Deep slate report — 20260929

1. Evidence receipt and gaps

Sources received

All seven manifest entries were checked:

Source	Status	Purpose	Receipt finding
prediction review — /app/output/reports/prediction_review_20260929.json	loaded	Official outcomes and scoring errors	Present, 102,938 bytes
forensic deep audit — /app/output/reports/PREDICTION_SYSTEM_DEEP_AUDIT_20260929.json	loaded	Probability, provenance, starter, and decision contracts	Present, 75,641 bytes
continuous learning — /app/output/learning/continuous_learning_daily_20260929.json	loaded	Stage quality, calibration, lineage, score distribution, starter comparisons	Present, 681,575 bytes
detailed report — /app/output/reports/detailed_20260929.html	loaded	Full game and simulation rendering	Present, 96,839 bytes
ultra detailed player audit — /app/output/reports/20260929_ultra_detailed_player_audit.html	loaded	Player-level prediction trace	Present, 8,510,517 bytes
structural audit — /app/output/reports/audit_report_20260929.html	loaded	Invariant findings	Present, 518,619 bytes
integrity revision audit — /app/output/reports/audit_report_20260929_integrity_revision.html	missing	Post-Luna remediation proof for immutable revision artifacts	Missing; byte count is null

Slate coverage

All four listed games were reviewed:

1. BOS@NYY
2. CHC@SD
3. CHW@HOU
4. PHI@ATL

Coverage is internally complete for the game-level prediction audit:

• games_reviewed: 4
• hits: 4
• misses: 0
• expected_games: 4
• canonical_artifacts_loaded: 4
• continuous_learning_games: 4
• forensic_deep_audit_games: 4
• games_with_starter_comparisons: 4
• structural_audits_completed: 4
• all_games_structurally_audited: true

Gaps and contradictions

1. Missing integrity revision audit. The only missing manifest artifact is the immutable revision/remediation proof. This is a documentation and artifact-verification gap, not evidence that the predictions themselves were altered.
2. No market cross-check. market_crosscheck.available is false because no source-labelled moneyline covers were available.
3. No actual losses. The loss-specific fields are null or empty:

• misses: 0
• loss_primary_drivers: {}
• mean_team_run_error_on_losses: null
• mean_total_run_bias_on_losses: null

Consequently, no game can be reconstructed as a winner miss.

1. Calibration terminology is inconsistent. The bucket shows average confidence 0.5208, observed hit rate 1.0, and error -0.4792, correctly indicating underconfidence under the stated definition (“average published confidence minus observed hit rate”). Elsewhere, the self-assessment calls the slate “confidence_overconfident” and recommends lowering caps. That recommendation is not supported by this four-game bucket alone.
2. The top-pick streak data contains an invalid date. top_confidence_pick.per_slate includes one entry with date: "" and result unknown. The aggregate loss_streak: 0 remains supported by the dated current entry, but the sequence is not clean enough to treat every historical position as fully auditable.
3. The model reports 100% winner accuracy while run-model validation is weak. This is not a contradiction: all four winners were correct, but exact scores and margins were poor.
4. “Forecast-file integrity: clean” appears for every game, while the manifest still lacks the separate integrity revision audit. These refer to different checks. The game artifacts passed their structural audits; the separate revision-proof artifact remains unavailable.

---

2. Executive diagnosis

Winner accuracy

Excellent on this slate, but weak evidence of broad model quality.

• Winner accuracy: 4/4 = 1.000
• Every published winner matched the official winner.
• The continuous-learning summary also reports accuracy: 1.0.
• The model’s historical walk-forward accuracy is 0.570149, and its holdout winner accuracy is 0.600414.
• The current slate therefore beat both historical reference levels, but the sample is only four games.

The correct winner was sometimes produced by a very small edge:

• CHC@SD: published home edge 0.0158
• PHI@ATL: published home edge 0.0108

Those should not be treated as strong directional forecasts.

Calibration quality

Poorly measured on this slate and not suitable for a permanent recalibration from four observations.

The only reported confidence bucket is:

• Average published confidence: 0.5208
• Observed hit rate: 1.000
• Calibration error: -0.4792
• Weighted absolute calibration error: 0.4792
• Four games, four hits

By the supplied definition, the negative signed error indicates underconfidence, not overconfidence. The four correct calls do not establish that the confidence system is generally too conservative; the observed hit rate is an unusually noisy four-game realization.

The high-severity single_slate_calibration_instability finding is therefore appropriate: the bucket is badly aligned on this slate, but the registered guardrail correctly says not to change calibration weights from a single four-game sample.

Score accuracy

Materially weak. Correct winners did not imply accurate scores.

Published versus actual scores:

• Predicted margin MAE: 4.25
• Signed margin bias: -3.25, showing systematic underestimation of winning margins
• All four predicted margins were ≤2
• Two actual margins were ≥8
• Mean predicted margin: 1.25
• Mean actual margin: 5.50
• Actual margins: 9, 8, 3, 2
• Predicted margins: 2, 1, 1, 1

The score_margin_compression issue is correctly rated high. It is a material run-model risk, not a critical contract failure. The guardrail is also correct: one or two blowouts do not justify globally widening every score prediction.

The validated run signal is weak:

• Home-runs R²: 0.002143
• Away-runs R²: 0.030936
• Total-runs R²: 0.002415
• Training rows: 2,415

The weak_run_model_signal finding is therefore a serious modeling concern, but its proposed remedy is feature/target and training work, not a small nightly parameter adjustment.

Starter accuracy

Mixed but not a slate-wide starter failure.

Across the games:

• BOS@NYY: both starters were reasonably projected; worst ER error 1
• CHC@SD: Michael King was materially better than projected, with projected ER 3 versus actual 0; ER error -3
• CHW@HOU: starter lines were incomplete for Hagen Smith in the settled per-player feed, although the game-level starter comparison contains projected and actual values
• PHI@ATL: both starter ER projections were imperfect, with worst ER error 2; no clear starter-driven winner failure

Cross-game evidence does not justify changing pitcher weights:

• player.pitcher.v4 form weights / player.pitcher_prior_games helped CHC@SD but risked harming BOS@NYY and CHW@HOU.
• The lever tension is 1 helped vs 2 hurt.
• The cross-slate starter ERA bias is nonpersistent: same sign on 6 of 11 slates, mean 0.291, z-score 0.43, status hold.

Artifact integrity

Game-level structural integrity passed; one separate revision artifact is missing.

• All four game structural audits: PASS
• Structural summary: critical: 0, error: 0, warning: 0, info: 31
• All four game diagnostic observations say the forecast file was clean and not responsible for an incorrect pick.
• Probability lineage records no unexplained material delta.
• The simulation produced no winner flips, but this is a functional/diagnostic concern rather than a file-integrity defect.
• The missing integrity_revision_audit prevents a complete immutable-revision receipt.

Causal issue register reconciliation

pipeline_winner_instability — high

• Evidence: historical_ml → variance_applied changed winner accuracy by -0.5000, losing two hits.
• Affected games: CHC@SD and PHI@ATL.
• Exact direction changes:
• CHC@SD: SD → TIE
• PHI@ATL: ATL → TIE
• Action: preserve before-stage and after-stage picks in the trace and list exact flips.
• Guardrail: later recovery does not make the unstable intermediate stage harmless.

This is a real process/model risk even though enhanced prediction later restored both winners.

score_margin_compression — high

• Evidence: all four predicted margins were at most two runs; two actual margins were at least four runs; mean predicted margin 1.25 versus actual 5.50.
• Affected games: BOS@NYY and CHC@SD are the clearest examples.
• Action: evaluate tail coverage and margin MAE over several recent out-of-sample slates before widening the spread.
• Guardrail: do not widen every score prediction because of one blowout.

single_slate_calibration_instability — high

• Evidence: weighted absolute calibration error 0.4792 across four games.
• Affected games: all four.
• Action: monitor the same buckets over the rolling window.
• Guardrail: do not make a permanent calibration change from one slate.

weak_run_model_signal — high

• Evidence: home/away/total R² of 0.0021/0.0309/0.0024.
• Affected games: all four.
• Action: improve features or target and run an isolated backtest.
• Guardrail: low positive R² does not make exact scores reliable.

complete_slate_audit_coverage — info

• Evidence: all four games structurally audited.
• Action: retain the per-game audit receipt.
• Guardrail: complete coverage does not negate findings inside the games.

probability_publication_delta_explained — info

• Evidence: one material movement for CHW@HOU, explained by score-coherence and/or recency transformations.
• Action: continue displaying every probability layer and named transformation.
• Guardrail: a published probability above or below a calibrated probability is not a bypass when the transformation is recorded.

simulation_inert_on_slate — info on this slate, escalated across slates

• Evidence: zero flip candidates and zero published flips on the current slate.
• Current action: measure simulation contribution over rolling slates.
• Persistent cross-slate status: escalated after 11 consecutive slates; this is now an established diagnostic/code failure pattern, not merely a one-slate observation.

---

3. Every game

BOS@NYY

Prediction and result

• Predicted: NYY 4, BOS 2; winner NYY
• Actual: NYY 9, BOS 0; winner NYY
• Winner: correct
• Classification: right_winner_wrong_score
• Published home probability: 0.540
• System confidence edge: +0.08 home
• Conviction: lean
• Brier score: 0.2116
• Team-run absolute error: 7
• Total-run absolute error: 3 — predicted total 6, actual total 9
• Signed total bias: -3

This is explicitly a correct-winner/wrong-score game.

Pregame edges and controls

• Season home edge: +0.3398
• Recent home edge: +0.2300
• Pitcher edge: -0.0121
• Composite home edge: +0.5778
• Season quality edge: +0.4366
• Recent quality edge: +0.16
• Season margin adjustment: 0.0
• Pitcher adjustment applied:
• BOS runs versus NYY pitcher factor: 1.0034
• NYY runs versus BOS pitcher factor: 0.9913
• Starter-experience shrinkage: not applied; minimum starter-games control was 26, shrink factor 1.0
• Score-volatility action: no_change
• Confidence path:
• Raw pick probability: 0.5175
• Governor input: 0.5175
• Governed/calibrated probability: 0.5554
• Published: 0.540
• Recency guardrail applied
• No unexplained material delta

Starter projection versus actual

Starter	Projected	Actual	Error
Payton Tolle, BOS	2 ER, 17 outs, 3.18 ERA	1 ER, 14 outs, 1.93 ERA	ER -1, outs -3, ERA -1.25
Cam Schlittler, NYY	0 ER, 18 outs, 0.00 ERA	0 ER, 19 outs, 0.00 ERA	ER 0, outs +1, ERA 0

Starter identity matching passed.

Score path and player adjustments

The primary score miss was offensive, not directional:

• NYY lineup scored 2 above projection in the matchup breakdown, while the total NYY offense was reported as under-projected by 5 in the contributing-factor summary.
• BOS lineup scored 4 below projection.
• BOS bullpen allowed 1 ER; NYY bullpen allowed 0.
• Ben Rice was the largest positive player deviation:
• Actual: 4-for-5, 2 HR, 6 RBI
• Projected: 1 H, 1 HR
• Run impact: +2.32
• Season xwOBA: 0.378, classified elite
• George Lombard Jr.: +0.94 run impact
• Austin Wells: +0.94
• Willson Contreras: -1.40
• Roman Anthony: -0.80

Only seven of 27 material player deviations across the slate were supported by the hitter’s own contact-quality profile; Rice was one of the supported cases.

Audit and causes

• Structural audit: PASS
• Findings: none
• Risk score: 0
• Artifact defects: none; seven informational notes
• Primary cause: offensive/run-environment miss, especially NYY offense exceeding projection
• Secondary cause: compressed predicted margin
• Starter cause: not material; the starter projection was broadly accurate
• Important guardrail: the game was a correct directional call, not a score-quality success

---

CHC@SD

Prediction and result

• Predicted: SD 5, CHC 4; winner SD
• Actual: SD 8, CHC 0; winner SD
• Winner: correct
• Classification: toss_up_went_our_way
• Published home probability: 0.507896, rounded diagnostic value 0.508
• System confidence edge: +0.0158 home, reported as approximately 0.02
• Conviction: toss_up
• Brier score: 0.242166
• Team-run absolute error: 7
• Total-run absolute error: 1 — predicted total 9, actual total 8
• Signed total bias: +1

This is also a correct-winner/wrong-score game.

Pregame edges and controls

• Season home edge: -0.2355
• Recent home edge: +0.1580
• Pitcher edge: -0.0325
• Composite home edge: +0.0185
• Season quality edge: -0.2355
• Recent quality edge: -0.6358
• Pitcher adjustment applied:
• CHC runs factor: 0.9899
• SD runs factor: 0.9574
• Starter-experience shrinkage: not applied; minimum starter-games control 22, shrink factor 1.0
• Score-volatility action: no_change
• Confidence path:
• Raw pick probability: 0.530
• Governor input: 0.530
• Governed/calibrated: 0.5109
• Published: 0.507896
• Recency guardrail applied
• No unexplained material delta

Starter projection versus actual

Starter	Projected	Actual	Error
Matthew Boyd, CHC	5 ER, 18 outs, 7.50 ERA	4 ER, 12 outs, 9.00 ERA	ER -1, outs -6, ERA +1.50
Michael King, SD	3 ER, 17 outs, 4.76 ERA	0 ER, 21 outs, 0.00 ERA	ER -3, outs +4, ERA -4.76

King’s start was a major starter projection miss. The forensic synthesis classifies it as variance: no credible venue, opponent-strength, or experience signal explained the performance.

Score path and player adjustments

• CHC offense was over-projected by 4.
• SD offense was under-projected by 3.
• CHC lineup run delta: -5
• SD lineup run delta: +3
• Bullpen ER: CHC 0, SD 0
• Fernando Tatis Jr.: +1.73, actual 2-for-3, 1 HR versus projected 1 H
• Xander Bogaerts: +1.54
• Manny Machado: +0.78
• Pete Crow-Armstrong: -1.40
• Alex Bregman: -0.78

The game illustrates why a near-even call can still have a poor score: the model’s directional edge was only approximately 1.6 percentage points, and SD’s 8–0 result was driven by both the King outlier and offensive deviations.

Audit and causes

• Structural audit: PASS
• Findings: none
• Risk score: 0
• Artifact defects: none; eight informational notes
• Primary cause: starter variance, particularly King’s 0 ER over 21 outs versus 3 projected ER
• Secondary cause: offensive deviations and compressed margin
• Pipeline finding: variance application temporarily moved the winner from SD to TIE; enhanced prediction recovered SD. That recovery does not erase the intermediate regression.

---

CHW@HOU

Prediction and result

• Predicted: CHW 6, HOU 5; winner CHW
• Actual: CHW 6, HOU 3; winner CHW
• Winner: correct
• Classification: solid_correct_call
• Published away probability: 0.530; home probability 0.470
• System confidence edge: -0.06 home, equivalent to a +0.06 away edge
• Conviction: lean
• Brier score: 0.2209
• Team-run absolute error: 2
• Total-run absolute error: 2 — predicted total 11, actual total 9
• Signed total bias: +2

This was the best score-quality game of the slate, though it still missed the exact total by two runs.

Pregame edges and controls

• Season home edge: -0.4495
• Recent home edge: -0.1540
• Pitcher edge: -0.0020
• Composite home edge: -0.4701
• Season quality edge: -0.4495
• Recent quality edge: -0.571
• Pitcher adjustment was not applied:
• Away factor: 1.000
• Home factor: 0.998
• Diagnostic flag: large_savant_run_environment_swing
• Starter-experience shrinkage: not applied; minimum starter-games control 18, shrink factor 1.0
• Score-volatility action: no_change
• Confidence path:
• Raw pick probability: 0.530
• Governor input: 0.530
• Governed/calibrated: 0.5513
• Score-coherence probability: 0.551288
• Published: 0.530
• Recency guardrail applied
• The material movement was explained; no bypass occurred

Starter projection versus actual

Starter	Projected	Actual	Error
Hagen Smith, CHW	1 ER, 15 outs, 1.80 ERA	0 ER, 9 outs, 0.00 ERA	ER -1, outs -6, ERA -1.80
AJ Blubaugh, HOU	0 ER, 15 outs, 0.00 ERA	0 ER, 2 outs, 0.00 ERA	ER 0, outs -13, ERA 0

The settled per-player feed lacks a pitching line for Hagen Smith, despite the game-level starter comparison containing projected and actual values. The report explicitly marks this as starter_line_missing for the player-level matchup breakdown. That is a data-coverage limitation, not a structural audit error.

The HOU starter’s short outing was also materially different in workload, although not in ER.

Score path and player adjustments

• CHW lineup run delta: +1
• HOU lineup run delta: -3
• HOU bullpen allowed 1 ER
• Colson Montgomery: +1.40
• Yordan Alvarez: -1.26
• Jose Altuve: -1.11; weak season contact-quality tier, xwOBA 0.271
• Munetaka Murakami: -0.93
• Jeremy Peña: -0.93

This was the game most supportive of the existing recent/season run-mean weighting: the team-run error was only two, and the away-side directional fundamentals were strongly favorable.

Audit and causes

• Structural audit: PASS
• Findings: none
• Risk score: 0
• Artifact defects: none; eight informational notes
• Primary cause of residual score error: modest offense/bullpen and workload variance
• Secondary cause: run-environment swing, reflected in the diagnostic flag
• No evidence supports changing starter weights based on this game

---

PHI@ATL

Prediction and result

• Predicted: ATL 3, PHI 2; winner ATL
• Actual: ATL 5, PHI 3; winner ATL
• Winner: correct
• Classification: toss_up_went_our_way
• Published home probability: 0.505394, rounded diagnostic value 0.505
• System confidence edge: +0.0108 home, approximately 0.01
• Conviction: toss_up
• Brier score: 0.244635
• Team-run absolute error: 3
• Total-run absolute error: 3 — predicted total 5, actual total 8
• Signed total bias: -3

This is a correct-winner/wrong-score game.

Pregame edges and controls

• Season home edge: +0.5590
• Recent home edge: +0.4060
• Pitcher edge: +0.0494
• Composite home edge: +0.6976
• Season quality edge: +0.679
• Recent quality edge: +0.336
• Season margin adjustment: +0.2348
• Pitcher adjustment applied:
• PHI runs factor: 0.9507
• ATL runs factor: 1.0001
• Starter-experience shrinkage: not applied; minimum starter-games control 27, shrink factor 1.0
• Score-volatility action: no_change
• Confidence path:
• Raw pick probability: 0.5264
• Governor input: 0.5264
• Governed/calibrated: 0.5075
• Published: 0.505394
• Recency guardrail applied
• No unexplained material delta

Starter projection versus actual

Starter	Projected	Actual	Error
Jesús Luzardo, PHI	3 ER, 20 outs, 4.05 ERA	1 ER, 15 outs, 1.80 ERA	ER -2, outs -5, ERA -2.25
Chris Sale, ATL	2 ER, 19 outs, 2.84 ERA	3 ER, 19 outs, 4.26 ERA	ER +1, outs 0, ERA +1.42

The starter projection was imperfect on both sides but did not clearly drive the result.

Score path and player adjustments

• PHI lineup run delta: -1
• ATL lineup run delta: +1
• PHI bullpen allowed 3 ER, including Jhoan Duran:
• Projected 0 ER, actual 3 ER
• Actual 2 H, 6 outs versus projected 4 outs
• ATL bullpen allowed 0 ER
• Michael Harris II: +0.94, supported by elite season contact quality, xwOBA 0.342
• Ozzie Albies: +0.95
• Matt Olson: -1.40
• Kyle Schwarber: -0.93
• The total was under-projected by three runs.

Audit and causes

• Structural audit: PASS
• Findings: none
• Risk score: 0
• Artifact defects: none; eight informational notes
• Primary cause of score error: bullpen/run realization, especially PHI’s three bullpen ER
• Secondary cause: compressed total and margin
• Winner direction was also temporarily degraded by the variance stage, which changed ATL to TIE before enhanced prediction restored ATL.

---

4. Miss reconstruction

Game-level losses

There were no published winner losses. All four games were winner hits, so there is no loss to reconstruct in the usual sense.

There were, however, score and intermediate-stage failures:

• BOS@NYY and CHC@SD had seven runs of team-run absolute error.
• PHI@ATL had a three-run total error.
• CHW@HOU had a two-run total error.
• historical_ml → variance_applied moved two correct eventual winners toward a tie:
• CHC@SD: SD → TIE
• PHI@ATL: ATL → TIE

A later recovery does not erase this intermediate winner-direction regression.

Earliest wrong-direction stage

The earliest measured regression was:

• Stage: historical_ml → variance_applied
• Winner accuracy: 1.0 → 0.5
• Winner hits: 4 → 2
• Margin MAE: 4.562 → 4.750
• Team-run MAE: 2.281 → 2.375
• Regressed games: CHC@SD and PHI@ATL

The next stage, variance_applied → enhanced_prediction, recovered both:

• CHC@SD: TIE → SD
• PHI@ATL: TIE → ATL
• Winner accuracy returned to 1.0
• Margin MAE improved by 0.25, from 4.75 to 4.50
• Team-run MAE worsened by 0.125, from 2.375 to 2.5

Therefore the proper diagnosis is intermediate winner-direction instability, not “no problem because the final picks were correct.”

A home-away confidence edge of 0.12, if encountered in another slate, would be moderate—not a near coin flip. On this slate, the largest edge was only 0.08; CHC@SD and PHI@ATL were approximately 0.02 and 0.01.

Cause classification

Category	Evidence on this slate	Assessment
Bad input	One settled player-level starter line missing for Hagen Smith; no structural errors	Limited data gap, not a demonstrated winner cause
Bad weighting	Variance stage regressed two winner directions; recent/season weighting helped two large score misses but could disturb CHW@HOU	Material process risk at variance stage; run-weight change not yet justified
Calibration	Four-game calibration error 0.4792; bucket is underconfident by stated definition	Unstable measurement, not enough evidence for permanent recalibration
Starter error	King was 3 ER better than projected; PHI/ATL starter errors; BOS starters broadly accurate	Some individual variance; pitcher lever net tension is unfavorable
Offense/bullpen miss	NYY offense exceeded projection; CHC offense underperformed and SD offense exceeded projection; PHI bullpen allowed 3 ER	Major source of score error
Run-environment miss	Predicted margins were compressed; total errors were -3,+1,+2,-3	Material score-model issue
Irreducible variance	King’s outlier start and player-level deviations unsupported by stable profiles	Important explanation; do not tune directly to isolated performances

Pipeline-stage quality across all stages

The stored stages compare directly to official final scores:

Stage	Winner accuracy	Team-run MAE	Total-run MAE	Margin MAE	Interpretation
historical_ml	1.000	2.281	0.6485	4.562	Baseline
variance_applied	0.500	2.375	0.750	4.750	Winner regression; score regression
enhanced_prediction	1.000	2.500	2.000	4.500	Recovers winners, but total/team score metrics worsen versus baseline
embedded_savant_run_environment	1.000	2.375	2.250	4.250	Margin improves; total MAE worsens versus enhanced
fundamental_edge_target_review	1.000	2.375	2.250	4.250	No additional change
capped_target	1.000	2.375	2.250	4.250	No additional change
simulation_result	1.000	2.375	2.250	4.250	No winner intervention
published_final	1.000	2.375	2.250	4.250	No change from simulation stage

Transition audit:

• historical_ml → variance_applied: regression in CHC@SD and PHI@ATL; no improvements.
• variance_applied → enhanced_prediction: both winner directions recovered; no winner regressions.
• enhanced_prediction → embedded_savant_run_environment: margin MAE improved by 0.25; no winner change.
• All later transitions were inert for winner direction and mostly inert for scores.
• simulation_result → published_final: no change.

The later recovery is useful, but it does not cancel the unstable variance-stage behavior.

---

5. Cross-game counterfactuals

Lever tension: score variance multiplier / floor

• Helps: none recorded in the explicit lever-tension table.
• Hurts: none recorded.
• Current slate evidence: all four predicted margins were too narrow; BOS@NYY and CHC@SD were especially compressed by seven runs each.
• Net conclusion: The slate supports investigating a bounded widening, but not an unrestricted production change. The explicit 0 helped vs 0 hurt entry means the counterfactual has not been demonstrated game by game.
• Guardrail: the score_margin_compression finding specifically says to check several recent out-of-sample slates first.

Lever tension: pitcher form weights / prior games

• Helps: CHC@SD, where Michael King allowed 0 ER against 3 projected.
• Hurts: BOS@NYY, where the starter projection was accurate within one ER; CHW@HOU, where starter ER projections were also close within one ER.
• Net evidence: 1 helped vs 2 hurt; do not reweight this lever from the current slate.
• Reason: King’s result was explicitly classified as an outlier start without a credible explanatory signal.

Lever tension: recent/season run-mean weighting

• Helps: BOS@NYY and CHC@SD, each with team-run error 7.
• Hurts: CHW@HOU, where team-run error was only 2; more aggressive reweighting could disturb an already good call.
• Net evidence: 2 helped vs 1 hurt, but the evidence is about score realization, not winner accuracy. A broad weighting change is not yet justified.

Tempting one-game fixes that must not be made

1. Do not widen all score distributions solely because BOS@NYY and CHC@SD were blowouts.
2. Do not increase pitcher-form weighting solely because Michael King outperformed his projection.
3. Do not lower all confidence caps solely because the current four-game bucket has a 0.4792 absolute calibration error; under the stated definition the bucket is underconfident.
4. Do not treat all hitter deviations as learnable. Only 7/27 material deviations were supported by the hitter’s own season contact quality.
5. Do not infer that simulation is harmless because the final winners were unchanged. It has been inert for 11 consecutive slates.
6. Do not tune the near-zero run model with a nightly variance knob alone. The registered diagnosis calls for feature/target work and an isolated retraining backtest.
7. Do not change the pitcher or run-mean levers without preserving the counterfactual effects on BOS@NYY, CHW@HOU, and other unaffected games.

---

6. Data and contract defects

Critical findings

There are no game-level critical findings:

• Structural audits: critical: 0
• Structural audit summary: critical: 0

However, the missing integrity revision audit is a receipt gap that prevents complete artifact-proof verification. It is not classified by the supplied audit as a critical prediction defect.

Error findings

There are no game-level error findings:

• Every game structural audit has error: 0.
• Every game status is PASS.

Warning findings

There are no warning findings:

• Every game has warning: 0.
• Slate summary has warning: 0.

Informational findings

There are 31 informational findings in the structural summary. Per-game counts are:

• BOS@NYY: 7
• CHC@SD: 8
• CHW@HOU: 8
• PHI@ATL: 8

These include the informational notes associated with clean artifact status, recorded transformations, and diagnostics. Informational status does not mean the model was accurate; it means the contract/audit did not identify a structural violation.

Prediction-model defects versus presentation/artifact bugs

Prediction/model defects or risks

• Intermediate variance-stage winner regression.
• Compressed score margins.
• Near-zero holdout run-model signal.
• Simulation module functionally inert across 11 slates.
• Player-level projections showed weak separation:
• H spread ratio 0.442
• SO spread ratio 0.438
• H correlation 0.129
• SO correlation 0.055
• RBI correlation 0.030
• Slate totals were over-projected for several player metrics:
• AB +10.1%
• BB +31.6%
• SO +17.5%
• RBI +18.5%
• HR +28.6%
• 2B +50.0%

Data/artifact/presentation issues

• Missing immutable revision audit file.
• Missing settled per-player pitching line for Hagen Smith in CHW@HOU.
• No source-labelled moneyline cover for market comparison.
• Invalid blank date in one historical top-pick entry.
• The narrative “confidence_overconfident” label conflicts with the supplied calibration definition and current bucket result.
• No evidence of a file corruption or published-score overwrite: game-level audits passed and all probability movements were explained.

Probability lineage audit

The complete pick-side lineage is available for all four games:

BOS@NYY

raw_model_pick_probability 0.5973 → governor input 0.5175 → calibrated/governor output 0.5554 → score coherence 0.555379 → published 0.540

Transformation: score coherence followed by matchup-recency guardrail. No unexplained delta.

CHC@SD

raw 0.5143 → governor input 0.5300 → calibrated/governed 0.5109 → score coherence 0.510929 → published 0.507896

Transformation: published value matches the calibrated/governed value. No unexplained delta.

CHW@HOU

raw 0.5898 → governor input 0.5300 → calibrated/governed 0.5513 → score coherence 0.551288 → published 0.530

Transformation: score coherence followed by matchup-recency guardrail. The material movement is recorded and explained; it is not a governor bypass.

PHI@ATL

raw 0.5079 → governor input 0.5264 → calibrated/governed 0.5075 → score coherence 0.507466 → published 0.505394

Transformation: published probability matches the calibrated/governed value. No unexplained delta.

Conclusion: no material lineage bypass is demonstrated. A bypass should be called only if a material movement lacks a recorded transformation; that condition is not present here.

---

7. Cross-slate persistence

A single slate is noisy, but several findings have multi-slate evidence.

Finding streaks

Finding	Consecutive-slate count	Present now?	Interpretation
high_raw_probability_cluster	0	No	Historical incidence in 2 of 11 window slates, but not current
pipeline_winner_instability	1	Yes	Current slate shows the issue; incidence 6 of 11, but only a one-slate current streak
score_margin_compression	1	Yes	Current slate shows it; incidence 3 of 11
simulation_inert_on_slate	11	Yes	Established persistent failure, not single-slate noise
single_slate_calibration_instability	1	Yes	Current slate shows it; incidence 7 of 11, but current streak is only one
weak_run_model_signal	1	Yes	Current slate shows it; incidence 10 of 11, but the registered streak field is one

The escalated_findings entry must be treated as established:

Simulation inertness

• Streak: 11 consecutive slates
• Incidence: 11/11
• Target lever recorded: simulation.allow_score
• Target metrics: team_run_mae, total_run_mae
• Parameter actionability: false
• Required direction: confirm the simulation is running rather than silently no-oping
• Metric risk: simulation is not contributing independent score or winner-distribution information

This is a code/diagnostic failure, not a reason to force a parameter move.

Raw structural blocks

Margin distribution

Current slate:

• Actual margins ≥4: 2
• Actual margins: 2, 3, 8, 9
• Predicted margins ≤2: 4
• Predicted margins: 1, 1, 1, 2
• BOS@NYY signed margin error: -7
• CHC@SD signed margin error: -7
• CHW@HOU signed margin error: -2
• PHI@ATL signed margin error: -1

The trend block labels margin bias regressing, with mean -1.9532.

Pipeline transitions

The current slate confirms an intermediate winner-direction regression:

• Historical ML to variance: -2 winner hits
• Variance to enhanced: +2 winner hits
• Later stages: no winner-direction changes

Calibration monitoring

• Current bucket: 0.50–0.55
• Average confidence: 0.5208
• Observed hit rate: 1.0
• Signed error: -0.4792
• Weighted absolute error: 0.4792

The trend labels weighted calibration error regressing, with mean 0.2201. This makes calibration a monitoring priority, but the current four-game direction is underconfidence by the stated definition.

High-raw cluster

• Threshold: 0.65
• Current games: 0
• Current hits: 0
• Current cluster rows: none

There is no current high-raw-confidence cluster.

Simulation control

• Observed games: 4
• Flip candidates: 0
• Published flips: 0
• Status: diagnostic-only without independent support

The 11-slate streak shows that this is not an isolated no-op.

Metric trend directions

Each supplied metric trend is explicit:

• Margin bias: regressing; mean -1.9532
• Mean team-run error on losses: improving; mean 5.308

There are no current losses, so this trend is inherited from prior slates.

• Weighted calibration error: regressing; mean 0.2201
• Winner accuracy: improving; mean 0.6062

The winner-accuracy improvement must not be used to dismiss score-model and pipeline risks.

Top-confidence-pick loss streak

• Reported loss streak: 0
• Current dated top pick, 20260929: hit
• Historical sequence includes misses on 20260919, 20260921, 20260922, and 20260926.
• One historical row has a blank date and unknown, which is a data-quality defect in the streak receipt.
• There is no current consecutive loss streak.

---

8. Adjustments for today

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

1. Highest priority: diagnose and repair simulation inertness

• Pattern: simulation_inert_on_slate
• Persistence: 11 consecutive slates; strongest evidence
• Registered parameter/target: simulation.allow_score
• Action type: code/diagnostic, not a parameter move
• Direction: verify enabled execution and remove silent no-op behavior
• Concrete action: separate target-realization search from win-probability estimation. Use an independent, unsteered batch of approximately 200–500 draws, or an equivalent closed-form distribution, without early exit. Register the batch size.
• Expected metrics: lower team_run_mae and/or total_run_mae; restore measurable simulation contribution and winner-distribution diagnostics.
• Helps: potentially all games, especially games with large score-tail uncertainty such as BOS@NYY and CHC@SD.
• Hurts: no direct game is identified in the current lever-tension record, but a poorly implemented simulation could introduce noise across the entire slate.
• Rollback monitor: revert if independent simulation worsens rolling team-run MAE, total-run MAE, or winner Brier without improving tail coverage over the shadow evaluation window.
• Reason not to force a parameter change: the issue is registered as param_actionable: false; target lever is a code control, not a demonstrated tuning knob.

2. Operator-level run-model repair

• Pattern: weak run-model signal
• Persistence: incidence 10/11 slates, although current consecutive streak field is 1
• Registered targets: training features and target contract; no bounded nightly lever is adequate
• Action: add and validate market no-vig probability where source-labelled, park/weather environment, bullpen rest/fatigue, and platoon-adjusted lineup quality; retrain in an isolated same-fingerprint backtest and increment the model schema version if feature semantics change.
• Expected metrics: improve home/away/total holdout R² and reduce team_run_mae and total_run_mae.
• Helps: all games in future slates, particularly large score misses such as BOS@NYY and CHC@SD.
• Hurts: possible regression in CHW@HOU-style games if added features overfit; require shadow comparison and rollback.
• Reason not to use a small parameter tweak: the supplied self-assessment explicitly says nightly variance/confidence tuning cannot fix near-zero feature signal.

3. Bounded shadow experiment for margin compression

• Pattern: score_margin_compression
• Current evidence: all four predicted margins were ≤2; BOS@NYY and CHC@SD missed actual margins by 7 each.
• Exact registered parameter: score.variance_multiplier

Use score.variance_floor only as a secondary shadow arm, not simultaneously.

• Direction: increase modestly.
• Magnitude: test a small +5% multiplier increase in shadow mode; do not promote automatically.
• Expected metrics: improve margin MAE, tail coverage, and total-run calibration; monitor team-run MAE and Brier score for deterioration.
• Helps: BOS@NYY and CHC@SD most directly; potentially PHI@ATL’s three-run total miss.
• Hurts: no explicit current-game hurt entry is recorded, but CHW@HOU already had team-run error only 2, so widening may make that class of game less precise.
• Gate: require improvement over multiple recent out-of-sample slates, not just this slate.
• Rollback: revert if total-run MAE or team-run MAE worsens materially without improving tail coverage.

4. Calibration caps: monitor, do not immediately promote

• Pattern: current calibration instability; weighted error trend is regressing.
• Registered parameters: confidence.base_cap, confidence.moderate_cap, confidence.empirical_max_weight
• Direction suggested by self-assessment: lower caps / reduce confidence concentration.
• Current recommendation: run a shadow calibration experiment only; do not make a production change from this slate.
• Why: the current four-game bucket is underconfident under the stated definition, despite the self-assessment label “overconfident.” The current slate has no high-conviction games and all four winners hit.
• Expected metrics: lower rolling Brier score and weighted calibration error if the persistent trend is genuine.
• Helps: future high-confidence clusters.
• Hurts: may unnecessarily flatten genuinely useful edges such as the 0.08 BOS@NYY lean or the strong CHW@HOU away direction.
• Rollback: use rolling calibration and Brier gates, not this four-game result.

5. Do not adjust pitcher-form weights today

• Lever: player.pitcher.v4 form weights / player.pitcher_prior_games
• Reason: one game helps, two games are threatened, and the cross-slate starter bias is nonpersistent.
• Expected consequence of changing it: possible improvement on outlier starts such as CHC@SD, but degradation on BOS@NYY and CHW@HOU.
• Action: retain current setting and continue collecting starter ER/outs MAE.

6. Do not make a recent/season run-mean production change today

• Lever: recent.* / season.* run-mean weighting
• Reason: it helps two large score misses but risks CHW@HOU; the evidence concerns score error, not winner failure.
• Action: use the lever only in a bounded shadow experiment alongside the variance test, not as an immediate promotion.

---

9. Bottom line

Known

• All four published winners were correct.
• Three games had materially wrong scores or margins; BOS@NYY and CHC@SD were the clearest examples.
• The score model systematically compressed margins: predicted margins 1,1,1,2 versus actual margins 2,3,8,9.
• The run-model holdout signal is near zero: total R² 0.002415.
• The variance stage temporarily moved two eventual winners in the wrong direction.
• All game-level structural audits passed with zero critical, error, or warning findings.
• Probability lineage is complete and records every material transformation; no unexplained publication bypass is shown.
• Simulation inertness has persisted for 11 consecutive slates and is an established code/diagnostic issue.
• No top-confidence-pick loss streak is currently present.

Uncertain

• Whether the current calibration error reflects a true confidence defect or a four-game sample artifact.
• Whether widening the score distribution will improve rolling performance or merely reduce precision in games like CHW@HOU.
• Whether starter-form weighting can improve outlier handling without harming accurate starter projections.
• How much of the score error is reducible versus ordinary baseball variance.
• The immutable revision status, because the dedicated integrity revision audit is missing.

Single safest next action

Run a non-promoting shadow diagnostic that verifies the simulation module is actually executing an independent, non-steered probability distribution, while preserving the current published pipeline.

This is safest because the failure has the strongest persistence evidence—11 consecutive slates—is registered as a code/diagnostic problem rather than a speculative parameter issue, and can be evaluated with rollback monitors on team-run MAE, total-run MAE, margin-tail coverage, and winner Brier score without changing today’s production picks.