Luna deep analysis — slate 20260914

Coverage receipt: COMPLETE · 9/8 sections; 10/10 games; 5/5 causes; 5/6 sources

1. Evidence receipt and gaps

Source receipt

All seven source-manifest entries were reviewed as declared:

SourceStatusDeclared purpose
prediction_review_20260914.jsonLoadedOfficial outcomes and scoring errors
PREDICTION_SYSTEM_DEEP_AUDIT_20260914.jsonLoadedProbability, provenance, starter, and decision contracts
continuous_learning_daily_20260914.jsonLoadedStage quality, calibration, lineage, score distribution, and starter comparisons
detailed_20260914.htmlLoadedFull game and simulation rendering
20260914_ultra_detailed_player_audit.htmlLoadedPlayer-level prediction trace
audit_report_20260914.htmlLoadedInvariant findings
audit_report_20260914_integrity_revision.htmlMissingPost-Luna remediation proof for immutable revision artifacts

The missing integrity-revision artifact is the only explicit source-manifest gap. It does not invalidate the game outcomes or structural audits, but it prevents verification of the claimed post-remediation immutable-revision proof.

Slate coverage

The slate contained 10 games, and all 10 were reviewed:

ATL@CHC, BAL@NYM, CHW@CLE, DET@TOR, LAD@CIN, MIA@ARI, NYY@MIN, SD@COL, SEA@LAA, and SF@STL.

Coverage was complete for the ordinary prediction artifacts:

The aggregate structural audit reports 0 critical, 0 error, 0 warning, and 74 informational findings. Every game-level structural audit passed with zero defects.

Internal contradictions and gaps

  1. Missing integrity-revision audit. The manifest marks the artifact missing even though all ordinary structural audits passed.
  2. Probability-publication lineage is incomplete for two games. BAL@NYM and MIA@ARI have unexplained calibrated-to-published changes of at least two percentage points. This is explicitly registered as critical. It is not yet evidence that the governor was bypassed.
  3. The lineage contains smaller unexplained deltas. ATL@CHC has a published-minus-calibrated delta of -1.1176 percentage points, marked nonmaterial and noncritical. The register’s critical threshold is ≥2 percentage points.
  4. Starter comparison is technically complete but not semantically complete for every starter. CHW@CLE lacks a settled projection line for Sean Newcomb; MIA@ARI lacks one for Corbin Burnes. Therefore, “10/10 starter comparisons” means the game had a starter-comparison record, not that every starter had a complete projected line.
  5. The reported “confidence edge” is not uniformly the half-distance from 50%. The game diagnostics explicitly warn against treating a 0.12 edge as a near coin flip. The published probability and the signed home-minus-away edge must be interpreted separately.
  6. Simulation evidence is operationally weak. The simulation generated zero winner flips and zero flip candidates, and this has persisted for 11 slates. That is a persistent diagnostic/code concern, not evidence that simulation had no theoretical value.

2. Executive diagnosis

Overall result

The winner result was acceptable, but it was materially better than the score and probability quality. Correct winners must not be treated as evidence that the score forecasts, confidence levels, or internal causal pipeline were sound.

Winner accuracy

The model correctly selected seven winners:

It missed:

The strongest winner call was NYY@MIN, with a 0.14 away edge, strong negative home composite evidence, and a correct 8–3 prediction direction. However, the model’s top-confidence selection lost on this slate: BAL@NYM was the top-confidence miss. The loss streak for the top-confidence pick is currently 1 slate.

Calibration quality

Calibration was poor on this slate:

The register correctly classifies this as high, not critical: the evidence shows poor single-slate alignment, but the sample is too small to justify changing calibration weights immediately. The high-raw cluster contained three games and went 2/3, but its empirical hit rates were only approximately 0.536–0.569 despite calibrated probabilities of 0.6705–0.7259. This is a warning against high raw confidence, but not a complete rolling calibration curve.

Score accuracy

Score accuracy was substantially weaker than winner accuracy:

The validated run models explain almost none of holdout run variance. This is a high model risk and a persistent 11-slate issue, not a critical evidence-contract failure.

Starter accuracy

Starter projections were mixed:

Starter errors were often variance rather than a clear systematic defect. However, SEA@LAA provides a credible process concern: Kade Anderson had only four prior appearances and produced a 10.38 ERA against a 1.80 projection. The registered recommendation to reduce confidence for thin-history starters is directionally reasonable, but the current cross-slate evidence does not establish a global starter-ERA bias. The cross-slate starter-ERA pattern is explicitly marked nonpersistent: only 6 of 12 slates shared the sign, with z = 0.32.

Artifact integrity

Artifact integrity was strong for the loaded artifacts:

However, integrity is not completely closed because:

Causal issue register reconciliation

Critical: unexplained published probability deltas

This is a contract/provenance failure, not necessarily a model-quality failure. The data show:

The transformation may be a legitimate confidence-control step, but it is not recorded in the lineage.

High: single-slate calibration instability

High: weak run-model signal

Informational: complete audit coverage

Informational: simulation inert on slate

3. Every game

The signed confidence edge below is the published home-minus-away edge. For away picks, the published pick probability is 1 - published_home_probability.

ATL@CHC — correct winner, wrong score

Evidence edges and controls

Starter comparison

Score path

Historical ML 4.27–4.61 → variance 4–5 → enhanced 4–7 → Savant 4–7 → fundamental review 4–7 → capped 4–7 → simulation 4–7 → published 5–6.

Player-level adjustments

Audit and causes

BAL@NYM — wrong winner and wrong score

Evidence edges and controls

Starter comparison

Both starters outperformed the run projections, but the game’s principal failure was the much lower-than-expected overall run environment.

Score path

Historical 4.27–4.61 → variance 4–5 → enhanced 4–7 → Savant approximately 4.03–7.50 → fundamental 4–7 → capped 4–7 → simulation 4–7 → published 4–7.

Player-level adjustments

Audit and causes

CHW@CLE — wrong winner and wrong score

Evidence edges and controls

Starter comparison

Score path

Historical 4.55–4.38 → variance 5–4 → enhanced 4–5, with margin-recency and one-run-margin controls → Savant approximately 3.81–4.98 → fundamental 4–5 → capped 4–5 → simulation 4–5.

Player-level adjustments

Audit and causes

DET@TOR — correct winner, wrong score

Evidence edges and controls

Starter comparison

Score path

Historical 4.55–4.38 → variance 5–4 → enhanced 4–5 → Savant 3.81–4.98 → fundamental 4–5 → capped 4–5 → simulation 4–5.

Player-level adjustments

The documented hitter deviations were modest. Dillon Dingler was -0.93 runs, while Spencer Torkelson was +0.78.

Audit and causes

LAD@CIN — correct winner, wrong score

A 0.16 away edge is moderate, not near coin-flip territory.

Evidence edges and controls

Starter comparison

Both starters were better than their projections, especially Lodolo relative to an unusually high projected ERA.

Score path

Historical approximately 5–4 → variance 6–4 → enhanced 8–3 → Savant approximately 7.78–2.83 → fundamental 8–3 → capped 8–3 → simulation 8–3.

Player-level adjustments

Audit and causes

MIA@ARI — wrong winner and wrong score

Evidence edges and controls

Starter comparison

Score path

Historical 4.80–4.71 → variance 5–5 → enhanced 5–4 → Savant approximately 5.49–4.19 → fundamental 5–4 → capped 5–4 → simulation 5–4.

Player-level adjustments

ARI hitters were systematically under-projected in the result:

This was a high-variance offensive game, with both teams exceeding many individual projections.

Audit and causes

NYY@MIN — correct winner, wrong score

The 0.14 away edge is moderate and should not be described as a near coin flip.

Evidence edges and controls

Starter comparison

Both starters outperformed projections.

Score path

Historical approximately 4–5 → variance 5–5 → enhanced 6–4 → Savant approximately 6.20–4.15 → fundamental 6–4 → capped 6–4 → simulation 6–4.

Player-level adjustments

Audit and causes

SD@COL — correct winner, wrong score

Evidence edges and controls

Starter comparison

Sugano was the principal pitching miss.

Score path

Historical approximately 5–5 → variance 6–5 → enhanced 8–5 → Savant approximately 7.74–4.67 → fundamental 8–5 → capped 8–5 → simulation 8–5.

Player-level adjustments

Audit and causes

SEA@LAA — correct winner, wrong score

Evidence edges and controls

Starter comparison

Score path

Historical approximately 4–4 → variance 5–4 → enhanced 5–6 → Savant approximately 5.23–6.04 → fundamental 5–6 → capped 5–6 → simulation 5–6.

Player-level adjustments

Audit and causes

SF@STL — correct winner, wrong score

Evidence edges and controls

Starter comparison

Score path

Historical approximately 4–5 → variance 4–5 → enhanced 4–5 → Savant approximately 3.85–4.99 → fundamental 4–5 → capped 4–5 → simulation 4–5.

Player-level adjustments

Audit and causes

4. Miss reconstruction

Earliest wrong-direction stage for each loss

BAL@NYM

CHW@CLE

MIA@ARI

Loss classification

GameBad inputBad weightingCalibrationStarter errorOffense/bullpen missRun-environment missIrreducible variance
BAL@NYMIncomplete explanation of publication transformationYes, conviction escalationYes, strong 0.26 edge lostMostly variance; both starters betterNYM offense -6 versus projectionPrimary; 11 projected runs vs 3 actualYes
CHW@CLEMissing settled Newcomb lineYes, enhanced stage reversed sideLow-confidence toss-upNewcomb line unavailable; Williams +1 ERCHW offense +1, CLE offense -3Secondary, only -1 total biasPrimary toss-up variance
MIA@ARIBurnes starter line incompleteYes, ARI offense underweightedToss-upAlcantara close; Burnes incompleteARI offense +4; bullpen contributedPrimary; 9 projected runs vs 15 actualYes

A confidence edge of 0.12, where present on other games, is moderate—not a near coin flip. In this slate, the 0.14 and 0.16 edges on NYY@MIN, LAD@CIN, and SD@COL were meaningful moderate edges, whereas CHW@CLE and SF@STL were approximately 0.01 toss-ups.

Pipeline-stage quality across all stages

StageWinner accuracyWinner hitsTeam-run MAETotal-run MAEMargin MAEChange versus historical ML
Historical ML60%6/102.08283.34331.8811Baseline
Variance applied50%5/101.95003.30001.7000Margin/team-run improved; winner regressed
Enhanced prediction70%7/102.10003.40002.0000Winner improved; runs/margin worsened
Embedded Savant environment60%6/102.15003.30002.3000Winner and margin regressed
Fundamental edge review70%7/102.10003.40002.2000Winner recovered; margin remained worse
Capped target70%7/102.10003.40002.2000No change
Simulation result70%7/102.10003.40002.2000No change
Published final70%7/102.10003.40002.2000No change

Important transition findings:

The recovery at the fundamental stage does not erase the intermediate Savant-stage regression. The pipeline had a genuine intermediate winner-direction failure even though the final published winner was restored.

The simulation’s zero-flip result is especially important: it did not independently validate or correct any winner on this slate.

5. Cross-game counterfactuals

Lever tension 1: widen run distributions

Lever tension 2: pitcher form weights and prior-game shrinkage

Lever tension 3: recent/season run-mean weighting

Tempting one-game fixes that must not be made

  1. Do not globally increase NYM confidence because of the strong recent-form signal in BAL@NYM; that would worsen calibration risk after the 0.26-confidence miss.
  2. Do not flip toss-up handling based on CHW@CLE; its published edge was only 0.0113 and the outcome was explicitly classified as toss-up variance.
  3. Do not broadly raise ARI offense weights solely because Nootbaar and Lawlar exceeded projections in MIA@ARI; only 21 of 55 material player deviations were supported by the hitter’s own season contact quality.
  4. Do not lower all starter projections because of Kade Anderson or Sugano; several starter projections were reasonably close, and the cross-slate starter bias is nonpersistent.
  5. Do not use the correct winners in LAD@CIN or SD@COL to claim the run model is sound; both had substantial score error despite correct direction.
  6. Do not treat the zero simulation flips as evidence that simulation should be removed. First determine whether the simulation is no-oping because of the documented early-exit/design issue.

6. Data and contract defects

Audit finding inventory

Across the structural audit summaries:

Every per-game structural audit passed. No structural invariant failure was recorded in the game artifacts.

The broader forensic findings are nevertheless material:

Critical contract/provenance finding

unexplained_probability_publication_delta

High model findings

single_slate_calibration_instability

weak_run_model_signal

Informational findings

complete_slate_audit_coverage

simulation_inert_on_slate

Prediction-model defects versus presentation/artifact bugs

Prediction-model/process defects

  1. Near-zero run-model signal.
  2. Persistent simulation no-op/inertness.
  3. Poor single-slate probability calibration.
  4. Thin-history starter uncertainty.
  5. Incomplete settled starter lines for Newcomb and Burnes.
  6. Potential overreliance on recent/season run means in the enhanced stage.
  7. Player-level projections have weak separation:
  8. H spread ratio 0.386
  9. H correlation 0.213
  10. 2B metric worse than flat baseline: MAE 0.264 versus flat MAE 0.231, with 17.4% overprojection.

Presentation/provenance/artifact defects

  1. Missing integrity-revision audit artifact.
  2. Unexplained publication transformations in two probability lineages.
  3. Inconsistent probability fields in MIA@ARI, where the control trace and lineage show different intermediate values.
  4. Starter-comparison records that exist structurally but lack a complete settled projection line for one starter in two games.

The clean forecast artifact does not prove the underlying model was correct. It only establishes that the saved artifact was not corrupted in the ordinary structural sense.

Complete pick-side probability lineage audit

The required chain is:

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

Games with fully recorded transformations

Materially incomplete lineages

No bypass should be asserted until these transformations are traced, but these are genuine contract defects that block promotion.

7. Cross-slate persistence

The trend file contains 11 prior/current slates, with 10 prior slates loaded. The current slate is therefore not isolated for persistent findings.

Finding streaks

FindingConsecutive-slate countPresent now?Interpretation
high_raw_probability_cluster0NoOne occurrence in window; not persistent
pipeline_winner_instability0NoTwo occurrences in window, but not current
score_margin_compression0NoOne occurrence in window; not persistent
simulation_inert_on_slate11YesEstablished persistent failure
single_slate_calibration_instability1YesCurrent and recurring in 5/11 window slates, but only one consecutive slate
weak_run_model_signal11YesEstablished persistent failure

The two 11-slate findings are escalated findings and must be treated as established failures, not as single-slate noise.

Escalated finding: simulation inertness

The self-assessment identifies a likely root cause: the simulation uses the same draws for target-realization search and win-probability estimation, and exits early after finding a target-realizing play. With approximately two selected draws, the calculated win rate cannot function as an independent probability. The recommended fix is code/diagnostic work in the simulation distribution builder and draw-loop early exit, not a forced nightly parameter move.

Escalated finding: weak run-model signal

The registered evidence is strong:

The recommended remedy is not nightly hyperparameter adjustment. The self-assessment proposes adding market no-vig probability where legitimately sourced, park/weather environment, bullpen rest/fatigue, and platoon-adjusted lineup quality, followed by same-fingerprint holdout testing and a model-schema version bump.

Current non-escalated calibration finding

single_slate_calibration_instability:

This is concerning, but the causal register correctly says not to change calibration weights from one slate alone. The self-assessment’s auto-action proposal to lower caps is therefore best treated as a bounded shadow experiment, not an unmonitored promotion.

Metric trends

These trends show an important divergence: winner accuracy and margin bias are improving, while calibration and loss-score error are worsening.

Top-confidence pick loss streak

Cross-slate bias holds

The two explicit cross-slate bias findings remain nonactionable:

  1. Starter ERA error:
  2. 6/12 slates share the sign.
  3. Mean 0.144, z = 0.32.
  4. Status: hold.
  5. Lineup run delta:
  6. 7/12 slates share the sign.
  7. Mean -0.163, z = -0.82.
  8. Status: hold.

They should not be treated as established persistent biases.

8. Adjustments for today

Recommendations are ranked by persistence, scope, expected metrics, affected games, and regression risk.

1. Fix and independently validate the simulation module

2. Run an isolated feature/target backtest for the run model

3. Bounded calibration-cap shadow experiment

4. Bounded run-variance shadow experiment

5. Thin-history starter shadow control

Defects that require abstention rather than tuning

9. Bottom line

Known

  1. The slate produced a 7–3 winner record.
  2. Winner accuracy was better than the validated baseline on this slate, but one slate cannot establish durable improvement.
  3. The three losses were:
  4. one strong-conviction run-environment failure: BAL@NYM;
  5. one genuine toss-up with a small wrong-side lean: CHW@CLE;
  6. one toss-up with a six-run underprojection and incomplete starter data: MIA@ARI.
  7. Exact run prediction is weak. Holdout R² is approximately 0.01–0.02, and this weakness has persisted for 11 slates.
  8. Simulation contribution is not independently trustworthy and has been inert for 11 slates.
  9. Calibration is poor on the current slate and is trending worse across the monitoring window.
  10. Ordinary artifact coverage is complete and all structural audits passed.
  11. Two probability lineages have critical unexplained publication transformations.
  12. The integrity-revision audit artifact is missing.

Uncertain

  1. Whether the unexplained probability deltas represent a legitimate but undocumented starter/recency transformation or an actual publication-contract bypass.
  2. Whether recent/season run weighting is beneficial after controlling for slate composition; the current lever tension is mixed.
  3. Whether starter prior weighting should be globally changed; the cross-slate starter bias is not persistent.
  4. Whether widening variance will improve realized score metrics or merely make forecasts less specific; it should remain a shadow experiment.
  5. The extent to which the reported player-level deviations are learnable: only 21 of 55 material deviations were supported by the hitter’s own season contact quality.

Safest single next action

Do not promote any model or probability change. First reconcile the complete pick-side probability lineage for BAL@NYM and MIA@ARI, while independently validating the simulation distribution code in shadow.

That is the safest immediate action because it addresses the only critical contract failure and the strongest persistent operational failure before allowing calibration or variance adjustments to mask a provenance or no-op simulation defect.