Luna deep analysis — slate 20260907

Coverage receipt: COMPLETE · 9/8 sections; 11/11 games; 5/5 causes; 5/6 sources

1. Evidence receipt and gaps

Sources received

All seven source_manifest entries were checked against the supplied evidence:

SourceStatusPurpose
prediction_review_20260907.jsonLoadedOfficial outcomes and scoring errors
PREDICTION_SYSTEM_DEEP_AUDIT_20260907.jsonLoadedProbability, provenance, starter, and decision contracts
continuous_learning_daily_20260907.jsonLoadedPipeline-stage quality, calibration, probability lineage, score distribution, and starter comparisons
detailed_20260907.htmlLoadedFull game and simulation rendering
20260907_ultra_detailed_player_audit.htmlLoadedPlayer-level prediction trace
audit_report_20260907.htmlLoadedInvariant findings
audit_report_20260907_integrity_revision.htmlMissingPost-Luna remediation proof for immutable revision artifacts

The missing integrity-revision audit is the only explicitly missing source. It prevents independent confirmation of immutable revision-artifact remediation, although the supplied structural and per-game audits report no critical or error findings.

Slate coverage

All 11 expected games were reviewed:

ARI@KC, ATL@PHI, CHC@MIL, CIN@LAD, CLE@BAL, LAA@BOS, MIN@DET, NYM@MIA, STL@SF, TOR@OAK, and WAS@SD.

Coverage is internally complete for the principal slate artifacts:

This agrees with coverage_matrix.all_games_structurally_audited=true.

Gaps and contradictions

  1. Missing integrity revision audit. integrity revision audit is marked missing; no bytes are supplied.
  1. Four unexplained probability-publication deltas. The causal register identifies ARI@KC, CIN@LAD, LAA@BOS, and NYM@MIA as having unexplained calibrated-to-published deltas of at least two percentage points. This is a critical evidence/contract failure until traced.
  1. Probability-lineage representation is not uniformly intuitive. Some records show a calibrated probability above 0.50 while the published pick probability is below 0.50 because the selected side changes from the home side to the away side. That is not necessarily a defect, but the lineage must clearly identify whether each probability is:
  2. probability of the home team,
  3. probability of the selected side, or
  4. probability after side inversion.
  1. ARI starter identity inconsistency. The game-level starter line lists away starter Jose Cabrera, with identity_match=false, actual 0.00 ERA over 6 outs, and no settled pitching line. The player-forensics block instead identifies Noah Cameron as the relevant Kansas City starter and explicitly reports the Arizona starter as missing. This is a real data-reconciliation issue, even though that game’s structural audit reports zero findings.
  1. Audit finding names are not fully supplied. The structural summary reports 0 critical, 0 error, 0 warning, and 81 info, but the individual 81 informational finding texts are not present in the supplied object. They therefore cannot be enumerated by title without inventing evidence.
  1. No market cross-check. No source-labelled moneyline was available for any game. Consequently, market agreement or disagreement cannot be evaluated.
  1. No actual player results may be inferred beyond supplied lines. The analysis below uses only the listed player, starter, lineup, and run-deviation records.

2. Executive diagnosis

Overall result

The slate finished 6–5, or 54.55% winner accuracy. That is below the supplied walk-forward baseline accuracy of 55.53% and below the historical validation winner accuracy of 59.16%.

The system’s own headline is therefore directionally correct:

The slate should not be described as a good prediction merely because it got six winners correct.

Winner accuracy

Winner performance was weak but not catastrophically below chance:

The five misses were: ARI@KC, MIN@DET, NYM@MIA, STL@SF, and TOR@OAK.

Calibration quality

Calibration was poor on this slate:

The current slate’s confidence was anti-predictive at the upper end: strong convictions lost three of four times. This is material calibration evidence, but the causal register correctly warns not to fit a stable calibration curve from one slate.

The high-raw cluster was specifically:

That is a warning against treating raw or governed confidence near 0.70 as reliable.

Score accuracy

Score accuracy was materially worse than winner accuracy:

The exact-score system is therefore low-confidence. A correct winner frequently coexisted with a poor score forecast.

Starter accuracy

Starter projections were highly volatile:

Artifact integrity

Artifact integrity was generally good, but not clean in the broadest contractual sense:

Therefore:

Causal issue register reconciliation

Critical: unexplained_probability_publication_delta

This is a contract failure, not merely a calibration problem.

High: single_slate_calibration_instability

High: weak_run_model_signal

Info: complete_slate_audit_coverage

Info: simulation_inert_on_slate

3. Every game

The probability below is the published probability for the selected side unless otherwise noted. A positive confidence edge means the home side was favored; a negative edge means the away side was favored.

ARI@KC — miss; wrong winner and wrong score

Edges and controls

Starter projection versus actual

Score path and player adjustments

Audit and causal assessment

ATL@PHI — correct winner, wrong score

Edges and controls

Starters

Score path and player adjustments

Audit and causal assessment

CHC@MIL — correct winner, poor score

Edges and controls

Starters

Score path and player adjustments

Audit and causal assessment

CIN@LAD — correct winner, wrong score

Edges and controls

Starters

Score path and player adjustments

Audit and causal assessment

CLE@BAL — correct winner, wrong score

Edges and controls

Starters

Score path and player adjustments

Audit and causal assessment

LAA@BOS — strong correct winner, close score

Edges and controls

Starters

Score path and player adjustments

Audit and causal assessment

MIN@DET — miss; wrong winner and wrong score

Edges and controls

Starter projections

Score path and player adjustments

Audit and causal assessment

NYM@MIA — miss; wrong winner and wrong score

Edges and controls

Starters

Score path and player adjustments

Audit and causal assessment

STL@SF — strong miss; wrong winner and wrong score

Edges and controls

Starters

Score path and player adjustments

Audit and causal assessment

TOR@OAK — strongest miss; wrong winner and wrong score

Edges and controls

Starters

Score path and player adjustments

Audit and causal assessment

WAS@SD — correct winner, wrong score

Edges and controls

Starters

Score path and player adjustments

Audit and causal assessment

4. Miss reconstruction

Earliest wrong-direction stages

LossEarliest stage moving toward wrong resultReconstruction
ARI@KCHistorical ML already favored KCThe wrong side was present before later controls; subsequent stages maintained KC. Starter identity/feed inconsistency is a data-quality concern, but no later winner-direction flip is documented.
MIN@DETenhanced_predictionHistorical ML was approximately tied; variance stayed tied; enhanced prediction changed TIE → MIN, which was wrong.
NYM@MIAHistorical ML already favored MIAVariance, enhanced, Savant, and simulation all maintained MIA. This was a small-edge miss rather than a later regression.
STL@SFenhanced_predictionHistorical ML and variance favored SF; enhanced prediction changed SF → STL, and later stages preserved the wrong side.
TOR@OAKHistorical ML already favored TORAll stages reinforced TOR, with Savant and score-volatility adjustments widening the wrong margin.

A confidence edge of 0.12 would be moderate, not a near coin flip. More generally, the system’s own edges must not be interpreted as half-scale distances from 50%. The ARI edge was approximately 0.16 and still lost; STL and TOR had much stronger edges and also lost. Conversely, several 0.01–0.06 edges won because they were genuinely uncertain calls.

Loss-type classification

Bad input

Bad weighting or transition

Calibration

Starter error

Offense/bullpen miss

Run-environment miss

The slate overprojected total scoring in:

It underprojected:

This supports the persistent weak-run-signal diagnosis rather than a single consistent high- or low-scoring bias.

Irreducible variance

The player forensics identify:

Several major starter outcomes were also explicitly labelled outliers. A model change based on every individual over- or underperformance would learn noise.

Pipeline-stage quality across all stages

The stored stage metrics were:

StageWinner accuracyTeam-run MAETotal-run MAEMargin MAE
historical_ml6/11 = 0.54551.57822.34212.1950
variance_applied5/11 = 0.45451.59092.45452.0909
enhanced_prediction6/11 = 0.54551.81822.90912.0000
embedded_savant_run_environment6/11 = 0.54551.90913.09092.1818
fundamental_edge_target_review6/11 = 0.54551.90913.09092.1818
capped_target6/11 = 0.54551.90913.09092.1818
simulation_result6/11 = 0.54551.90913.09092.1818
published_final6/11 = 0.54551.90913.09092.1818

Transition audit

  1. Historical ML → variance applied
  2. Winner accuracy fell from 6 to 5.
  3. CIN@LAD, CLE@BAL, and MIN@DET moved toward ties.
  4. CLE@BAL is explicitly listed as regressed.
  5. Margin MAE improved by 0.1041, but winner accuracy worsened by 0.091.
  1. Variance applied → enhanced prediction
  2. Winner accuracy recovered from 5 to 6.
  3. CIN@LAD and CLE@BAL improved.
  4. STL@SF regressed from SF to STL.
  5. Team-run MAE increased by 0.2273.
  6. This transition contains a documented intermediate winner-direction regression; the later aggregate recovery does not erase it.
  1. Enhanced prediction → embedded Savant run environment
  2. No winner-direction changes.
  3. Team-run MAE worsened by 0.0909.
  4. Total-run MAE worsened to 3.0909.
  5. Margin MAE worsened by 0.1818.
  1. All later transitions
  2. No winner-direction changes.
  3. No measurable score changes.
  4. simulation_result → published_final made no change to any recorded metric.

The simulation therefore did not intervene in winner direction on this slate, and this same inertness has persisted for 11 consecutive slates.

5. Cross-game counterfactuals

Lever tension: widen run distributions

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

Games helped

Game hurt

Net evidence

Lever tension: pitcher form weights and prior-game shrinkage

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

Games helped

Games hurt

Net evidence

Lever tension: recent/season run-mean weighting

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

Games helped

Games hurt

Net evidence

Tempting one-game fixes that must not be made

  1. Do not reverse the STL pick rule. STL@SF was a severe miss, but a single game does not justify suppressing strong negative composite edges.
  1. Do not permanently downgrade Toronto or Oakland. TOR@OAK was the strongest miss, but the evidence supports broader confidence/variance control, not a team-specific rule.
  1. Do not tune to individual player box scores. Only 18 of 72 material player deviations were supported by player contact-quality evidence; the other 54 are classified as variance.
  1. Do not remove starter adjustments entirely. Several games show useful starter information, and the pitcher-form lever helped four games with no listed hurts.
  1. Do not use one slate to alter calibration weights aggressively. The causal register explicitly says the slate is too small for a stable confidence curve.
  1. Do not assume the probability governor was bypassed. Trace pick-side math for the four unexplained deltas first.

6. Data and contract defects

Critical findings

1. Unexplained published probability transformations

Affected games:

The lineage says “no recorded transformation explains the publication delta” for all four. This is a critical contract failure because a material probability transformation is missing from evidence.

Error findings

No error findings were reported in the supplied structural audits:

This does not mean the four probability-lineage issues are resolved; they are raised in a separate causal contract register.

Warning findings

No warning findings were reported:

Informational findings

Prediction-model defects versus artifact bugs

Prediction-model defects or risks

Presentation/artifact/contract bugs

The per-game forecast files themselves were reported clean, so the primary defect is not a rendering or saved-file corruption issue.

Complete probability lineage audit

The required conceptual path is:

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

Correctly recorded transformations

Examples:

Material unresolved transformations

The following have no recorded transformation explaining the published delta:

These are bypass-like symptoms, but they must not be called a governor bypass until pick-side math is traced. The causal register’s guardrail is appropriate.

No bypass conclusion for other games

No material unexplained delta is present for:

For those games, the published value matches the calibrated/governed or documented post-governor value within rounding.

Coverage reconciliation

Coverage is complete at the game level:

The slate is therefore structurally covered, but not contractually clean because coverage does not resolve the four probability deltas, the missing integrity revision audit, or ARI’s starter inconsistency.

7. Cross-slate persistence

The supplied cross-slate window contains 11 slates, with 10 prior slates loaded.

Escalated findings — established failures

These are not single-slate noise.

simulation_inert_on_slate

The self-assessment provides a specific suspected cause: the simulation distribution is built from the same draws used to search for a target-realizing play-by-play, and the loop can terminate after approximately two selected draws. That makes the estimated win rate non-probabilistic and explains zero flip candidates over 11 slates.

weak_run_model_signal

The raw evidence remains poor:

The self-assessment specifically recommends adding, where valid and source-labelled:

It also requires a model schema-version bump when feature semantics change.

Every finding streak

FindingConsecutive countIncidencePresent now?Interpretation
high_raw_probability_cluster03/11NoHistorical recurrence, not current
pipeline_winner_instability04/11NoHistorical recurrence, not current
score_margin_compression03/11NoHistorical recurrence, not current
simulation_inert_on_slate1111/11YesEstablished persistent failure
single_slate_calibration_instability16/11YesCurrent finding, not yet persistent by consecutive-slate count
weak_run_model_signal1111/11YesEstablished persistent failure

The fact that single_slate_calibration_instability occurred in 6 of 11 slates but has only a one-slate consecutive streak is important: its incidence is notable, but it has not reached the escalation status of the two 11-slate failures.

Metric trends

Each reported direction is:

Top-confidence-pick loss streak

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

The full sequence includes several earlier misses, but the supplied explicit streak is two consecutive losses. This is concerning when combined with the current high-conviction miss rate of 75%, though still weaker evidence than the 11-slate escalations.

Cross-slate bias findings

The two listed cross-slate bias findings are not actionable yet:

  1. Starter ERA error positive
  2. Mean: +0.281
  3. Same-sign share: 0.625
  4. 5 of 8 slates
  5. z = 0.95
  6. Status: hold
  7. No change to context.opponent_strength_weight
  1. Lineup run delta negative
  2. Mean: -0.06
  3. Same-sign share: 0.625
  4. z = -0.3
  5. Status: hold
  6. No change to score.variance_floor

The source correctly treats these as not separable from slate-to-slate noise.

8. Adjustments for today

Recommendations are ranked by persistence and scope. Persistent issues receive code/diagnostic actions rather than forced parameter moves, because the supplied escalation records explicitly state that they are not parameter-actionable.

1. Simulation inertness — highest persistence and scope

2. Weak run-model signal — persistent feature/target failure

3. Upper-tail confidence overstatement — strongest current single-slate driver

4. Score-distribution shadow experiment

5. Starter venue/form experiment

6. No direct change to context.opponent_strength_weight

The cross-slate evidence says the positive starter ERA bias is not separable from noise:

This lever should remain unchanged pending more evidence.

9. Bottom line

Known

Uncertain

Safest next action

Repair and independently validate the simulation probability path before promoting any model or confidence change. The simulation inertness is persistent, has a concrete code-level explanation, and can be tested without committing to a risky one-game or one-slate parameter adjustment. In parallel, keep confidence-cap and score-variance changes in shadow mode only, with Brier score, calibration error, winner accuracy, team-run MAE, and total-run MAE monitored for rollback.