Luna deep analysis — slate 20260903

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

1. Evidence receipt and gaps

Sources received

All six available source-manifest artifacts were marked loaded:

  1. Prediction review — /app/output/reports/prediction_review_20260903.json, 225,604 bytes; official outcomes and scoring errors.
  2. Forensic deep audit — /app/output/reports/PREDICTION_SYSTEM_DEEP_AUDIT_20260903.json, 151,096 bytes; probability, provenance, starter, and decision contracts.
  3. Continuous learning — /app/output/learning/continuous_learning_daily_20260903.json, 1,175,839 bytes; pipeline-stage quality, calibration, probability lineage, score distributions, and official starter comparisons.
  4. Detailed report — /app/output/reports/detailed_20260903.html, 215,805 bytes; game and simulation rendering.
  5. Ultra-detailed player audit — /app/output/reports/20260903_ultra_detailed_player_audit.html, 18,544,069 bytes; player-level prediction trace.
  6. Structural audit — /app/output/reports/audit_report_20260903.html, 1,125,883 bytes; invariant findings.

One manifest entry was missing:

That missing artifact is a receipt gap. It does not establish that the predictions were wrong, but it prevents a complete independent verification of revision immutability and post-Luna remediation claims.

Slate coverage

All nine expected games were reviewed:

Coverage is internally complete for the available evidence:

Contradictions and unresolved gaps

  1. MIA@KC probability publication delta: calibrated pick probability was 0.5271, while published pick probability was 0.547642, a +0.020542 delta. The lineage explicitly says there is “no recorded transformation” explaining it. This is the registered critical issue.
  2. MIA@KC probability representation: the game-level diagnostic says published home probability was 0.452, consistent with a 0.547642 away-pick probability. Thus, the numerical values may be complementary side representations rather than a mathematical contradiction. The problem is that the transformation and side orientation were not recorded clearly enough to reconcile the value.
  3. MIL@CHC structural warning: Logan Henderson’s pitch-count summary recorded 3 hits allowed versus 2 in the pitching box score. This is a warning-level source inconsistency, not a demonstrated cause of the miss.
  4. Artifact warning: the aggregate structural audit reports one warning and 68 informational findings, but the supplied aggregate does not identify the exact warning. The per-game evidence identifies the MIL pitch-count warning; the canonical artifact warning should still be reconciled explicitly.
  5. No player-level actuals: MIA@KC, MIL@CHC, OAK@SEA, SF@PIT, and TB@TEX state that per-player actuals were unavailable. Player-level attribution for those games is therefore limited to projected adjustments and team/starter outcomes.
  6. No market cross-check: no source-labelled moneyline was available for any game. Market agreement or disagreement cannot be assessed.
  7. No actual causal proof from one slate: several causal claims are appropriately framed as drivers or risks rather than proven causal effects. Persistent findings in the cross-slate section are different: those have been escalated and must be treated as established operational failures.

2. Executive diagnosis

Outcome summary

The system went 5–4, with winner accuracy of 55.56%.

That result is only modestly above guessing and below the validated historical walk-forward baseline:

The self-assessment correctly says the slate beat coin-flipping but did not beat the validated baseline.

Winner accuracy

Winner accuracy is therefore weak-to-flat, not a success merely because five calls were correct.

Calibration quality

Calibration was unstable and poor in the most consequential bucket:

The confidence system was not demonstrated to be globally anti-predictive, but the strong-confidence loss and the cross-slate confidence findings justify bounded confidence reduction and wider score distributions. A published home-away edge of 0.12, such as CHW@HOU, is moderate—not a near-coin-flip justification for ignoring all directional information, and not strong evidence of a robust edge.

Score accuracy

Score accuracy was the weakest model dimension.

Slate-level facts:

The run models have near-zero holdout explanatory power:

This supports treating exact scores as low-confidence estimates. It does not prove that every specific score miss came from the run model, nor does it justify arbitrary score-tail tuning without backtesting.

Starter accuracy

Starter projections were materially wrong in several losses:

The starter-weighting lever has a mixed slate result: four games identified as helped versus three games potentially harmed. That is not strong enough to justify an aggressive global reweighting from this slate alone, especially because the persistent cross-slate run-model findings are stronger than this one-slate starter balance.

Artifact integrity

Artifact integrity was mostly good but not clean:

The distinction is important: prediction artifacts can pass structural QA while still containing a material probability-publication contract defect.

Causal issue register reconciliation

Critical: unexplained published probability delta

This is a contract/evidence failure, not proof that the wrong side was selected because of the publication delta.

High: artifact-integrity warning

High: weak run-model signal

Informational: complete audit coverage

Informational: simulation inertness

3. Every game

BOS@BAL — correct winner, wrong score

Evidence edges and controls

Starters

Score path

Historical ML was approximately BOS 4.95, BAL 4.03; variance stage rounded to 5–4. Later stages did not change the published score. The score missed by one run for each team but retained the correct one-run margin.

Player adjustments and audit

No player-level adjustments were recorded. Savant adjustment was applied. Structural audit: PASS, zero defects, seven informational findings. Forecast-file integrity was clean.

CHW@HOU — correct winner, wrong score

Evidence edges and controls

Starters

Score path and audit

Predicted margin was one run; actual margin was four. No player-level adjustments were listed. Savant adjustment was applied. Structural audit: PASS, zero defects, eight informational findings.

MIA@KC — winner miss

Evidence edges and controls

Starters

Score path

Historical ML: MIA approximately 4.95, KC 4.03. The model retained MIA 5–4 through variance, enhancement, Savant, fundamental review, cap, and simulation. The simulation did not alter the winner or score.

Player adjustments and audit

Twenty-four material player changes were recorded. Examples include:

Actual player-level results were unavailable, so these adjustments cannot be validated individually. Structural audit: PASS, zero defects, seven informational findings; forecast integrity clean.

MIL@CHC — winner miss and largest score failure

Evidence edges and controls

Starters

Score path

This was the clearest score-path deterioration:

The transition from variance to enhanced prediction increased the score total by four runs while preserving the wrong MIL direction. Later stages did not recover the error.

Player adjustments and audit

Twenty-two material player changes were recorded, including:

Actual player-level results were unavailable.

Structural audit: WARNING, risk score 10. The concrete warning was a pitch-count mismatch for Logan Henderson: summary hits allowed 3, box score 2. It was not identified as the reason for the winner miss.

OAK@SEA — strong-confidence winner miss

Evidence edges and controls

Starters

Score path

The score and winner moved toward SEA before the later stages, and no later stage corrected that direction.

Player adjustments and audit

Twenty-six material player changes were recorded. Examples:

Actual player-level results were unavailable. Structural audit: PASS, zero defects, seven informational findings; forecast integrity clean.

SF@PIT — correct winner, wrong score

Evidence edges and controls

Starters

Score path and audit

Prediction remained 5–6; actual PIT won 5–2. No player-level actuals were available. Savant adjustment applied. Structural audit: PASS, zero defects, seven informational findings; forecast integrity clean.

STL@LAD — correct winner, wrong score

Evidence edges and controls

Starters

Score path and audit

The system correctly selected LAD but projected a five-run total versus the actual five-run total? The predicted total was 9 and actual total 5, so total bias was +4. Structural audit: PASS, zero defects, seven informational findings; forecast-file integrity clean.

TB@TEX — winner miss

Evidence edges and controls

Starters

Score path and audit

Prediction remained TB 5–4 from historical ML through publication. Actual was TB 0–TEX 6. No player-level actuals were available. Twenty-four material player changes were recorded, including Jorge Mateo, Ezequiel Duran, Cal Quantrill, Corey Seager, and others, but their actual effects cannot be validated.

Structural audit: PASS, zero defects, seven informational findings; forecast integrity clean.

TOR@CLE — correct winner, wrong score

Evidence edges and controls

Starters and audit

4. Miss reconstruction

MIA@KC

Earliest wrong-direction stage

The historical ML stage already pointed toward the wrong result:

The earliest model movement toward the wrong result was therefore the historical model / pre-enhancement directional baseline, reinforced by season-quality and head-to-head treatment.

Error classification

The −0.0953 home-minus-away edge was a lean, not strong conviction. It should not be interpreted as a near coin flip, but it also did not support aggressive confidence.

MIL@CHC

Earliest wrong-direction stage

Historical ML already favored MIL:

The first clear score-quality regression occurred at the variance-to-enhanced transition: total score rose from 9 to 13 runs, while winner direction remained wrong. The winner-direction error itself was present at historical ML.

Error classification

OAK@SEA

Earliest wrong-direction stage

The historical ML stage favored SEA 5.04–4.02. Variance retained SEA at 5–4, and enhancement widened SEA’s margin to 6–4. No later stage corrected it.

Thus the earliest wrong-direction evidence was the historical model, with later enhancement and season-quality weighting reinforcing the error.

Error classification

TB@TEX

Earliest wrong-direction stage

Historical ML favored TB approximately 4.56–3.63. All stages retained TB at 5–4. No stage moved toward TEX.

The wrong direction therefore entered at the historical model, and the later controls failed to reverse it.

Error classification

Pipeline-stage audit across all stages and transitions

StageWinner accuracyWinner hitsTeam-run MAETotal-run MAEMargin MAEChange versus historical
historical_ml55.56%52.00062.50172.7246baseline
variance_applied44.44%42.05562.55562.7778worse
enhanced_prediction55.56%52.33333.11112.8889worse
embedded_savant_run_environment55.56%52.33333.11112.8889unchanged from enhanced
fundamental_edge_target_review55.56%52.33333.11112.8889unchanged
capped_target55.56%52.33333.11112.8889unchanged
simulation_result55.56%52.33333.11112.8889unchanged
published_final55.56%52.33333.11112.8889unchanged

Transitions:

  1. Historical → variance: BOS moved BOS → TIE. No games improved. Winner accuracy fell by one hit, team-run MAE rose 0.055, and margin MAE rose 0.0532.
  2. Variance → enhanced: BOS recovered TIE → BOS. No games regressed in winner direction, but team-run MAE rose 0.2777 and margin MAE rose 0.1111.
  3. Enhanced → embedded Savant: no changes.
  4. Embedded Savant → fundamental review: no changes.
  5. Fundamental review → capped target: no changes.
  6. Capped target → simulation: no changes.
  7. Simulation → published final: no changes.

The BOS recovery does not erase the intermediate variance-stage regression. Likewise, later unchanged stages do not demonstrate that they validated the earlier score expansion. The simulation contributed no winner-direction intervention on this slate.

5. Cross-game counterfactuals

Lever tension: widen score distributions

Lever tension: pitcher form weights and prior games

Lever tension: recent/season run-mean weighting

One-game fixes that must not be made

6. Data and contract defects

Critical findings

Unexplained probability publication delta

This is a prediction-contract/evidence defect, not necessarily a model-side selection defect.

Error findings

Warning findings

  1. MIL@CHC pitch-count mismatch
  2. Logan Henderson’s summary reports hits allowed 3; pitching box score reports 2.
  3. Rule: PITCH_COUNT_PITCHER_MATCH.
  4. Source: simulation.
  5. Risk score: 10.
  6. This is a data/audit consistency defect. It was not demonstrated to cause the prediction miss.
  1. Aggregate artifact warning
  2. Structural totals contain one warning.
  3. The supplied evidence identifies the MIL mismatch as the concrete warning, but the aggregate canonical warning should be reconciled explicitly.
  1. Missing integrity revision audit
  2. Not marked as a structural warning in the totals, but it is a receipt and artifact-integrity gap because the manifest explicitly marks the revision-proof artifact missing.

Informational findings

Model defects versus presentation/artifact defects

Prediction-model/process defects

Presentation/artifact/contract defects

Complete pick-side probability lineage audit

The evidence provides the following chain for all nine games:

Eight games have a recorded transformation and unexplained_material_delta: false. MIA@KC is the sole material bypass/lineage exception.

Important examples:

A bypass should be called only for MIA because its material delta lacks a recorded transformation. The other deltas are documented, even where the values appear surprising.

Coverage matrix reconciliation supports complete game coverage, but not a clean slate: full structural coverage coexists with the MIA contract defect, the MIL warning, and the missing integrity revision artifact.

7. Cross-slate persistence

Six prior slates were loaded, forming a seven-slate window from 2026-08-28 through 2026-09-03. Persistent escalations are not single-slate noise.

Finding streaks

FindingConsecutive-slate countPresent now?Interpretation
high_raw_probability_cluster0NoOccurred in 2 of 7 slates, but not currently persistent
pipeline_winner_instability0NoOccurred in 3 of 7, not currently present
score_margin_compression0NoOccurred in 3 of 7, not currently present as a registered streak, though this slate shows the same structural symptom
simulation_inert_on_slate7YesEscalated established failure
single_slate_calibration_instability0NoOccurred in 5 of 7, but no current consecutive streak
weak_run_model_signal7YesEscalated established failure

The current slate independently exhibits margin compression, but the finding-streak record reports score_margin_compression with zero consecutive slates. Those two statements should not be conflated: the raw current-slate margin data show compression, while the registered streak counter does not classify it as currently persistent.

Escalated finding: simulation inertness

The self-assessment identifies a concrete root cause: the same draws are used to search for a target-realizing play-by-play, and the loop may early-exit after finding one, with a default draw count around two. That makes the resulting win rate non-probabilistic and the flip signal vacuous. This requires operator/code action.

Escalated finding: weak run-model signal

The correct action is not to blindly alter the three training hyperparameters. The registered response is a feature-contract and target investigation involving market no-vig probability where available, park/weather, bullpen rest/fatigue, and platoon-adjusted lineup quality, followed by same-fingerprint holdout comparison.

Metric trends

All directions must be stated explicitly:

The top-confidence loss streak is a material warning even though the high-raw cluster is not present this slate.

Raw structural evidence

8. Adjustments for today

Recommendations are ranked by persistence and evidence strength first, then scope and regression risk.

1. Repair and independently validate the simulation path

2. Investigate the persistent near-zero run model

3. Apply a small score-variance shadow increase for the strongest single-slate driver

4. Reconcile MIA probability lineage before any promotion

5. Do not globally reweight pitcher form today

9. Bottom line

Known

Uncertain

Safest next action

The single safest action is to repair and independently verify the simulation/probability pipeline before promotion: separate score-target realization from an independently sampled winner-probability distribution, eliminate the early-exit/no-op behavior, register the draw-count parameter, and reconcile MIA’s published probability lineage. This directly addresses a seven-slate established failure and a critical evidence-contract defect without making an unvalidated global model adjustment.