direction stage: enhanced_prediction, where Milwaukee was expanded to 8 runs after a 4–4 baseline.
• Classification: run-environment miss and bad score weighting; not a major starter miss.
• Evidence: actual starters were each only one ER from projection, but Milwaukee scored 3 rather than 8.
• Important: the recency guardrail materially reduced the published probability from calibrated 0.6861 to 0.54, and this transformation is recorded.

OAK@SEA

• Earliest wrong-direction stage: the historical and enhanced stages already favored SEA; the Savant environment increased SEA to 8 runs.
• Classification: starter projection plus run-environment miss.
• Evidence: Springs projected 5 ER and allowed 0; Kirby projected 6 and allowed 3. Actual OAK won 6–2.
• Model issue: the run environment did not merely miss the winner; it materially overstated total scoring by six runs.

SF@NYM

• Earliest wrong-direction stage: enhanced_prediction moved from a 4–4 baseline to NYM 5–4, and later stages preserved it.
• Classification: starter projection plus run-environment underprojection.
• Evidence: Thornton projected 1 ER and allowed 6; Molina projected 2 and allowed 0. SF scored 9 rather than 4.
• Pipeline issue: variance stage had an intermediate winner-direction regression, later recovered to the wrong final side.

STL@COL

• Earliest wrong-direction stage: historical/variance/enhanced stages favored STL.
• Classification: toss-up variance, small-edge miss, and COL offense underprojection.
• Evidence: published STL probability was only 0.509201; actual COL won 10–7.
• Not primarily: starter error; both starters were within one ER of projection.

Loss taxonomy

Across the seven losses:

• Starter projection: primary in LAA@PIT, OAK@SEA, SF@NYM; materially contributory in ATL@PHI.
• Run-environment overprojection: primary in MIL@CIN; secondary in OAK@SEA.
• Run-environment underprojection: secondary in SF@NYM; also affected STL@COL and correct-winner score misses.
• Wrong conviction side: explicit in DET@CLE and LAA@PIT.
• Toss-up variance: explicit in STL@COL.
• Bad input: no direct evidence of malformed team/starter identities; all starter identity matches were true and all structural audits passed.
• Bad weighting: evident in the variance/enhanced transitions, especially the 3-hit degradation at historical_ml → variance_applied.
• Irreducible variance: most defensible for STL@COL, and partially for extreme starter outcomes, but not a sufficient explanation for the persistent run-model weakness.

A published home-away confidence edge of 0.12 is moderate, not a near coin flip. The system’s confidence labels correctly distinguish toss-up, lean, moderate, and strong; a 0.11–0.12 edge should not be treated as negligible.

Pipeline-stage audit across all stages

Transitions

• historical_ml → variance_applied
• Winner accuracy: -0.20
• Changed direction: MIL@CIN, NYY@SD, SF@NYM, TB@TEX to ties.
• Regressed games recorded: NYY@SD, SF@NYM, TB@TEX.
• This is the most important intermediate winner-direction defect.

• variance_applied → enhanced_prediction
• Winner accuracy: +0.20
• Recovered: CHC@MIA, NYY@SD, TB@TEX, TOR@KC.
• Regressed: ATL@PHI.
• The recovery improves the final count but does not erase the earlier regression.

• enhanced_prediction → embedded_savant_run_environment
• No winner changes.
• Margin MAE worsened by 0.0666; team-run MAE worsened by 0.0333.

• All later transitions:
• No winner changes.
• No score improvement.
• simulation_result → published_final was inert.

5. Cross-game counterfactuals

Lever tension: widen run distributions

Lever: score.variance_multiplier / score.variance_floor.

Helps

• DET@CLE: lost a 0.11-confidence pick; wider tails could reduce overconfidence in a narrow CLE projection.
• LAA@PIT: lost a 0.27-confidence pick; wider tails could better represent starter uncertainty.

Hurts

• CHC@MIA: strong 0.48 edge and correct winner; widening could soften a good call.

Net evidence

• 2 helps versus 1 hurts.
• The net supports a small bounded shadow increase, not an aggressive global change.
• However, persistent weak run-model signal means variance tuning cannot substitute for better predictive features.

Lever tension: pitcher form weights / prior games

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

Helps

• ARI@HOU: starter ER off by 3.
• BOS@BAL: starter ER off by 3.
• LAA@PIT: starter ER off by 6.
• OAK@SEA: starter ER off by 5.
• SF@NYM: starter ER off by 5.

Hurts

• CHC@MIA, MIL@CIN, MIN@CHW, STL@COL, TOR@KC: starter projections were each only one ER off, so broad reweighting could degrade already adequate cases.

Net evidence

• Exactly 5 helps versus 5 hurts.
• No broad parameter move is supported. A targeted starter-form diagnostic or shadow split is preferable.

Lever tension: recent/season run-mean weighting

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

Helps

• BOS@BAL: team-run error 4.
• DET@CLE: team-run error 7.
• LAA@PIT: team-run error 7.
• MIL@CIN: team-run error 5.
• OAK@SEA: team-run error 6.
• SF@NYM: team-run error 5.
• STL@COL: team-run error 4.
• TOR@KC: team-run error 5.
• WAS@LAD: team-run error 4.

Hurts

• ARI@HOU, CHC@MIA, MIN@CHW, TB@TEX: team-run errors were only 1–2, so changing means risks disturbing good score estimates.

Net evidence

• 9 helps versus 4 hurts.
• The net favors investigation and a bounded shadow experiment, but the near-zero holdout R² means the correct response is feature/target work, not simply increasing recency or season weights.

Tempting one-game fixes that should not be made

• Do not reverse all moderate picks because DET@CLE lost. The 0.11 edge is moderate, not a near coin flip.
• Do not globally suppress strong probabilities because LAA@PIT lost; CHC@MIA was a strong, accurate call.
• Do not broadly reweight starters from LAA@PIT, OAK@SEA, or SF@NYM; five games had good starter projections.
• Do not remove score expansion entirely because MIL@CIN and OAK@SEA overprojected runs; CHC@MIA benefited from a strong score-tail adjustment.
• Do not tune the simulation to force winner flips. Its nine-slate persistence requires a code/diagnostic fix, not outcome-specific intervention.
• Do not treat a later winner recovery as evidence that the variance stage is safe.

6. Data and contract defects

Critical findings

1. Unexplained probability publication delta
2. ATL@PHI: +3.2979 pp.
3. LAA@PIT: -2.3371 pp.
4. WAS@LAD: -4.8469 pp.
5. Required action: block promotion until each transformation is recorded and reconciled.

1. Missing integrity revision audit
2. The manifest explicitly marks the integrity revision audit as missing.
3. This is a source-receipt gap, even though the loaded structural audits passed.

Error findings

• No structural audit errors were reported.
• Structural audit summary: 0 errors.

Warning findings

• No structural audit warnings were reported.
• Structural audit summary: 0 warnings.

Informational findings

• Structural audit summary contained 111 informational findings.
• Every game’s audit reported zero critical/error/warning findings, with either 100 or 101 “good” checks and 7 or 8 informational notes.
• The player feed had no rows for the slate, preventing actual player-level attribution.
• No source-labelled market data was available.
• Simulation flip candidates and published flips were both zero.

Prediction-model defects versus artifact bugs

Prediction-model/process defects

• Near-zero run-model holdout signal.
• Winner-direction regression at the variance stage.
• Persistent simulation inertness.
• Starter projection instability in several specific games.
• Run-environment over- and under-dispersion.
• High raw-probability cluster underperformance, currently only a medium one-slate finding.

Presentation/artifact/contract defects

• Three material calibrated-to-published probability deltas have no recorded transformation.
• The missing integrity revision audit is a receipt/verification defect.
• SF@NYM has a smaller undocumented delta: its transformation text says no recorded transformation, but the delta is below the material threshold.
• These contract defects are distinct from the model’s poor run accuracy.

Coverage reconciliation

The coverage matrix says the slate is fully covered:

• 15 expected and 15 audited.
• 15 canonical artifacts loaded.
• 15 starter comparisons.
• 15 continuous-learning records.

Therefore, the slate is coverage-complete but not contract-clean. Full coverage does not permit calling the slate clean because three probability-lineage defects remain.

Complete pick-side probability lineage

The expected path is:

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

Material examples:

• CHC@MIA: raw 0.7723 → calibrated/governed 0.7404 → published 0.73953. Transformation recorded; no bypass.
• MIL@CIN: raw 0.7085 → calibrated 0.6861 → governor input 0.54 → governed/published 0.54 after matchup-recency guardrail. Large delta is recorded; not a bypass.
• ARI@HOU: raw 0.5085 → governor input 0.53 → calibrated/governed 0.5083 → published 0.517148 after score coherence/recency. Recorded.
• ATL@PHI: calibrated 0.5050 → published 0.537979 with no recorded transformation. Bypass/contract failure cannot be conclusively called a governor bypass, but the publication transformation is missing.
• LAA@PIT: calibrated 0.6591 → published 0.635729 with no recorded material transformation. Same conclusion.
• WAS@LAD: calibrated 0.6219 → published 0.573431 with no recorded material transformation. Same conclusion.
• SF@NYM: calibrated 0.5050 → published 0.523578; text says no recorded transformation, but delta is 1.8578 pp and therefore below the registered material threshold.

The evidence supports calling a material undocumented transformation, not asserting a governor bypass without further math tracing.

7. Cross-slate persistence

The trend window contains nine slates: 2026-08-28 through 2026-09-05, with eight prior slates loaded.

Finding streaks

• high_raw_probability_cluster: consecutive count 1; present now. Not established across slates.
• pipeline_winner_instability: consecutive count 1; present now. It is current but not yet persistent by the registered streak count.
• score_margin_compression: consecutive count 0; not present now, despite three incidences in the nine-slate window.
• simulation_inert_on_slate: consecutive count 9; present now and escalated. Established failure, not single-slate noise.
• single_slate_calibration_instability: consecutive count 0; not present now.
• weak_run_model_signal: consecutive count 9; present now and escalated. Established failure, not single-slate noise.

Escalated findings

Simulation inertness

• Streak: 9 consecutive slates.
• Target lever: simulation.allow_score.
• Metric harmed: team-run MAE and total-run MAE; also prevents meaningful simulation-based winner intervention.
• Direction: confirm the simulation module is actually running and not silently no-oping.
• Interpretation: this is a code/diagnostic failure, not evidence for a parameter increase.

The self-assessment identifies a likely root cause: the simulation uses draws selected while searching for a target-realizing play-by-play, exits early with roughly two draws, and then treats those selected draws as a probability distribution. That makes the win-rate estimate vacuous.

Weak run-model signal

• Streak: 9 consecutive slates.
• Target levers listed: training.xgb.n_estimators, training.xgb.max_depth, training.xgb.learning_rate.
• Metric harmed: team-run MAE and total-run MAE.
• Direction: investigate features and target using an isolated training backtest.
• Guardrail: the finding explicitly says not to force a nightly parameter move; the issue is feature/target quality, not a simple hyperparameter shortage.

The self-assessment recommends adding, where legitimately available:

• source-labelled no-vig market probability,
• park/weather run environment,
• bullpen rest/fatigue,
• platoon-adjusted lineup quality,

then retraining under a new model schema version and comparing same-fingerprint holdout R².

Metric trends

• margin_bias: improving. Series ended at -1.467 after a worse earlier period.
• mean_team_run_error_on_losses: improving. Latest 5.286, though still materially high.
• weighted_calibration_error: improving. Latest 0.100, down from several earlier values above 0.20.
• winner_accuracy: regressing. Latest 0.5333; nine-slate mean 0.5654.
• top_confidence_pick loss streak: 0.
• 2026-08-28 hit
• 2026-08-29 hit
• 2026-08-30 miss
• 2026-08-31 hit
• 2026-09-01 miss
• 2026-09-02 miss
• 2026-09-03 miss
• 2026-09-04 hit
• 2026-09-05 hit

The top-confidence pick has had intermittent misses but no current consecutive-loss streak.

8. Adjustments for today

Recommendations are ranked by persistence, scope, expected metric effect, and regression risk.

1. Fix and independently validate the simulation path — highest evidence strength

• Pattern: simulation inertness, nine consecutive slates.
• Registered lever: simulation.allow_score, but this should not be used as a blind parameter move.
• Bounded change: code/diagnostic action rather than parameter tuning:
• verify the simulation module executes;
• separate target-realization search from independent winner-probability simulation;
• use an independent, unsteered batch of approximately 200–500 draws or a closed-form run-difference calculation;
• remove the early exit from the probability batch;
• log before-stage and after-stage winner probabilities.
• Expected metric: improve the diagnostic value of simulation; reduce winner-direction opacity and potentially improve team-run/total-run MAE.
• Helps: games with unstable intermediate winner direction, especially NYY@SD, SF@NYM, and TB@TEX.
• Hurts/risk: an independent simulation may soften a good strong call such as CHC@MIA; therefore use a shadow result and rollback monitor.
• Rollback monitor: no promotion unless independent-simulation contribution improves rolling winner Brier or score reconciliation without increasing team-run MAE.

2. Run the persistent run-model feature/target backtest — high evidence strength

• Pattern: weak run-model signal, nine consecutive slates.
• Registered parameters: training.xgb.n_estimators, training.xgb.max_depth, training.xgb.learning_rate are listed, but the recommended action is not a blind nightly move.
• Bounded change: perform an isolated same-fingerprint backtest after adding or validating park/weather, bullpen fatigue, platoon-adjusted lineup quality, and legitimately source-labelled market features.
• Expected metrics: increase away/home/total holdout R² and reduce team-run and total-run MAE.
• Helps: the broad score-error group, especially DET@CLE, LAA@PIT, MIL@CIN, OAK@SEA, SF@NYM, TOR@KC, and WAS@LAD.
• Hurts/risk: could disturb already-close score forecasts in MIN@CHW, ARI@HOU, CHC@MIA, and TB@TEX.
• Rollback monitor: require out-of-sample improvement in all three run targets or at least no degradation in winner Brier before adoption. Bump the model schema version when feature semantics change.

3. Shadow a small confidence-cap reduction for the current high-raw cluster — lower evidence strength

• Pattern: high raw probability cluster, one consecutive slate only; 1/3 hit.
• Exact registered parameter: confidence.base_cap.
• Bounded change: reduce confidence.base_cap by 0.02 in shadow mode only.
• Expected metric: lower calibration error and possibly improve Brier score by reducing overstatement in LAA@PIT and MIL@CIN.
• Helps: LAA@PIT, MIL@CIN, and potentially moderate/high conviction misses such as DET@CLE.
• Hurts: CHC@MIA, where the 0.48 edge was correct and could be unnecessarily softened.
• Rollback monitor: because this finding has only a one-slate streak, revert if rolling Brier or weighted calibration error worsens over the next registered evaluation window.
• Do not simultaneously alter multiple confidence caps, because that would make attribution impossible.

No direct starter-weight change is recommended today: the pitcher lever is 5–5 in cross-game tension, and the evidence does not support a net direction.

9. Bottom line

Known

• The slate was fully covered structurally: 15/15 games, 8 winners correct.
• Winner accuracy was only 53.33%, below the 55.85% walk-forward baseline.
• Exact-score performance was weak, with near-zero validated run-model R² and a 3.333-run margin MAE.
• The variance stage caused a measurable three-hit winner-direction regression before later recovery.
• Seven games were missed; the most consequential concrete errors were severe starter projection failures in LAA@PIT, OAK@SEA, and SF@NYM, plus run-environment overprojection in MIL@CIN.
• Two critical probability-lineage defects are not merely theoretical: ATL@PHI, LAA@PIT, and WAS@LAD each have material unexplained calibrated-to-published deltas.
• Simulation inertness and weak run-model signal are established nine-slate patterns.
• Calibration and margin bias are improving, but winner accuracy is regressing.

Uncertain

• The exact code path producing the three material probability publication deltas is not established; the evidence shows missing transformations, not conclusively a governor bypass.
• Player-level actual attribution is unavailable because the slate-wide player feed contained no rows.
• The individual starter misses may include irreducible variance, but their recurrence and magnitude justify targeted diagnostic work.
• The one-slate high-raw cluster does not establish a permanent confidence-shrinkage rule.

Single safest next action

Run the operator-level simulation diagnostic and contract reconciliation before promoting another slate: independently verify that the simulation executes with an unsteered probability batch, and trace every calibrated-to-published probability transformation for ATL@PHI, LAA@PIT, and WAS@LAD. This is safer than tuning scores or confidence because it addresses an established code/process failure and the registered critical publication defects without weakening the proven strong call in CHC@MIA.