Luna deep analysis — slate 20260902

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

1. Evidence receipt and gaps

Source manifest

All seven manifest entries were reviewed as evidence records:

SourceStatusSizePurpose
prediction review — /app/output/reports/prediction_review_20260902.jsonLoaded372,146 bytesOfficial outcomes and scoring errors
forensic deep audit — /app/output/reports/PREDICTION_SYSTEM_DEEP_AUDIT_20260902.jsonLoaded240,269 bytesProbability, provenance, starter, and decision contracts
continuous learning — /app/output/learning/continuous_learning_daily_20260902.jsonLoaded1,925,532 bytesPipeline-stage quality, calibration, probability lineage, score distribution, starter comparisons
detailed report — /app/output/reports/detailed_20260902.htmlLoaded352,626 bytesFull game and simulation rendering
ultra detailed player audit — /app/output/reports/20260902_ultra_detailed_player_audit.htmlLoaded30,625,934 bytesPlayer-level prediction trace
structural audit — /app/output/reports/audit_report_20260902.htmlLoaded1,878,790 bytesInvariant findings
integrity revision audit — /app/output/reports/audit_report_20260902_integrity_revision.htmlMissingNot suppliedPost-Luna remediation proof for immutable revision artifacts

The missing integrity-revision audit is a genuine evidence gap. It prevents independent confirmation that any post-generation remediation or immutable revision process was correctly recorded.

Game coverage

All 15 expected games were reviewed:

  1. ATL@WAS
  2. BAL@COL
  3. CHW@HOU
  4. DET@MIN
  5. MIA@KC
  6. MIL@CHC
  7. NYM@TB
  8. NYY@LAA
  9. OAK@TEX
  10. PHI@ARI
  11. SD@CIN
  12. SEA@BOS
  13. SF@PIT
  14. STL@LAD
  15. TOR@CLE

Coverage is complete according to the matrix:

The slate was therefore fully covered, but it was not defect-free.

Internal gaps and contradictions

  1. Missing integrity revision audit. The manifest marks the post-Luna immutable-revision proof as missing.
  2. SF@PIT probability lineage defect. The causal register identifies one unexplained calibrated-to-published delta of at least two percentage points, affecting SF@PIT. Its published probability was 0.54559 versus calibrated probability 0.5223, a +2.329 percentage-point delta, with unexplained_material_delta: true.
  3. MIA@KC minor lineage inconsistency. The record says “no recorded transformation explains the publication delta,” but marks unexplained_material_delta: false. The delta is only +0.9808 percentage points, below the material threshold, so this is a traceability warning rather than the critical SF-type defect.
  4. Pipeline-stage accounting needs careful interpretation. variance_applied → enhanced_prediction reduced winner accuracy from 9/15 to 5/15, a four-hit regression. The transition lists nine direction changes, of which one improved ATL@WAS and five are explicitly classified as regressions; three additional changes were recoveries or non-loss changes. The causal register correctly focuses on the net four-hit degradation, but the full flip list must remain in the trace.
  5. Player-level actuals are unavailable for all games where player_level_forensics.available is false. Player adjustments are recorded, but post-game player-level attribution cannot be independently verified for those games.
  6. Structural-audit totals differ in composition by game. The slate total is 0 critical, 0 error, 2 warnings, and 110 informational findings. Two games carry warnings; the remainder pass structurally. This does not invalidate the model findings.

2. Executive diagnosis

Overall result

This was not merely a bad-score slate. The published winner layer also failed materially, and an intermediate model stage degraded winner direction.

Winner accuracy

The published system correctly selected five winners:

It missed ten.

The historical ML stage would have selected 11/15 correctly. After variance_applied, accuracy fell to 9/15. After enhanced_prediction, it fell further to 5/15. All later stages remained at 5/15.

That makes the most important winner-direction finding the high-severity pipeline instability at variance_applied → enhanced_prediction. This is a material model/process risk, not an artifact issue.

Calibration quality

Published confidence was materially overconfident:

The two moderate-confidence buckets performed especially poorly. The evidence supports bounded empirical shrinkage experiments, but the causal register correctly warns against treating one slate as a stable calibration curve. Cross-slate evidence later establishes this as a persistent pattern.

No high-raw-probability cluster was present: threshold 0.65, games 0, hits 0.

Score accuracy

Score accuracy was the weakest broad model dimension.

The score-margin compression finding is high severity because it is broad, repeated, and persistent across six prior slates in the trend record. It is not a justification for indiscriminately widening every forecast; it supports a bounded shadow experiment on registered variance parameters.

The low run-model R² confirms that exact scores are weakly supported by the current features. Winner and score systems should not be treated as equally reliable outputs.

Starter accuracy

Starter projections were mixed rather than uniformly defective.

Several games had meaningful starter ER errors:

Other games had relatively accurate starter ER projections, with worst error only one run:

The lever tension was nearly balanced: pitcher-form reweighting helped seven games and risked hurting six. The evidence does not justify a broad pitcher-weight change today.

Artifact integrity

Artifact integrity was mostly sound but not clean:

The causal register must therefore be reconciled as follows:

Critical: unexplained probability publication delta

High: artifact-integrity warning

High: pipeline winner instability

High: score-margin compression

High: single-slate calibration instability

High: weak run-model signal

Informational: complete audit coverage

Informational: simulation inertness

3. Every game

Notation: edge is the absolute home-minus-away confidence distance; signed edge is home probability direction. “Path” lists the meaningful published score stages.

ATL@WAS — hit, correct winner/wrong score

BAL@COL — miss

CHW@HOU — hit, correct winner/wrong score

DET@MIN — miss

MIA@KC — hit, correct winner/wrong score

MIL@CHC — hit, solid correct call but wrong score

NYM@TB — miss

NYY@LAA — hit, solid correct call but wrong score

OAK@TEX — miss

PHI@ARI — miss

SD@CIN — miss

SEA@BOS — miss

SF@PIT — miss

STL@LAD — miss

TOR@CLE — miss

4. Miss reconstruction

Earliest wrong-direction movement for every loss

GameEarliest stage moving toward wrong resultDiagnosis
BAL@COLHistorical ML already favored BALBad weighting/input balance or irreducible close-game variance; starter projection also overstated Rogers’ ER
DET@MINenhanced_prediction: DET → MINBad weighting/process regression; recent/H2H/low-score logic overrode the prior direction. The 0.12 edge was moderate, not near coin-flip
NYM@TBHistorical ML already favored TBBad winner weighting or irreducible variance; score model severely underpredicted NYM offense
OAK@TEXHistorical ML slightly favored OAK, then variance_applied made it TEX and enhanced_prediction retained TEXEarliest wrong direction: variance stage. Later enhancement preserved the error. The 0.12 edge was moderate
PHI@ARIHistorical ML favored ARI, then variance_applied moved to PHI and enhanced_prediction retained PHIEarliest wrong direction: variance stage. Starter and low-run-environment errors worsened the miss
SD@CINHistorical ML, variance, and enhanced stages all favored SDLean-side miss with modest starter and Cincinnati-offense error; likely variance/weighting plus irreducible outcome noise
SEA@BOSHistorical ML slightly favored SEA; variance_applied remained close, enhanced_prediction moved to BOSEarliest clear wrong-direction stage: enhanced prediction. Toss-up variance was important; winner confidence was only 0.04
SF@PITHistorical ML favored SF; variance produced a tie; enhancement moved to PITEarliest directional degradation began at variance through SF → tie, then enhancement completed the wrong-side move
STL@LADHistorical ML favored LAD; variance produced a tie; enhancement restored LADEarliest movement away from the actual winner began at historical ML itself; variance did not create the final wrong side but did remove a correct historical direction
TOR@CLEHistorical ML favored TOR; variance preserved TOR; enhancement moved to CLEEarliest wrong-direction stage: enhanced prediction. The later score environment did not recover it

Cause classification

Pipeline-stage quality across all stages

StageWinner accuracyWinner hitsTeam-run MAEMargin MAETotal-run MAE
historical_ml73.33%112.93943.87774.0109
variance_applied60.00%92.93333.73334.1333
enhanced_prediction33.33%53.06674.66674.0000
embedded_savant_run_environment33.33%53.13334.53334.2667
fundamental_edge_target_review33.33%53.13334.53334.2667
capped_target33.33%53.13334.53334.2667
simulation_result33.33%53.13334.53334.2667
published_final33.33%53.13334.53334.2667

Transitions:

A later unchanged or recovered stage does not erase the enhanced-stage regression.

5. Cross-game counterfactuals

Lever tension: score variance multiplier/floor

Lever: score.variance_multiplier / score.variance_floor Proposed direction: widen run distributions Targets: Brier score and calibration error

Helps:

Hurts: none listed in the slate-level tension record.

Net evidence: 4 helps, 0 hurts. This is favorable, but the counterfactual is mainly about distribution and calibration, not guaranteed winner improvement. It should be tested as a bounded shadow experiment.

Lever tension: pitcher form/prior weighting

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

Helps:

Hurts or creates risk:

Net evidence: 7 helps versus 6 hurts. No broad move is supported. The lever risks disturbing several games where starter projection was already within one ER.

Lever tension: recent/season run-mean weighting

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

Helps or could reduce error in:

Hurts or risks disturbing:

Net evidence: 12 helps versus 1 hurts. This is favorable for run MAE and total-run MAE, but the cross-slate weak-R² finding argues for isolated training/backtesting rather than an immediate production weighting change.

One-game fixes that must not be made

6. Data and contract defects

Audit findings by severity

Critical

No structural audit file reports a critical finding. However, the causal register contains one critical contract failure:

Error

Warning

  1. OAK@TEX — Cody Bradford pitch-count mismatch
  2. Summary H 3 versus box H 2
  3. Summary BF 19 versus box BF 18
  4. Rule: PITCH_COUNT_PITCHER_MATCH
  5. Source: simulation
  6. Risk score: 10
  1. PHI@ARI — Andrew Painter pitch-count mismatch
  2. Summary H 4 versus box H 3
  3. Summary BF 21 versus box BF 20
  4. Rule: PITCH_COUNT_PITCHER_MATCH
  5. Source: simulation
  6. Risk score: 10

These are artifact/stat-contract defects, not demonstrated prediction causes.

Informational

Model defects versus presentation/artifact defects

Prediction-model/process defects

Presentation/artifact/contract defects

The evidence does not support blaming the two pitch-count warnings for the misses.

Probability-lineage audit

The available lineage records contain the following fields:

Most games have a recorded transformation such as “score coherence followed by matchup-recency guardrail,” and the governor is marked applied.

Material bypass:

Non-material traceability issue:

No other material unexplained publication delta is recorded.

7. Cross-slate persistence

The trend window covers six slates, with five prior slates loaded.

Escalated findings

These are established failures, not single-slate noise.

Score-margin compression

Simulation inertness

Calibration instability

Weak run-model signal

Finding streaks

Raw structural blocks

Margin distribution

Pipeline transitions

Calibration monitoring

Current buckets remain overconfident:

The trend direction is marked flat, but at a poor level rather than a healthy level.

High raw cluster

Simulation control

Metric-trend directions

Top-confidence-pick loss streak

The top-confidence pick has a current loss streak of 2:

8. Adjustments for today

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

1. Bounded shadow experiment for persistent score compression

score.variance_floor could be tested separately, not simultaneously, to preserve causal interpretability.

2. Bounded shadow experiment for persistent calibration overconfidence

Do not simultaneously change confidence.base_cap or confidence.moderate_cap; those are alternative shadow levers.

3. Code/diagnostic action for six-slate simulation inertness

4. Isolated training/backtest for six-slate weak run-model signal

5. Winner-pipeline diagnostic for the strongest single-slate driver

6. Critical probability-lineage remediation

9. Bottom line

Known:

Uncertain:

The single safest next action is:

Block promotion and trace the complete probability lineage for SF@PIT while simultaneously preserving before/after winner picks at every pipeline stage, especially the variance_applied → enhanced_prediction transition.

That action addresses the only critical contract failure and the strongest single-slate process risk without making an unvalidated production tuning change.