# Luna deep analysis — slate 20260907 > Coverage receipt: **COMPLETE** · 9/8 sections; 11/11 games; 5/5 causes; 5/6 sources # 1. Evidence receipt and gaps ## Sources received All seven `source_manifest` entries were checked against the supplied evidence: | Source | Status | Purpose | |---|---:|---| | `prediction_review_20260907.json` | Loaded | Official outcomes and scoring errors | | `PREDICTION_SYSTEM_DEEP_AUDIT_20260907.json` | Loaded | Probability, provenance, starter, and decision contracts | | `continuous_learning_daily_20260907.json` | Loaded | Pipeline-stage quality, calibration, probability lineage, score distribution, and starter comparisons | | `detailed_20260907.html` | Loaded | Full game and simulation rendering | | `20260907_ultra_detailed_player_audit.html` | Loaded | Player-level prediction trace | | `audit_report_20260907.html` | Loaded | Invariant findings | | `audit_report_20260907_integrity_revision.html` | **Missing** | Post-Luna remediation proof for immutable revision artifacts | The missing integrity-revision audit is the only explicitly missing source. It prevents independent confirmation of immutable revision-artifact remediation, although the supplied structural and per-game audits report no critical or error findings. ## Slate coverage All 11 expected games were reviewed: `ARI@KC`, `ATL@PHI`, `CHC@MIL`, `CIN@LAD`, `CLE@BAL`, `LAA@BOS`, `MIN@DET`, `NYM@MIA`, `STL@SF`, `TOR@OAK`, and `WAS@SD`. Coverage is internally complete for the principal slate artifacts: - 11 expected games - 11 canonical artifacts loaded - 11 continuous-learning games - 11 forensic deep-audit games - 11 structural audits - 11 starter comparisons - 11 games with structural QA This agrees with `coverage_matrix.all_games_structurally_audited=true`. ## Gaps and contradictions 1. **Missing integrity revision audit.** `integrity revision audit` is marked `missing`; no bytes are supplied. 2. **Four unexplained probability-publication deltas.** The causal register identifies `ARI@KC`, `CIN@LAD`, `LAA@BOS`, and `NYM@MIA` as having unexplained calibrated-to-published deltas of at least two percentage points. This is a critical evidence/contract failure until traced. 3. **Probability-lineage representation is not uniformly intuitive.** Some records show a calibrated probability above 0.50 while the published pick probability is below 0.50 because the selected side changes from the home side to the away side. That is not necessarily a defect, but the lineage must clearly identify whether each probability is: - probability of the home team, - probability of the selected side, or - probability after side inversion. 4. **ARI starter identity inconsistency.** The game-level starter line lists away starter `Jose Cabrera`, with `identity_match=false`, actual 0.00 ERA over 6 outs, and no settled pitching line. The player-forensics block instead identifies `Noah Cameron` as the relevant Kansas City starter and explicitly reports the Arizona starter as missing. This is a real data-reconciliation issue, even though that game’s structural audit reports zero findings. 5. **Audit finding names are not fully supplied.** The structural summary reports `0 critical`, `0 error`, `0 warning`, and `81 info`, but the individual 81 informational finding texts are not present in the supplied object. They therefore cannot be enumerated by title without inventing evidence. 6. **No market cross-check.** No source-labelled moneyline was available for any game. Consequently, market agreement or disagreement cannot be evaluated. 7. **No actual player results may be inferred beyond supplied lines.** The analysis below uses only the listed player, starter, lineup, and run-deviation records. --- # 2. Executive diagnosis ## Overall result The slate finished **6–5**, or **54.55% winner accuracy**. That is below the supplied walk-forward baseline accuracy of **55.53%** and below the historical validation winner accuracy of **59.16%**. The system’s own headline is therefore directionally correct: - Slate winner accuracy: **0.5455** - Walk-forward baseline: **0.555307** - Historical model validation accuracy: **0.591647** - Slate Brier score is not explicitly aggregated, but individual Brier values show substantial dispersion, including **0.489381 for TOR@OAK** and **0.392833 for STL@SF** - High-conviction games: 4 - High-conviction misses: 3 - High-conviction miss rate: **75%** The slate should not be described as a good prediction merely because it got six winners correct. ## Winner accuracy Winner performance was weak but not catastrophically below chance: - Correct winners: 6 of 11 - Incorrect winners: 5 of 11 - Correct-winner/wrong-score games: `ATL@PHI` and `WAS@SD` - Toss-up calls that happened to win: `CHC@MIL` and `CLE@BAL` - Strong calls that won: `LAA@BOS` - Strong calls that lost: `STL@SF` and `TOR@OAK` The five misses were: `ARI@KC`, `MIN@DET`, `NYM@MIA`, `STL@SF`, and `TOR@OAK`. ## Calibration quality Calibration was poor on this slate: - Weighted absolute calibration error: **0.2611** - The 0.65–0.70 bucket had average confidence **0.6996**, zero hits, and observed hit rate **0.0** - The 0.55–0.60 bucket had average confidence **0.5796**, zero hits, and observed hit rate **0.0** - The 0.60–0.65 bucket had average confidence **0.625**, one hit in two games - The 0.50–0.55 bucket had five hits in seven games, observed hit rate **0.7143**, and was underconfident by the stated definition The current slate’s confidence was anti-predictive at the upper end: strong convictions lost three of four times. This is material calibration evidence, but the causal register correctly warns not to fit a stable calibration curve from one slate. The high-raw cluster was specifically: - `TOR@OAK` - Calibrated probability: **0.6996** - Empirical historical hit rate for the relevant cluster: **0.5614** - Outcome: miss That is a warning against treating raw or governed confidence near 0.70 as reliable. ## Score accuracy Score accuracy was materially worse than winner accuracy: - Holdout run-model R²: - Away runs: **0.020536** - Home runs: **0.012404** - Total runs: **0.018362** - Pipeline final team-run MAE: **1.9091** - Pipeline final total-run MAE: **3.0909** - Pipeline final winner accuracy: **0.5455** - Actual margins of at least four runs: **1 of 11** - Predicted margins of two runs or less: **10 of 11** - Mean final margin bias: **-0.364** - Margin trend across slates is reported as improving, but the current model still compressed most games toward close outcomes. The exact-score system is therefore low-confidence. A correct winner frequently coexisted with a poor score forecast. ## Starter accuracy Starter projections were highly volatile: - Five of 20 listed starters exceeded projected ERA by at least 3.00 in the matchup-pattern analysis. - Two starters had fewer than 12 prior appearances in the feed: `Jonah Tong` with 3 and `Nick Pivetta` with 4. - Starter variance accounted for many large game-level deviations, but not all winner misses. - The strongest recurring starter-specific evidence was the credible venue split for `Chase Burns` in `CIN@LAD`; the supplied diagnosis recommends increasing `context.venue_adjustment_weight` in shadow validation. - The diagnosis does **not** support blindly tuning starters on every outlier. It explicitly classifies many of the largest deviations as variance. ## Artifact integrity Artifact integrity was generally good, but not clean in the broadest contractual sense: - Every game’s structural audit reported `PASS`. - Every game reported zero critical, error, and warning findings. - The aggregate structural summary reported: - Critical: 0 - Error: 0 - Warning: 0 - Informational: 81 - However, four probability-lineage records have unexplained material publication deltas. - The integrity-revision audit is missing. - ARI contains a starter identity/missing-line inconsistency. Therefore: - **Presentation/file integrity:** generally passed. - **Probability contract integrity:** failed for four games pending reconciliation. - **Starter-data integrity:** at least one game, ARI, requires reconciliation. - **Model quality:** weak run signal and poor slate calibration are independent of artifact cleanliness. ## Causal issue register reconciliation ### Critical: `unexplained_probability_publication_delta` - **Severity:** Critical - **Evidence:** Four games have unexplained calibrated-to-published deltas of at least two percentage points. - **Affected games:** `ARI@KC`, `CIN@LAD`, `LAA@BOS`, `NYM@MIA` - **Recommended action:** Block promotion until the missing probability transformation is recorded and reconciled. - **Guardrail:** Do not assume the probability governor was bypassed until the pick-side math is traced. This is a contract failure, not merely a calibration problem. ### High: `single_slate_calibration_instability` - **Severity:** High - **Evidence:** Weighted absolute calibration error **0.2611 across 11 games** - **Affected games:** all 11 - **Recommended action:** Observe the same buckets over the full rolling window before changing calibration weights. - **Guardrail:** Do not treat this slate as sufficient to establish a stable calibration curve. ### High: `weak_run_model_signal` - **Severity:** High - **Evidence:** Home/away/total run R² of **0.0124/0.0205/0.0184** - **Affected games:** all 11 - **Recommended action:** Treat exact scores as low confidence and investigate features/targets. - **Guardrail:** Do not infer that small positive R² makes exact scores reliable. ### Info: `complete_slate_audit_coverage` - **Severity:** Informational - **Evidence:** 11/11 structurally audited - **Affected games:** all 11 - **Recommended action:** Retain the per-game audit receipt. - **Guardrail:** Full coverage does not cancel the findings within audited games. ### Info: `simulation_inert_on_slate` - **Severity:** Informational on this record, but escalated to a persistent failure in cross-slate evidence - **Evidence:** 0 flip candidates and 0 published flips - **Affected games:** all 11 - **Recommended action:** Measure simulation contribution over rolling slates. - **Guardrail:** Zero winner flips is not itself a defect and does not prove that no score reconciliation occurred. --- # 3. Every game The probability below is the published probability for the selected side unless otherwise noted. A positive confidence edge means the home side was favored; a negative edge means the away side was favored. ## `ARI@KC` — miss; wrong winner and wrong score - **Predicted:** ARI 3, KC 5; winner KC - **Actual:** ARI 5, KC 4; winner ARI - **Published selected-side probability:** KC **0.579623** - **System confidence edge:** home-minus-away **+0.1592** - **Bucket:** Moderate - **Brier:** **0.335963** - **Team-run absolute error:** **3** - **Total-run bias:** **-1**; predicted total 8 versus actual total 9 - **Primary cause:** Wrong conviction side - **Secondary cause:** None ### Edges and controls - Season home-minus-away edge: **-0.4026** - Recent edge: **+0.204** - Pitcher edge: **-0.0608** - Composite edge: **+0.0558** - Starter-experience shrink applied: yes - KC starter history: 27 career games in feed - Opposing starter record was thin/missing in the settled feed - Score-volatility action: `no_change` - Published probability was lower than the pre-publication pipeline value: governed **0.618**, published **0.579623** ### Starter projection versus actual - ARI listed starter `Jose Cabrera`: projected ERA **3.86**, projected 2 ER/14 outs; listed actual 0.00 ERA, 0 ER/6 outs; identity mismatch and no settled pitching line - KC `Noah Cameron`: projected ERA **1.5**, projected 1 ER/18 outs; actual 0.00 ERA, 0 ER/18 outs - Forensic block separately records KC starter `Noah Cameron` with projected 1 ER and actual 0 ER, and identifies the ARI starter line as missing ### Score path and player adjustments - Historical ML: ARI 4.38, KC 5.08 - Variance-applied: 4–5 - Enhanced: 3–5 - Savant/environment: 3–5 - Final simulation: 3–5 - Important player adjustments included Nick Loftin, Jac Caglianone, Ildemaro Vargas, Bobby Witt Jr., Carter Jensen, Kyle Isbel, Nolan Arenado, and Jordan Lawlar. - Player-level result highlights included Salvador Pérez outperforming projection and Corbin Carroll/Bobby Witt Jr. underperforming it. ### Audit and causal assessment - Structural audit: PASS; zero critical/error/warning findings - Forecast artifact: reported clean - The miss is best classified as a moderate-confidence directional error with starter-feed inconsistency and player variance. The supplied forensic verdict says neither side broke materially and calls the result closer to variance. --- ## `ATL@PHI` — correct winner, wrong score - **Predicted:** ATL 3, PHI 4; winner PHI - **Actual:** ATL 0, PHI 1; winner PHI - **Published probability:** PHI **0.53** - **Confidence edge:** **+0.06** - **Bucket:** Lean - **Brier:** **0.2209** - **Team-run absolute error:** **6** - **Total-run bias:** **+6**; predicted 7 versus actual 1 - **Winner quality:** Right winner, wrong score - **Primary/secondary causes:** None assigned ### Edges and controls - Season edge: **-0.390** - Recent edge: **+0.030** - Pitcher edge: **-0.1125** - Composite edge: **-0.2745** - Recency guardrail: applied - Starter-experience shrink: not applied - Large Savant run-environment swing flag: present - Score-volatility action: `no_change` ### Starters - ATL Grant Holmes: projected 2 ER/14 outs, projected ERA 3.86; actual 0 ER/18 outs, 0.00 ERA - PHI Jesús Luzardo: projected 2 ER/19 outs, projected ERA 2.84; actual 0 ER/27 outs, 0.00 ERA - Both starters materially beat projections. ### Score path and player adjustments - Published score: 3–4 - The supplied score-path details are less extensive than for several other games, but the actual result was a one-run, one-total-run game. - Major player-level deviations included Michael Harris II, Lane Thomas, Drake Baldwin, Bryson Stott, and Ozzie Albies. - Both offenses scored well below projection. ### Audit and causal assessment - Structural audit: PASS - This is a correct-winner/wrong-score game. The score miss was driven by simultaneous low offensive output and unexpectedly dominant starting pitching, not by a wrong side. --- ## `CHC@MIL` — correct winner, poor score - **Predicted:** CHC 5, MIL 7; winner MIL - **Actual:** CHC 3, MIL 4; winner MIL - **Published probability:** MIL **0.506818** - **Confidence edge:** **+0.0136** - **Bucket:** Toss-up - **Brier:** **0.243228** - **Team-run absolute error:** **5** - **Total-run bias:** **+5**; predicted 12 versus actual 7 - **Winner quality:** Toss-up went our way ### Edges and controls - Season edge: **+0.1133** - Recent edge: **-0.346** - Pitcher edge: **+0.0785** - Composite edge: **+0.5085** - Recency guardrail: applied - Starter-experience shrink: not applied - Score-volatility action: `no_change` ### Starters - CHC Matthew Boyd: projected 4 ER/17 outs, ERA 6.35; actual 3 ER/21 outs, ERA 3.86 - MIL Robert Gasser: projected 4 ER/15 outs, ERA 7.20; actual 1 ER/14 outs, ERA 1.93 - Gasser was the largest starter miss: 3 fewer ER than projected. ### Score path and player adjustments - The exact score path is not included in the game-level block, but the final prediction was 5–7. - Player-level contributors included William Contreras, Pedro Ramirez, and Jackson Chourio. - CHC offense scored below projection; MIL was classified as not materially broken on the offensive side. ### Audit and causal assessment - Structural audit: PASS - This was a near-coin-flip call. The winner was correct, but the score model overestimated scoring by five runs. Starter variance, particularly Gasser’s unexpected performance, was central. --- ## `CIN@LAD` — correct winner, wrong score - **Predicted:** CIN 5, LAD 6; winner LAD - **Actual:** CIN 3, LAD 6; winner LAD - **Published probability:** LAD **0.532561** - **Confidence edge:** **+0.0651** - **Bucket:** Lean - **Brier:** **0.218499** - **Team-run absolute error:** **2** - **Total-run bias:** **+2** - **Winner quality:** Solid correct call ### Edges and controls - Season edge: **+1.2463** - Recent edge: **-0.108** - Pitcher edge: **+0.076** - Composite edge: **+1.0057** - Season-quality margin adjustment: **+0.5234** - Score-volatility action: supported blowout-margin expansion - No starter-experience shrink ### Starters - CIN Chase Burns: projected 3 ER/17 outs, ERA 4.76; actual 0 ER/9 outs, 0.00 ERA - LAD Emmet Sheehan: projected 2 ER/15 outs, ERA 3.60; actual 1 ER/17 outs, ERA 1.59 - Burns was three ER better than projected; the forensic assessment notes a credible away venue weakness that was underweighted. ### Score path and player adjustments - Final score path ended at CIN 5, LAD 6. - Important players: Teoscar Hernández exceeded projection by 1.42 runs; Shohei Ohtani underperformed by 1.40; Tyler Stephenson and other Cincinnati hitters also deviated. - The supplied forensic recommendation is to enable or raise `context.venue_adjustment_weight` in shadow validation. ### Audit and causal assessment - Structural audit: PASS - Correct winner, somewhat inflated Cincinnati score, and a starter projection miss. The venue-split signal is more actionable than the one-game player deviations. --- ## `CLE@BAL` — correct winner, wrong score - **Predicted:** CLE 3, BAL 4; winner BAL - **Actual:** CLE 4, BAL 6; winner BAL - **Published probability:** BAL **0.507765** - **Confidence edge:** **+0.0155** - **Bucket:** Toss-up - **Brier:** **0.242295** - **Team-run absolute error:** **3** - **Total-run bias:** **-3**; predicted 7 versus actual 10 - **Winner quality:** Toss-up went our way ### Edges and controls - Season edge: **-0.1158** - Recent edge: **-0.235** - Pitcher edge: **-0.0245** - Composite edge: **+0.0037** - Recency guardrail: applied - Large Savant run-environment swing flag: present - Score-volatility action: `no_change` ### Starters - CLE Joey Cantillo: projected 2 ER/14 outs, ERA 3.86; actual 4 ER/7 outs, ERA 15.43 - BAL Trevor Rogers: projected 3 ER/17 outs, ERA 4.76; actual 1 ER/22 outs, ERA 1.23 - Cantillo was the major negative starter miss, while Rogers outperformed projection. ### Score path and player adjustments - Player deviations included Coby Mayo’s two-homer performance, Rece Hinds exceeding projection, and below-projection results from Brayan Rocchio and David Fry. - The model underpredicted the ten-run environment by three runs. ### Audit and causal assessment - Structural audit: PASS - Correct winner at essentially even odds. Score error was driven by highly variable starter and player outcomes rather than a clear directional model failure. --- ## `LAA@BOS` — strong correct winner, close score - **Predicted:** LAA 3, BOS 5; winner BOS - **Actual:** LAA 2, BOS 5; winner BOS - **Published probability:** BOS **0.62315** - **Confidence edge:** **+0.2463** - **Bucket:** Strong - **Brier:** **0.142016** - **Team-run absolute error:** **1** - **Total-run bias:** **+1** - **Winner quality:** Strong correct call ### Edges and controls - Season edge: **+1.0127** - Recent edge: **+0.658** - Pitcher edge: **+0.0198** - Composite edge: **+1.125** - Season-quality margin adjustment: **+0.4253** - Score-volatility action: `no_change` ### Starters - LAA Grayson Rodriguez: projected 3 ER/15 outs, ERA 5.40; actual 5 ER/12 outs, ERA 11.25 - BOS Brayan Bello: projected 2 ER/13 outs, ERA 4.15; actual 0 ER/15 outs, ERA 0.00 - Both starter outcomes favored Boston, despite Rodriguez being worse than projected and Bello much better. ### Score path and player adjustments - Player-level effects included Moisés Ballesteros, Jarren Duran, Mike Trout, Isiah Kiner-Falefa, and Josh Lowe. - The score was close to the actual result despite starter-level deviations. ### Audit and causal assessment - Structural audit: PASS - This is evidence against blindly widening every run distribution. The variance lever that could help several high-confidence misses would also soften this strong correct call. --- ## `MIN@DET` — miss; wrong winner and wrong score - **Predicted:** MIN 5, DET 4; winner MIN - **Actual:** MIN 4, DET 5; winner DET - **Published probability:** selected away side MIN **0.53**; home probability **0.470** - **Confidence edge:** home-minus-away **-0.06** - **Bucket:** Lean - **Brier:** **0.2809** - **Team-run absolute error:** **2** - **Total-run bias:** **0** - **Primary cause:** Mixed - **Secondary cause:** None ### Edges and controls - Season edge: **+0.257** - Recent edge: **-0.550** - Pitcher edge: **-0.046** - Composite edge: **-0.2773** - Recency guardrail: applied - Supported blowout-margin expansion was applied - Head-to-head history contained several high-scoring and blowout games ### Starter projections - MIN Joe Ryan: projected 4 ER/17 outs, ERA 6.35; actual 2 ER/12 outs, ERA 4.50 - DET Troy Melton: projected 2 ER/18 outs, ERA 3.00; actual 4 ER/17 outs, ERA 6.35 - DET’s starter was worse than projected by 2 ER; MIN’s starter was better by 2 ER. The starter results actually supported the MIN side, but not by enough to offset other factors. ### Score path and player adjustments - Historical ML: approximately 4.38–4.38 - Variance-applied: 4–4 - Enhanced prediction: 5–4, moving the winner toward MIN - Savant/environment: approximately 5–4 - Final simulation: 5–4 - The earliest winner-direction regression was the `variance_applied → enhanced_prediction` transition. - Player adjustments included Colt Keith, Troy Melton, Luke Keaschall, Brooks Lee, Kody Clemens, Zach McKinstry, Riley Greene, and Ryan Jeffers. ### Audit and causal assessment - Structural audit: PASS - This is a wrong-winner/wrong-score game. The model’s enhanced stage converted a tie into a MIN prediction, despite recent form favoring MIN but season quality favoring DET. The evidence supports a weighting/transition issue more than a pure starter miss. --- ## `NYM@MIA` — miss; wrong winner and wrong score - **Predicted:** NYM 4, MIA 5; winner MIA - **Actual:** NYM 9, MIA 4; winner NYM - **Published probability:** MIA **0.520771** - **Confidence edge:** **+0.0415** - **Bucket:** Lean - **Brier:** **0.271202** - **Team-run absolute error:** **6** - **Total-run bias:** **-4**; predicted 9 versus actual 13 - **Primary cause:** Lean-side miss - **Secondary cause:** None ### Edges and controls - Season edge: **+0.291** - Recent edge: **-0.042** - Pitcher edge: **+0.0386** - Composite edge: **+0.4145** - Inexperienced-starter shrink applied; shrink factor **0.5125** - The selected side remained MIA after the shrink - Score-volatility action: `no_change` ### Starters - NYM Jonah Tong: projected 3 ER/10 outs, ERA 8.10; actual 4 ER/15 outs, ERA 7.20 - MIA Eury Pérez: projected 3 ER/16 outs, ERA 5.06; actual 5 ER/12 outs, ERA 11.25 - Pérez was worse than projected by 2 ER and was the main starter-side deterioration. ### Score path and player adjustments - Historical ML: NYM 3.50, MIA 4.81 - Variance-applied: 3–5 - Enhanced: 4–5 - Savant/environment: 4–5 - Final: 4–5 - NYM offense exceeded projection by 5 runs. - Important deviations included Juan Soto, Nick Morabito, A.J. Ewing, Graham Pauley, Bo Bichette, and others. - Several of the positive batter deviations were not necessarily supported by strong prior profiles; the audit identifies only 18 of 72 material player deviations across the slate as learnable. ### Audit and causal assessment - Structural audit: PASS - This was a small-edge loss, not a near-certain miss. A confidence edge of approximately 0.04 is a lean. The score model also missed the offensive breakout. --- ## `STL@SF` — strong miss; wrong winner and wrong score - **Predicted:** STL 7, SF 5; winner STL - **Actual:** STL 4, SF 5; winner SF - **Published probability:** STL **0.626764** - **Confidence edge:** home-minus-away **-0.2535** - **Bucket:** Strong - **Brier:** **0.392833** - **Team-run absolute error:** **3** - **Total-run bias:** **+3**; predicted 12 versus actual 9 - **Primary cause:** Wrong conviction side - **Secondary cause:** None ### Edges and controls - Season edge: **-0.5161** - Recent edge: **-0.372** - Pitcher edge: **-0.1027** - Composite edge: **-0.9628** - Season-quality margin adjustment: **-0.2168** - Supported blowout-margin expansion applied - Final away margin was two runs, with the model selecting STL strongly ### Starters - STL Michael McGreevy: projected 2 ER/16 outs, ERA 3.38; actual 2 ER/20 outs, ERA 2.70 - SF Logan Webb: projected 4 ER/16 outs, ERA 6.75; actual 1 ER/18 outs, ERA 1.50 - Webb beat projection by 3 ER and was the central starter miss. ### Score path and player adjustments - Historical ML: 4.26–4.78, favoring SF - Variance-applied: 4–5, still SF - Enhanced prediction: 7–5, flipping to STL - Savant/environment: approximately 7–5 - Final: 7–5 - The earliest winner-direction regression occurred at `variance_applied → enhanced_prediction`. - Player adjustments included Bryan Torres, Leonardo Bernal, Turner Hill, Iván Herrera, Andrew Knizner, Jonah Cox, and Jung Hoo Lee. ### Audit and causal assessment - Structural audit: PASS - This is one of the clearest process/model risks. The model moved from an initially SF-favorable prediction to a strong STL conviction without a corresponding realized result. The starter miss was real, but the high conviction magnified its cost. --- ## `TOR@OAK` — strongest miss; wrong winner and wrong score - **Predicted:** TOR 6, OAK 2; winner TOR - **Actual:** TOR 5, OAK 6; winner OAK - **Published probability:** TOR **0.699558** - **Confidence edge:** home-minus-away **-0.3991** - **Bucket:** Strong - **Brier:** **0.489381** - **Team-run absolute error:** **5** - **Total-run bias:** **-3**; predicted 8 versus actual 11 - **Primary cause:** Wrong conviction side - **Secondary cause:** Lean-side miss ### Edges and controls - Season edge: **-0.9732** - Recent edge: **-0.906** - Pitcher edge: **-0.1731** - Composite edge: **-1.8626** - High-margin-pick diagnostic: present - Large Savant run-environment swing: present - Supported blowout-margin expansion applied - High-raw cluster member; historical empirical hit rate for the cluster was **0.5614** ### Starters - TOR Dylan Cease: projected 2 ER/18 outs, ERA 3.00; actual 4 ER/15 outs, ERA 7.20 - OAK Jacob Lopez: projected 5 ER/15 outs, ERA 9.00; actual 4 ER/18 outs, ERA 6.00 - Cease was worse than projected by 2 ER; Lopez was better than projected by 1 ER. ### Score path and player adjustments - Historical ML: TOR 5.90, OAK 3.64 - Variance-applied: 6–4 - Enhanced: 6–3 - Savant/environment: 6–2 - Final simulation: 6–2 - The score path repeatedly reinforced the already incorrect TOR conviction. - Player adjustments included Dylan Cease, Nathan Lukes, Zack Gelof, Alejandro Kirk, Carlos Cortes, Myles Straw, Vladimir Guerrero Jr., and Jonah Heim. - OAK offense was underprojected by four runs. ### Audit and causal assessment - Structural audit: PASS - This is the most damaging individual miss: - Highest published confidence - Highest Brier error among the slate - High-raw cluster failure - Large margin error - It supports lowering confidence caps and widening tails, but not a one-game reversal rule for Toronto/Oakland. --- ## `WAS@SD` — correct winner, wrong score - **Predicted:** WAS 5, SD 6; winner SD - **Actual:** WAS 2, SD 3; winner SD - **Published probability:** SD **0.529654** - **Confidence edge:** **+0.0593** - **Bucket:** Lean - **Brier:** **0.221225** - **Team-run absolute error:** **6** - **Total-run bias:** **+6**; predicted 11 versus actual 5 - **Winner quality:** Right winner, wrong score ### Edges and controls - Season edge: **-0.0369** - Recent edge: **-0.3847** - Pitcher edge: **-0.0934** - Composite edge: **-0.1419** - Starter-experience/recency guardrail applied - SD starter had only four prior appearances in the feed - Large Savant run-environment swing flag present - Score-volatility action: `no_change` ### Starters - WAS Jake Irvin: projected 3 ER/14 outs, ERA 5.79; actual 3 ER/17 outs, ERA 4.76 - SD Nick Pivetta: projected 4 ER/15 outs, ERA 7.20; actual 0 ER/15 outs, ERA 0.00 - Pivetta was four ER better than projected; the thin four-appearance history explains high projection uncertainty. ### Score path and player adjustments - Important player deviations included CJ Abrams, Dylan Crews, Jackson Merrill, Luis Campusano, and Keibert Ruiz. - Both offenses scored below projection. - Merrill’s positive deviation was identified as one of the more learnable player-level deviations because of above-average xwOBA. ### Audit and causal assessment - Structural audit: PASS - Correct winner/wrong score. The supplied recommendation is to reduce confidence for low-history starters, prefer priors, and validate run-parameter changes in shadow. --- # 4. Miss reconstruction ## Earliest wrong-direction stages | Loss | Earliest stage moving toward wrong result | Reconstruction | |---|---|---| | `ARI@KC` | Historical ML already favored KC | The wrong side was present before later controls; subsequent stages maintained KC. Starter identity/feed inconsistency is a data-quality concern, but no later winner-direction flip is documented. | | `MIN@DET` | `enhanced_prediction` | Historical ML was approximately tied; variance stayed tied; enhanced prediction changed `TIE → MIN`, which was wrong. | | `NYM@MIA` | Historical ML already favored MIA | Variance, enhanced, Savant, and simulation all maintained MIA. This was a small-edge miss rather than a later regression. | | `STL@SF` | `enhanced_prediction` | Historical ML and variance favored SF; enhanced prediction changed `SF → STL`, and later stages preserved the wrong side. | | `TOR@OAK` | Historical ML already favored TOR | All stages reinforced TOR, with Savant and score-volatility adjustments widening the wrong margin. | A confidence edge of **0.12** would be moderate, not a near coin flip. More generally, the system’s own edges must not be interpreted as half-scale distances from 50%. The ARI edge was approximately 0.16 and still lost; STL and TOR had much stronger edges and also lost. Conversely, several 0.01–0.06 edges won because they were genuinely uncertain calls. ## Loss-type classification ### Bad input - ARI has the clearest input/contract issue: - Away starter identity mismatch - Missing settled pitching line for the listed away starter - No other game has a reported structural input error. ### Bad weighting or transition - `MIN@DET`: `enhanced_prediction` changed a tie into the wrong MIN direction. - `STL@SF`: `enhanced_prediction` changed an SF direction into the wrong STL direction. - `TOR@OAK`: strong season/recent/compound edges were allowed to produce a four-run predicted margin, despite the broader slate showing severe margin compression and high uncertainty. ### Calibration - The most consequential calibration miss was `TOR@OAK` at published probability **0.699558**. - `STL@SF` also received a strong 0.626764 probability and lost. - Three of four high-conviction selections lost. - The appropriate conclusion is overconfidence at the upper end, not that every moderate-confidence call should be flattened. ### Starter error - `STL@SF`: Logan Webb, actual ERA 1.50 versus projected 6.75. - `TOR@OAK`: Dylan Cease, actual ERA 7.20 versus projected 3.00. - `MIN@DET`: Troy Melton actual ERA 6.35 versus projected 3.00. - `NYM@MIA`: Eury Pérez actual ERA 11.25 versus projected 5.06. - `ARI@KC`: the listed away starter is unresolved; KC’s Noah Cameron beat projection modestly. - Starter error explains substantial score and directional pressure, but these are mostly outlier outcomes rather than proven systematic starter-model failures. ### Offense/bullpen miss - `NYM@MIA`: NYM offense exceeded projection by 5 runs. - `TOR@OAK`: OAK offense exceeded projection by 4 runs relative to the matchup breakdown. - `STL@SF`: STL offense scored three fewer than projected. - `MIN@DET`: opposing offensive and starter effects were mixed. - Bullpen effects were material in some games, such as KC’s four ER from Craig Kimbrel and BAL’s three ER from Andrew Kittredge, but those did not independently explain all winner misses. ### Run-environment miss The slate overprojected total scoring in: - `ATL@PHI`: +6 - `CHC@MIL`: +5 - `WAS@SD`: +6 It underprojected: - `NYM@MIA`: -4 - `CLE@BAL`: -3 - `TOR@OAK`: -3 This supports the persistent weak-run-signal diagnosis rather than a single consistent high- or low-scoring bias. ### Irreducible variance The player forensics identify: - 72 material player deviations - Only 18 supported by the hitter’s own season contact quality - Learnable share: **25%** - Variance: **54** deviations Several major starter outcomes were also explicitly labelled outliers. A model change based on every individual over- or underperformance would learn noise. ## Pipeline-stage quality across all stages The stored stage metrics were: | Stage | Winner accuracy | Team-run MAE | Total-run MAE | Margin MAE | |---|---:|---:|---:|---:| | `historical_ml` | 6/11 = 0.5455 | 1.5782 | 2.3421 | 2.1950 | | `variance_applied` | 5/11 = 0.4545 | 1.5909 | 2.4545 | 2.0909 | | `enhanced_prediction` | 6/11 = 0.5455 | 1.8182 | 2.9091 | 2.0000 | | `embedded_savant_run_environment` | 6/11 = 0.5455 | 1.9091 | 3.0909 | 2.1818 | | `fundamental_edge_target_review` | 6/11 = 0.5455 | 1.9091 | 3.0909 | 2.1818 | | `capped_target` | 6/11 = 0.5455 | 1.9091 | 3.0909 | 2.1818 | | `simulation_result` | 6/11 = 0.5455 | 1.9091 | 3.0909 | 2.1818 | | `published_final` | 6/11 = 0.5455 | 1.9091 | 3.0909 | 2.1818 | ### Transition audit 1. **Historical ML → variance applied** - Winner accuracy fell from 6 to 5. - `CIN@LAD`, `CLE@BAL`, and `MIN@DET` moved toward ties. - `CLE@BAL` is explicitly listed as regressed. - Margin MAE improved by 0.1041, but winner accuracy worsened by 0.091. 2. **Variance applied → enhanced prediction** - Winner accuracy recovered from 5 to 6. - `CIN@LAD` and `CLE@BAL` improved. - `STL@SF` regressed from SF to STL. - Team-run MAE increased by 0.2273. - This transition contains a documented intermediate winner-direction regression; the later aggregate recovery does not erase it. 3. **Enhanced prediction → embedded Savant run environment** - No winner-direction changes. - Team-run MAE worsened by 0.0909. - Total-run MAE worsened to 3.0909. - Margin MAE worsened by 0.1818. 4. **All later transitions** - No winner-direction changes. - No measurable score changes. - `simulation_result → published_final` made no change to any recorded metric. The simulation therefore did not intervene in winner direction on this slate, and this same inertness has persisted for 11 consecutive slates. --- # 5. Cross-game counterfactuals ## Lever tension: widen run distributions **Lever:** `score.variance_multiplier / score.variance_floor` **Target metrics:** Brier score and calibration error ### Games helped - `ARI@KC`: moderate-confidence miss; wider tails reduce overconfidence. - `STL@SF`: strong-confidence miss; wider tails reduce overconfidence. - `TOR@OAK`: strongest miss; wider tails reduce overconfidence. ### Game hurt - `LAA@BOS`: strong correct call; widening the distribution would soften a good edge. ### Net evidence - 3 games helped versus 1 hurt. - The lever tension supports a bounded shadow experiment, not an unrestricted increase. - This lever is primarily about uncertainty and score tails; it should not be presented as a guaranteed winner-accuracy fix. ## Lever tension: pitcher form weights and prior-game shrinkage **Lever:** `player.pitcher.v4 form weights / player.pitcher_prior_games` **Target metrics:** starter ER MAE and starter outs MAE ### Games helped - `CHC@MIL`: Gasser was three ER better than projected. - `CIN@LAD`: Burns was three ER better than projected. - `STL@SF`: Webb was three ER better than projected. - `WAS@SD`: Pivetta was four ER better than projected. ### Games hurt - None listed. ### Net evidence - 4 helped, 0 hurt in the supplied slate-level tension table. - However, the same document says many starter/player deviations are variance and that persistent cross-slate starter ERA bias is not separable from noise. - A bounded shadow test is reasonable; direct promotion is not. ## Lever tension: recent/season run-mean weighting **Lever:** `recent.* / season.* run-mean weighting` **Target metrics:** team-run MAE, total-run MAE, accuracy ### Games helped - `ATL@PHI` - `CHC@MIL` - `NYM@MIA` - `TOR@OAK` - `WAS@SD` ### Games hurt - `CIN@LAD` - `LAA@BOS` - `MIN@DET` ### Net evidence - 5 helped versus 3 hurt. - The net count is mildly favorable, but the persistent cross-slate evidence says lineup-run bias is not separable from noise: mean **-0.06**, z-score **-0.3**, nonpersistent. - Do not make a large direct change from this slate. ## Tempting one-game fixes that must not be made 1. **Do not reverse the STL pick rule.** `STL@SF` was a severe miss, but a single game does not justify suppressing strong negative composite edges. 2. **Do not permanently downgrade Toronto or Oakland.** `TOR@OAK` was the strongest miss, but the evidence supports broader confidence/variance control, not a team-specific rule. 3. **Do not tune to individual player box scores.** Only 18 of 72 material player deviations were supported by player contact-quality evidence; the other 54 are classified as variance. 4. **Do not remove starter adjustments entirely.** Several games show useful starter information, and the pitcher-form lever helped four games with no listed hurts. 5. **Do not use one slate to alter calibration weights aggressively.** The causal register explicitly says the slate is too small for a stable confidence curve. 6. **Do not assume the probability governor was bypassed.** Trace pick-side math for the four unexplained deltas first. --- # 6. Data and contract defects ## Critical findings ### 1. Unexplained published probability transformations Affected games: - `ARI@KC`: calibrated 0.618; published 0.579623; delta -0.038377 - `CIN@LAD`: calibrated 0.5751; published 0.532561; delta -0.042539 - `LAA@BOS`: calibrated 0.6539; published 0.62315; delta -0.03075 - `NYM@MIA`: calibrated 0.5477; published 0.520771; delta -0.026929 The lineage says “no recorded transformation explains the publication delta” for all four. This is a critical contract failure because a material probability transformation is missing from evidence. ## Error findings No error findings were reported in the supplied structural audits: - Aggregate structural audit: `error: 0` - Every per-game structural audit: `error: 0` This does not mean the four probability-lineage issues are resolved; they are raised in a separate causal contract register. ## Warning findings No warning findings were reported: - Aggregate: `warning: 0` - Every per-game structural audit: `warning: 0` ## Informational findings - Aggregate informational findings: **81** - Per-game informational counts range from 7 to 8. - The supplied object does not include the individual informational finding texts, so the 81 entries cannot be enumerated by name without fabricating details. - Known informational conditions include: - Simulation produced zero winner flips - Full slate structural coverage - Score-volatility `no_change` decisions - Starter-recency and experience context - Forecast-file informational notes ## Prediction-model defects versus artifact bugs ### Prediction-model defects or risks - Near-zero run-model signal: - Total R² 0.018362 - Home R² 0.012404 - Away R² 0.020536 - Upper-tail confidence overstatement: - High-conviction miss rate 75% - TOR high-raw cluster miss - Persistent simulation inertness - Intermediate winner-direction regressions at the enhanced-prediction stage - Margin compression and poor score reliability ### Presentation/artifact/contract bugs - Missing integrity revision audit - Four unexplained calibrated-to-published probability deltas - ARI starter identity/missing-line inconsistency - No market cross-check because no source-labelled prices were supplied The per-game forecast files themselves were reported clean, so the primary defect is not a rendering or saved-file corruption issue. ## Complete probability lineage audit The required conceptual path is: **raw model → calibrated → governor → score coherence/recency → published** ### Correctly recorded transformations Examples: - `CHC@MIL`, `CLE@BAL`, `STL@SF`, `TOR@OAK`: published probability approximately matches calibrated/governed value. - `ATL@PHI`, `MIN@DET`, `WAS@SD`: lineage records score coherence followed by matchup-recency guardrail. - `ARI@KC`, `NYM@MIA`: starter-experience shrink is recorded in the game trace. ### Material unresolved transformations The following have no recorded transformation explaining the published delta: - `ARI@KC` - `CIN@LAD` - `LAA@BOS` - `NYM@MIA` These are bypass-like symptoms, but they must not be called a governor bypass until pick-side math is traced. The causal register’s guardrail is appropriate. ### No bypass conclusion for other games No material unexplained delta is present for: - `CHC@MIL` - `CLE@BAL` - `STL@SF` - `TOR@OAK` For those games, the published value matches the calibrated/governed or documented post-governor value within rounding. ## Coverage reconciliation Coverage is complete at the game level: - 11/11 expected games structurally audited - 11/11 with starter comparisons - 11/11 with continuous-learning records - 11/11 with forensic deep audits The slate is therefore structurally covered, but not contractually clean because coverage does not resolve the four probability deltas, the missing integrity revision audit, or ARI’s starter inconsistency. --- # 7. Cross-slate persistence The supplied cross-slate window contains 11 slates, with 10 prior slates loaded. ## Escalated findings — established failures These are not single-slate noise. ### `simulation_inert_on_slate` - Consecutive-slate count: **11** - Incidence in window: 11 of 11 - Present now: yes - Metric group: run model - Target lever listed: `simulation.allow_score` - Target metrics: `team_run_mae`, `total_run_mae` - Established direction of fix: confirm the simulation module is executing and not silently no-oping - Required action: code/diagnostic investigation, not a parameter promotion The self-assessment provides a specific suspected cause: the simulation distribution is built from the same draws used to search for a target-realizing play-by-play, and the loop can terminate after approximately two selected draws. That makes the estimated win rate non-probabilistic and explains zero flip candidates over 11 slates. ### `weak_run_model_signal` - Consecutive-slate count: **11** - Incidence in window: 11 of 11 - Present now: yes - Metric group: run model - Target levers listed: `training.xgb.n_estimators`, `training.xgb.max_depth`, `training.xgb.learning_rate` - Target metrics: `team_run_mae`, `total_run_mae` - Established direction of fix: investigate features and target through an isolated training backtest - Required action: code/data/training investigation, not a nightly knob adjustment The raw evidence remains poor: - Away R² 0.020536 - Home R² 0.012404 - Total R² 0.018362 The self-assessment specifically recommends adding, where valid and source-labelled: - No-vig market probability - Park/weather run environment - Bullpen rest/fatigue - Platoon-adjusted lineup quality It also requires a model schema-version bump when feature semantics change. ## Every finding streak | Finding | Consecutive count | Incidence | Present now? | Interpretation | |---|---:|---:|---:|---| | `high_raw_probability_cluster` | 0 | 3/11 | No | Historical recurrence, not current | | `pipeline_winner_instability` | 0 | 4/11 | No | Historical recurrence, not current | | `score_margin_compression` | 0 | 3/11 | No | Historical recurrence, not current | | `simulation_inert_on_slate` | 11 | 11/11 | Yes | Established persistent failure | | `single_slate_calibration_instability` | 1 | 6/11 | Yes | Current finding, not yet persistent by consecutive-slate count | | `weak_run_model_signal` | 11 | 11/11 | Yes | Established persistent failure | The fact that `single_slate_calibration_instability` occurred in 6 of 11 slates but has only a one-slate consecutive streak is important: its incidence is notable, but it has not reached the escalation status of the two 11-slate failures. ## Metric trends Each reported direction is: - **Margin bias:** improving Series ends at **-0.182**, better than earlier values as low as -3.5. - **Mean team-run error on losses:** improving Current value **3.8**, compared with earlier values up to 8.25. - **Weighted calibration error:** regressing Current value **0.2611**, worse than many preceding values, including 0.0937 and 0.1000. - **Winner accuracy:** flat Current slate **0.5455**, rolling mean **0.5667**, with substantial slate-to-slate noise. ## Top-confidence-pick loss streak The current top-confidence pick has a **loss streak of 2**: - 2026-09-06: miss - 2026-09-07: miss The full sequence includes several earlier misses, but the supplied explicit streak is two consecutive losses. This is concerning when combined with the current high-conviction miss rate of 75%, though still weaker evidence than the 11-slate escalations. ## Cross-slate bias findings The two listed cross-slate bias findings are not actionable yet: 1. **Starter ERA error positive** - Mean: **+0.281** - Same-sign share: 0.625 - 5 of 8 slates - z = 0.95 - Status: hold - No change to `context.opponent_strength_weight` 2. **Lineup run delta negative** - Mean: **-0.06** - Same-sign share: 0.625 - z = -0.3 - Status: hold - No change to `score.variance_floor` The source correctly treats these as not separable from slate-to-slate noise. --- # 8. Adjustments for today Recommendations are ranked by persistence and scope. Persistent issues receive code/diagnostic actions rather than forced parameter moves, because the supplied escalation records explicitly state that they are not parameter-actionable. ## 1. Simulation inertness — highest persistence and scope - **Finding:** `simulation_inert_on_slate` - **Persistence:** 11 consecutive slates - **Registered lever:** `simulation.allow_score` - **Parameter move:** **Abstain from changing the parameter.** - **Bounded action:** Inspect and repair the simulation execution path: - Separate target-realization search from independent win-probability simulation. - Use an independent unsteered batch of approximately 200–500 draws or an equivalent closed-form distribution. - Remove or revise the early-exit behavior around `MIN_DRAWS_BEFORE_CLOSE_EXIT`. - Register the batch size as a parameter. - Feed the independent win rate into `_distribution_agreement`. - **Expected metric:** `team_run_mae` and `total_run_mae` should become measurable and, if the module is contributing valid information, improve; winner-direction intervention should become nonzero when justified. - **Helps:** Broadly addresses all games, especially wrong-margin cases such as `TOR@OAK`, `STL@SF`, `ATL@PHI`, and `WAS@SD`. - **Hurts/regression risk:** An independent simulation may alter some correct calls, including `LAA@BOS`; rollback must monitor winner accuracy, Brier score, and score MAE. - **Evidence strength:** Highest, because the failure persisted 11 consecutive slates and has a concrete code-level diagnosis. ## 2. Weak run-model signal — persistent feature/target failure - **Finding:** `weak_run_model_signal` - **Persistence:** 11 consecutive slates - **Registered levers:** `training.xgb.n_estimators`, `training.xgb.max_depth`, `training.xgb.learning_rate` - **Parameter move:** **Do not tune these nightly parameters directly.** - **Bounded action:** Run the registered isolated training backtest after adding or validating: - No-vig market probability where source-labelled - Park/weather run environment - Bullpen fatigue/rest - Platoon-adjusted lineup quality - **Expected metric:** Improve holdout home/away/total R² and reduce team-run/total-run MAE. - **Helps:** Broad score failures such as `ATL@PHI`, `NYM@MIA`, `TOR@OAK`, and `WAS@SD`. - **Hurts/regression risk:** Feature changes could disturb close correct score calls such as `LAA@BOS` and `CIN@LAD`; require same-fingerprint holdout comparison and rollback. - **Evidence strength:** Highest, because all three run R² values are near zero and the pattern persisted 11 slates. ## 3. Upper-tail confidence overstatement — strongest current single-slate driver - **Finding:** Current high-confidence performance is anti-predictive; top-pick loss streak is 2 and high-conviction miss rate is 75%. - **Registered parameter:** `confidence.base_cap` - **Bounded change:** Reduce `confidence.base_cap` by a small amount, approximately **0.03**, in shadow only. - **Expected metric:** Lower weighted calibration error and improve Brier score by reducing overconfident losses. - **Helps:** `ARI@KC`, `STL@SF`, and especially `TOR@OAK`, all of which lost with meaningful conviction. - **Hurts:** `LAA@BOS`, a strong correct call whose published probability would be softened. - **Do not simultaneously make a large `moderate_cap` and variance-floor change:** isolate this confidence-cap experiment so the effect is identifiable. - **Evidence strength:** Strong for the current slate, but weaker than the 11-slate code failures because the calibration instability streak is only one slate. ## 4. Score-distribution shadow experiment - **Registered parameters:** `score.variance_multiplier` or `score.variance_floor` - **Bounded change:** Test a small increase in `score.variance_multiplier`, approximately **+0.05**, in shadow. - **Expected metric:** Lower Brier score, calibration error, and margin overconfidence; possibly improve total-run error. - **Helps:** `ARI@KC`, `STL@SF`, `TOR@OAK` - **Hurts:** `LAA@BOS`; potentially other strong correct calls. - **Evidence strength:** Moderate. The lever tension is 3 helped versus 1 hurt, but cross-slate lineup-run bias does not support a direct permanent change. This should remain a shadow experiment with rollback monitoring. ## 5. Starter venue/form experiment - **Registered parameter:** `context.venue_adjustment_weight` - **Bounded change:** Increase modestly in shadow, only for starters with a credible venue split. - **Expected metric:** Lower starter ERA MAE and potentially reduce total-run MAE. - **Helps most directly:** `CIN@LAD`, where the supplied forensic assessment explicitly says the venue weakness was underweighted. - **Potential harms:** Other games where venue splits are noisy or insufficiently sampled. - **Evidence strength:** Moderate and game-specific; do not promote from this slate alone. ## 6. No direct change to `context.opponent_strength_weight` The cross-slate evidence says the positive starter ERA bias is not separable from noise: - Mean +0.281 - z = 0.95 - Status: hold This lever should remain unchanged pending more evidence. --- # 9. Bottom line ## Known - The system went **6–5**, below the supplied walk-forward baseline. - Winner accuracy was materially better than score accuracy. - Two games had correct winners but incorrect scores: `ATL@PHI` and `WAS@SD`; other correct winners also had substantial run errors. - Three of four high-conviction calls lost: `STL@SF`, `TOR@OAK`, and one of the four strong/upper-confidence selections as counted in the supplied high-conviction set. - The worst individual miss was `TOR@OAK`, with published probability 0.699558 and Brier score 0.489381. - Run-model holdout signal is near zero and has persisted across 11 slates. - Simulation winner-direction intervention has been inert for 11 consecutive slates. - The enhanced-prediction transition introduced documented wrong-direction regressions in `MIN@DET` and `STL@SF`. - Four probability publications have unexplained material deltas. - One starter record in ARI has an identity/missing-line inconsistency. - Structural game coverage was complete, with zero reported critical/error/warning findings in the structural audit layer. ## Uncertain - Whether the four probability deltas represent an actual governor bypass, a side-orientation transformation, or an undocumented but valid post-processing step. - Whether the observed high-confidence failures reflect persistent calibration bias or a bad single slate; the upper-tail result is concerning, but only one current consecutive slate is recorded for calibration instability. - Whether pitcher-form or venue adjustments will improve future results rather than simply react to outliers. - How much of the run error is model deficiency versus irreducible player and starter variance. - Whether the missing integrity-revision artifact contains additional remediation evidence. ## Safest next action **Repair and independently validate the simulation probability path before promoting any model or confidence change.** The simulation inertness is persistent, has a concrete code-level explanation, and can be tested without committing to a risky one-game or one-slate parameter adjustment. In parallel, keep confidence-cap and score-variance changes in shadow mode only, with Brier score, calibration error, winner accuracy, team-run MAE, and total-run MAE monitored for rollback.