Luna deep analysis — slate 20260831

Coverage receipt: COMPLETE · 9/8 sections; 12/12 games; 7/7 causes; 6/6 sources

1. Evidence receipt and gaps

Source manifest

All seven manifest entries were checked against their recorded status:

SourcePathBytesStatusUse
Prediction review/app/output/reports/prediction_review_20260831.json300,127LoadedOfficial outcomes and scoring errors
Forensic deep audit/app/output/reports/PREDICTION_SYSTEM_DEEP_AUDIT_20260831.json195,470LoadedProbability, provenance, starter, and decision contracts
Continuous learning/app/output/learning/continuous_learning_daily_20260831.json1,934,747LoadedStage quality, calibration, lineage, margins, starter comparisons
Detailed report/app/output/reports/detailed_20260831.html284,841LoadedGame and simulation rendering
Ultra-detailed player audit/app/output/reports/20260831_ultra_detailed_player_audit.html4,202,742LoadedPlayer-level prediction trace
Structural audit/app/output/reports/audit_report_20260831.html257,116LoadedInvariant findings
Integrity revision audit/app/output/reports/audit_report_20260831_integrity_revision.htmlNot suppliedMissingPost-remediation proof for immutable revision artifacts

Coverage

All 12 expected games were reviewed:

The data marks learning eligibility as true because all game artifacts passed critical/error structural gates.

Gaps and contradictions

  1. Missing integrity revision audit. The manifest explicitly reports status: "missing" for the immutable-revision audit. This is an evidence gap and an artifact-integrity limitation, although the structural summary reports zero critical and zero error findings.
  2. Per-game audit contradiction. Each game has structural_audit.findings: [] and status: "PASS", but the diagnostic observations state that the structural audit exposed 7 or 8 non-pass findings. The summaries count only info findings—87 across the slate—but the underlying individual findings are not present in this JSON. Their content therefore cannot be independently enumerated or evaluated.
  3. BAL@COL probability-path naming inconsistency. The game-level trace records raw_pick_probability: 0.724282, while the probability-lineage row records raw_model_pick_probability: 0.7585 and governor_input_probability: 0.724282. The lineage is materially more complete; the game-level raw_pick_probability field appears to refer to a post-raw/pre-governor value rather than the true raw model output. This is a trace-labeling defect or schema ambiguity, not an unexplained publication bypass.
  4. Confidence-edge convention. The listed confidence_home_minus_away is a signed home-minus-away edge, not a probability itself. For example, BAL@COL has -0.4985, corresponding to a published away probability of approximately 0.749 and home probability of approximately 0.251. The reports correctly warn that the “0.50” strong edge is not a half-scale distance from 50%.
  5. No source data for actual player box-score detail beyond starter ER/outs and listed adjustment observations. No unprovided player performance should be inferred.

2. Executive diagnosis

Overall result

A correct winner was often accompanied by a poor score prediction. The slate produced 8 winner hits, but only two games were classified as solid_correct_call—CHW@HOU and MIA@WAS. The other six winner hits were either materially wrong scores or, in one case, a correct winner despite substantial starter/run error.

Separate diagnoses

Winner accuracy

Winner accuracy was acceptable on this slate at 66.7%, but not uniformly reliable:

The key process concern is not the final accuracy alone. The historical_ml → variance_applied transition regressed winner accuracy by 2 hits, from 8 to 6, before the enhanced stage recovered those two games. The affected direction changes were:

The later recovery does not erase the intermediate regression.

Calibration quality

Calibration was poor on this 12-game slate:

The bucket behavior is unstable rather than monotonic. The issue register appropriately warns not to infer a stable calibration curve from seven games or from this one slate. The high-raw cluster also deserves attention: it contained 3 games and 2 hits, while the associated empirical hit rates were only 0.5519, 0.5567, and 0.5766 despite calibrated probabilities from 0.661 to 0.7435.

Score accuracy

Score accuracy was the weakest prediction dimension:

The issue register’s score_margin_compression finding is therefore material and high severity. It does not justify mechanically widening every score. The correct response is rolling out-of-sample evaluation of variance width, tail coverage, margin MAE, and total-run MAE.

Starter accuracy

Starter projections were a major contributor to the four losses:

The starter lever tension reports that pitcher-form/prior weighting would have helped 9 games and hurt 1, but this is a small slate and does not establish causal superiority.

Artifact integrity

Structural coverage was complete and the aggregate structural audit reports:

However:

Therefore, artifact integrity is broadly passing at the critical/error gate but not fully clean as a forensic evidence package.

Causal issue register reconciliation

IDSeverityEvidenceAffected gamesActionGuardrail
pipeline_winner_instabilityHighhistorical_ml → variance_applied reduced winner accuracy by 0.1667, or 2 hitsBAL@COL, NYY@LAA; PHI@ARI also changed direction but recoveredIdentify exact boundary flips and preserve pre/post winnersLater recovery does not make the unstable intermediate transition harmless
score_margin_compressionHigh11/12 predicted margins ≤2; 6/12 actual margins ≥4; predicted mean 1.42 vs actual 4.92Listed six direct affected games: DET@MIN, MIL@CHC, NYY@LAA, OAK@TEX, SD@CIN, SF@ATLEvaluate distribution width, tail coverage, and margin MAE on rolling OOS slatesOne blowout does not justify widening every point estimate
single_slate_calibration_instabilityHighWeighted absolute calibration error 0.1661All 12Monitor buckets over registered rolling windowThis slate cannot establish a stable calibration curve
weak_run_model_signalHighRun R² only 0.0275/0.0378/0.0272All 12Investigate features/target through isolated training backtestPositive R² only beats the mean baseline; exact scores remain unreliable
complete_slate_audit_coverageInfo12/12 structurally auditedAll 12Retain per-game receiptCoverage does not erase internal findings
probability_publication_delta_explainedInfo8 material movements, all attributed to score coherence and/or recencyEight listed gamesContinue displaying all probability layersA higher published value is not automatically a governor bypass
simulation_inert_on_slateInfo on this slate, escalated persistent pattern across slates0 candidates and 0 published flipsAll 12Measure simulation contribution and verify executionZero flips is not itself a defect and does not prove no score work occurred

3. Every game

The pipeline score path for all games was:

historical_ml → variance_applied → enhanced_prediction → embedded_savant_run_environment → fundamental_edge_target_review → capped_target → simulation_result → published_final

Only the first two transitions changed winner direction on this slate. Simulation and later stages produced no winner flips.

BAL@COL — hit, correct winner/wrong score

CHW@HOU — hit, solid correct call

DET@MIN — hit, correct winner/wrong score

MIA@WAS — hit, solid correct call

MIL@CHC — miss

NYM@TB — miss

NYY@LAA — miss

OAK@TEX — hit, correct winner/wrong score

PHI@ARI — hit, correct winner/wrong score

SD@CIN — hit, correct winner/wrong score

SEA@BOS — hit, correct winner/wrong score

SF@ATL — miss

Correct-winner/wrong-score games

The six explicit right_winner_wrong_score games were:

CHW@HOU and MIA@WAS were classified as solid correct calls. MIL@CHC, NYM@TB, NYY@LAA, and SF@ATL were winner misses.

4. Miss reconstruction

MIL@CHC

Earliest wrong-direction stage

The prediction was already on the wrong side in the historical model and stayed wrong through publication. The data does not show a later stage moving the winner toward CHC. The earliest identifiable problem is therefore the base winner/run assessment, not a later transition.

Cause classification

NYM@TB

Earliest wrong-direction stage

The historical/base model already selected TB, and no later stage corrected it. The earliest evidence moving toward the wrong result was the high composite/season edge for TB:

The pitcher edge slightly favored NYM at −0.0457, but it did not overcome the broader TB signal.

Cause classification

NYY@LAA

Earliest wrong-direction stage

There was a documented intermediate regression:

The first clear move toward the wrong final result was the enhanced prediction’s recovery toward NYY, after the variance stage had already removed LAA’s direction. This is exactly the pipeline-instability issue: later recovery does not make the intermediate regression harmless.

The base evidence itself strongly favored NYY on season and recent quality:

Cause classification

SF@ATL

Earliest wrong-direction stage

The historical/base direction favored ATL and no later stage corrected it. The earliest wrong-direction evidence was the combination of:

The pitcher edge favored SF at −0.0832, but was insufficiently weighted to reverse the call.

Cause classification

Pipeline-stage quality across all stages

StageWinner accuracyWinner hitsTeam-run MAETotal-run MAEMargin MAE
historical_ml0.666783.12833.99104.7023
variance_applied0.500063.12503.91674.7500
enhanced_prediction0.666783.45834.25004.7500
embedded_savant_run_environment0.666783.41674.33334.6667
fundamental_edge_target_review0.666783.41674.33334.6667
capped_target0.666783.41674.33334.6667
simulation_result0.666783.41674.33334.6667
published_final0.666783.41674.33334.6667

Transitions:

The intermediate winner regression remains a genuine process risk even though enhanced prediction later recovered the final winner count.

5. Cross-game counterfactuals

Lever tension: widen run distributions

Lever: score.variance_multiplier / score.variance_floor

Helps:

Hurts:

Net evidence: 2 helps vs 2 hurts. The pattern supports a bounded shadow experiment, not an unrestricted production increase. It targets Brier score and calibration error, while score-margin compression is also relevant.

Lever tension: pitcher form/prior weighting

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

Helps:

Hurts:

Net evidence: 9 helps vs 1 hurts, but this is based on starter error observations, not a controlled counterfactual. It supports a diagnostic/backtest, not immediate aggressive reweighting.

Lever tension: recent/season run-mean weighting

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

Helps:

Hurts:

Net evidence: 9 helps vs 2 hurts for team-run and total-run metrics, but the reported “helps” do not prove a particular direction of weight change. The weak run-model R² argues for feature/target investigation before tuning this lever.

One-game fixes that must not be made

6. Data and contract defects

Critical findings

The structural audit summary reports 0 critical. No critical contract failure is evidenced in the supplied JSON.

The missing integrity revision audit is an evidence gap, but the data does not classify it as critical.

Error findings

Warning findings

Info findings

Prediction-model defects versus presentation/artifact defects

Prediction-model/process defects

  1. Intermediate winner instability at historical_ml → variance_applied.
  2. Compressed score margins and under-dispersed run distributions.
  3. Very weak run-model holdout signal.
  4. Starter projection errors materially affecting three misses.
  5. Calibration misalignment on the current slate.
  6. Simulation producing no winner-direction interventions over four consecutive slates, requiring execution diagnosis.

Presentation/artifact defects

  1. Missing integrity revision audit artifact.
  2. BAL raw probability field appears inconsistently named or populated between game trace and lineage.
  3. Per-game audit observations claim 7–8 non-pass findings while the embedded findings arrays are empty. This prevents independent audit of those findings.
  4. “PASS” status should not be interpreted as absence of all info-level or non-pass observations; the report itself says complete coverage does not erase findings.

Probability-lineage audit

The complete pick-side lineage contains 12 rows. The recorded layers are:

raw model → calibrated → governor input/output → score coherence and/or recency → published

The data identifies 8 material calibrated-to-published movements, and states that all were explained by named score-coherence and/or recency transformations. No material unexplained delta is marked in the lineage rows.

Examples:

There is no demonstrated governor bypass: every material published movement has a recorded transformation and unexplained_material_delta: false. The BAL field-label discrepancy is a trace consistency issue, not a material unexplained publication delta.

7. Cross-slate persistence

Three prior slates were loaded, producing a four-slate window from 20260828 through 20260831. A single slate is noise, but the escalated findings are explicitly established patterns and must not be treated as single-slate noise.

Raw structural evidence

Margin distribution

Current slate:

Pipeline transitions

The raw transition block shows a current winner-direction regression at historical_ml → variance_applied:

This finding has a two-slate consecutive streak and is present now.

Calibration monitoring

Current WACE is 0.1661. Bucket errors alternate between underconfidence and overconfidence, confirming instability rather than a simple universal overconfidence bias.

High-raw cluster

Current high-raw cluster:

The finding streak reports high_raw_probability_cluster at 0 consecutive slates, present current slate false. It is therefore not a current persistent finding.

Simulation control

Current slate:

Across slates, simulation inertness is established, not noise.

Finding streaks

FindingConsecutive slatesPresent now?Interpretation
high_raw_probability_cluster0NoNot persistent/current
pipeline_winner_instability2YesRecurrent intermediate winner-direction regression
score_margin_compression1YesCurrent issue, not yet a multi-slate streak
simulation_inert_on_slate4YesEstablished persistent failure
single_slate_calibration_instability1Yes; incidence 3 in windowCurrent calibration problem, but not a four-slate persistence pattern
weak_run_model_signal4YesEstablished persistent weak run-model signal

Escalated findings

simulation_inert_on_slate

weak_run_model_signal

Metric trends

Top-confidence-pick loss streak

There is no current top-confidence-pick loss streak.

8. Adjustments for today

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

1. Diagnose and repair simulation execution — highest persistence

2. Isolate the weak run model before tuning training parameters

3. Run a bounded shadow increase to score-distribution width

4. Do not move calibration caps today; use shadow monitoring

5. Starter model diagnostic, not an immediate global reweight

9. Bottom line

Known

Uncertain

Single safest next action

Perform the four-slate code-and-lineage diagnostic for simulation.allow_score and the weak run-model execution path before changing model weights. It addresses two established persistent findings, has a reversible shadow-gate path, and avoids making an unsupported one-slate calibration or score-distribution adjustment while also checking the most important artifact/process risk.