# Luna deep analysis — slate 20261008 > Coverage receipt: **COMPLETE** · 9/8 sections; 1/1 games; 5/5 causes; 7/7 sources ## 1. Evidence receipt and gaps **All seven `source_manifest` entries are present and marked `loaded`:** | Source | Bytes | Purpose | |---|---:|---| | prediction review — `/app/output/reports/prediction_review_20261008.json` | 17,218 | Official outcomes and scoring errors | | forensic deep audit — `/app/output/reports/PREDICTION_SYSTEM_DEEP_AUDIT_20261008.json` | 45,923 | Probability, provenance, starter and decision contracts | | continuous learning — `/app/output/learning/continuous_learning_daily_20261008.json` | 245,469 | Pipeline stages, calibration, probability lineage, score distribution and starter comparisons | | detailed report — `/app/output/reports/detailed_20261008.html` | 58,372 | Game and simulation rendering | | ultra detailed player audit — `/app/output/reports/20261008_ultra_detailed_player_audit.html` | 3,404,020 | Player-level prediction trace | | structural audit — `/app/output/reports/audit_report_20261008.html` | 244,042 | Invariant findings | | integrity revision audit — `/app/output/reports/audit_report_20261008_integrity_revision.html` | 126,889 | Structural audit of post-start revisions | **Game-level review:** one game is supplied in `games[]`: **CLE@CHW**. Its official result is CLE 9, CHW 5. The game record says `games_reviewed: 1`, `hits: 1`, `misses: 0`; the continuous-learning summary says one settled game, also one hit and zero misses. Thus the available game-level result is internally consistent. **Coverage-count contradiction:** the same evidence reports `expected_games: 1`, but `canonical_artifacts_loaded: 2`, `continuous_learning_games: 2`, `forensic_deep_audit_games: 2`, `structural_audits_completed: 2`, and `forensic_deep_audit_games: 2`. The causal register describes this as “2/1 games structurally audited.” These counts do not reconcile to the single supplied game record. The pipeline transitions also name **TB@NYY**, which is absent from `games[]` and has no supplied score, probability, or game-level audit here. I report that transition as evidence of a coverage/provenance gap, not as a second game that can be independently evaluated. There are no listed starter comparisons for a second game, and the only player-level trace supplied is for CLE@CHW. I do not infer any missing TB@NYY results or player outcomes. ## 2. Executive diagnosis ### Winner accuracy - The one supplied game was a **winner hit**: CLE was predicted to win and won 9–5. - Published away-win probability was **0.505**; this was a marginal pick, not evidence of a strong winner forecast. The confidence edge was only **0.01**, and conviction was `toss_up`. - Slate winner accuracy is reported as 1.0, but the settled sample is one. It must not be treated as evidence of reliable performance. ### Calibration quality - The game Brier score is **0.245025**, corresponding to the published 0.505 probability and a hit. - The calibration monitor has one settled observation in the 0.50–0.55 bucket: confidence 0.505, observed hit rate 1.0, signed calibration error **−0.495** under the stated “confidence minus hit rate” definition—i.e., underconfidence for this one observation. - Weighted absolute calibration error is reported as **0.495**. This is a large one-observation error, but the direction on the only bucket is underconfident. That conflicts with the self-assessment’s label `confidence_overconfident` and its “lower confidence caps” proposal. It also warrants caution with the cross-slate escalation’s prescribed increase in empirical shrinkage: the escalation is established as a recurring finding, but the direction of any adjustment is not validated by this single hit. - The supplied cross-slate record nonetheless reports six consecutive slates of `single_slate_calibration_instability`; its persistent status is actionable, while the present one-game calibration reading is not enough to select a production change. ### Score accuracy - Prediction: **CLE 5, CHW 4**, total 9, away margin 1. - Actual: **CLE 9, CHW 5**, total 14, away margin 4. - Winner was correct, but both the exact score and the run environment were wrong: total runs were under-projected by **5**, CLE runs by **4**, CHW runs by **1**. The listed team-run absolute error is 5; the stage-quality team-run MAE for this one game is 2.5 (the mean of the two team errors). - Predicted margin was 1 versus actual margin 4, an absolute margin error of 3. The evidence records signed margin error −3 in an away-margin convention and +3 in a home-minus-away convention; these are consistent under opposite sign conventions. - The single observed margin is compressed relative to the result (predicted margin ≤2; actual margin ≥4). It supports monitoring, not a general widening based on one blowout. ### Starter accuracy - CLE starter Parker Messick: projected **3 ER, 15 outs, 5.40 ERA**; recorded **2 ER, 7 outs, 7.71 ERA**. ER error is −1 (actual below projection), ERA error is +2.31 (actual ERA above projection), and the outs projection differs by 8. These measures point in different directions; they should not be reduced to one claim that the starter projection was simply “accurate.” - CHW starter Hagen Smith: projected **0 ER, 8 outs, 0.00 ERA**; recorded **0 ER, 2 outs, 0.00 ERA** in the starter-lines object. The settled per-player feed says there is **no pitching line** for him and marks the line missing. The zero-ERA comparison is therefore not independently supported by that feed; outs differ by 6. - The player-forensics summary attributes a 7.71 ERA to the CLE starter and reports the CHW starter’s line missing. Overall starter evidence is incomplete and should not be used to claim both starts were accurately forecast. ### Artifact integrity - The per-game forecast-file integrity observation says **clean: 0 defects, 1 informational note**; the game’s structural audit is `PASS`, with 0 critical, 0 error, 0 warning and 1 info. - No supplied finding identifies a corrupted saved score artifact as the cause of the game result. The prediction missed the final run totals despite a structurally clean game artifact. - However, probability provenance contains a raw-probability mismatch and the overall coverage counts do not reconcile. Those are auditability/contract concerns even though the supplied game artifact itself is marked clean. ### Causal issue register — every entry reconciled 1. **`score_margin_compression` — high, score distribution, CLE@CHW.** Evidence: 1/1 predicted margins ≤2; 1/1 actual margins ≥4; mean predicted margin 1.00 versus actual 4.00. Action: check tail coverage and margin MAE across recent out-of-sample slates before widening. Guardrail: a single blowout does **not** justify widening every score prediction. 2. **`single_slate_calibration_instability` — high, calibration, CLE@CHW.** Evidence: weighted absolute calibration error 0.4950 across one game. Action: watch confidence buckets over the rolling window before changing calibration weight. Guardrail: one slate is far too small to establish a stable calibration curve. 3. **`weak_run_model_signal` — high, model validation, CLE@CHW.** Evidence: home/away/total run-model R² = **0.0021/0.0309/0.0024**. Action: treat exact runs as low-confidence; investigate features/target rather than tuning score tails. Guardrail: small positive R² only means beating a mean-only reference; it does not make exact scores reliable. 4. **`complete_slate_audit_coverage` — info, coverage, CLE@CHW.** Evidence says “2/1 games structurally audited.” Action: retain a per-game audit receipt in each report. Guardrail: full audit coverage does not cancel game-level findings. The stated coverage count is internally inconsistent with the one supplied game record and must not be taken as proof that the unprovided game was reviewed here. 5. **`simulation_inert_on_slate` — info, simulation, CLE@CHW.** Evidence: 0 flip candidates and 0 published flips. Action: measure simulation contribution across rolling slates, separating score realization from winner flipping. Guardrail: zero winner flips is not itself a defect and does not mean no score-reconciliation work occurred. The cross-slate evidence separately escalates the recurring simulation issue, discussed below. ## 3. Every game ### CLE@CHW — winner hit, score miss **Prediction and result** - Predicted: **CLE 5, CHW 4; CLE winner**. - Actual: **CLE 9, CHW 5; CLE winner**. - **Correct-winner/wrong-score game:** yes. - Published home-win probability **0.495**, away/pick probability **0.505**. Home-minus-away edge **−0.01**; the system’s absolute edge is **0.01**, which is described as `toss_up`. The diagnostic correctly cautions that an edge of 0.01 is not “the half-scale distance from 50%.” - Conviction bucket: **toss_up**. No high-conviction pick is present. **Errors** - Brier: **0.245025**. - CLE team-run error: **4**; CHW team-run error: **1**; listed combined team-run absolute error: **5**; stage-quality team-run MAE: **2.5**. - Total-run error/bias: predicted 9 versus actual 14, so **−5** (under-projection). - Margin: predicted away margin 1; actual away margin 4; absolute margin error **3**. **Starter comparison** - CLE—Parker Messick, identity match: - Projected: 3 ER, 15 outs, 5.40 ERA. - Actual: 2 ER, 7 outs, 7.71 ERA. - ER error −1; ERA error +2.31; outs short of projection by 8. - CHW—Hagen Smith, identity match in the starter-lines object: - Projected: 0 ER, 8 outs, 0.00 ERA. - Recorded there: 0 ER, 2 outs, 0.00 ERA. - The player feed explicitly has no pitching line for Smith, so the zero-ERA comparison is not independently established. Outs were 6 fewer than projected. - The matchup breakdown identifies Smith’s settled-feed line as missing, with reason “no pitching line for this player in the settled per-player feed.” Do not interpret a projected/recorded 0.00 ERA as proof the projection was well supported. **Pregame edges, confidence and controls** - Season home-minus-away edge: **+0.1605**. - Recent home-minus-away edge: **+0.030**. - Pitcher home-minus-away edge: **−0.0005**. - Composite home edge: **−0.0853**. - Season quality edge: **+0.2685**; recent quality edge: **−0.04**. Season-quality margin adjustment was 0 because the quality edge was below the action threshold. - Raw probability in `model_trace`: **0.526027**; calibration sample size **571**, empirical hit rate **0.5499**; governed pick probability **0.505**. Governor was applied. - Recency guardrail was marked applied, but discount factor and reason are null. Do not infer a numerical recency discount. - Starter-experience shrinkage was not applied; minimum starter games 21 and shrink factor 1.0. - Score-coherence probability and published pick probability are both **0.505**. - A separate `probability_lineage` record says raw model pick probability was **0.500**, rather than 0.526027. That discrepancy is a provenance issue, not a basis for choosing whichever raw value better explains the final pick. **Score path** - Historical ML: winner accuracy 0/1; margin MAE 5.429; team-run MAE 2.7145; total-run MAE 5.1472. - Variance-applied: winner accuracy 0/1; margin MAE 5.000; team-run MAE 2.500; total-run MAE 5.000. Compared with historical ML, score errors improved, but winner accuracy did not. - Enhanced prediction: winner accuracy 1/1; margin MAE 3.000; team-run MAE 2.500; total-run MAE 5.000. - KPI, embedded Savant run environment, fundamental-edge target review, capped target, simulation result, and published final have the same listed metrics as enhanced prediction. No subsequent stage changed the measured result. - The transition table says variance → enhanced changed CLE from **CHW to CLE** and improved CLE@CHW; this is the provided evidence that the enhanced stage corrected the winner direction. It also names TB@NYY, which cannot be reconciled to the supplied game list. **Important player-level evidence** - CLE’s offense was under-projected by 4 runs. The player feed reports five material deviations, of which three are labeled learnable from the hitter’s own contact quality: - Angel Martínez: 0-for-3, 0 HR, 0 RBI against projections of 1.0 H and 1.0 HR; run impact **−1.73**; xwOBA **.299**, below average. Classified as model overrating a weak hitter. - José Ramírez: 2-for-5, 1 HR, 3 RBI against projections of 0.0 H and 0.0 HR; impact **+1.54**; xwOBA **.325**, above average. Classified as model underrating a strong hitter. - Steven Kwan: 1-for-3, 0 HR, 0 RBI against projections of 0.0 H and 0.0 HR; impact **+1.13**; xwOBA **.328**, above average. Also classified as model underrating a strong hitter. - Nathaniel Lowe: 2-for-4, no HR/RBI against projections of 2.0 H and 1.0 HR; impact **−1.55**. Listed as a top mover, not among the three deviations supported by the stated contact-quality assessment. - Travis Bazzana: 1-for-3, 1 RBI against projections of 0.0 H and 0.0 HR; impact **+0.78**. Also listed as a top mover, not among those three learnable cases. - The remaining named contributors have smaller listed impacts: Jo Adell +0.33, Austin Hedges −0.33, Brayan Rocchio +0.33, and Chase DeLauter −0.14. Their supplied lines are not a basis for inventing further causal claims. - Slate hitter totals also show projection weaknesses: AB 34 vs 30 actual (+13.3%); H 6 vs 8 (−25%); BB 5 vs 7 (−28.6%); SO 8 vs 2 (+300%); RBI 5 vs 7 (−28.6%); R 5 vs 7 (−28.6%); HR 2 vs 1 (+100%); 2B 0 vs 1 (−100%). For H, BB, SO, RBI, R and HR, a flat slate-mean hitter projection had lower reported MAE than the per-hitter projection. This is a one-slate diagnostic, not a stable player-model conclusion. **Audit and cause** - Structural audit: **PASS**, risk score 0; 108 good, 1 info, no critical/error/warning findings. Forecast-file integrity is reported clean, with one informational note. - No primary or secondary driver is assigned (`null`); player-forensics verdict counts assign two deviations to variance and none to offense or pitching as a side-level verdict. - This was a winner hit, but the CLE offense and total were materially under-projected. A correct winner does not make the score prediction good. ## 4. Miss reconstruction **Game losses:** none in the supplied game list. Thus there is no losing pick for which to assign a wrong-winner cause. CLE@CHW is a winner hit and a score miss, not a winner loss. The unexplained TB@NYY transition cannot be reconstructed as a game result from this evidence. **For the score miss in CLE@CHW, the earliest relevant evidence and distinctions are:** - **Bad input:** no direct evidence establishes a bad team-score input. The feed does document a missing CHW starter line, and the raw probability values conflict (see Section 6); neither proves that an input caused the score miss. - **Bad weighting / run signal:** the offense and total were under-projected, and several CLE hitter projections missed. The validated run models have near-zero holdout R². That is stronger evidence of weak run-prediction signal than of any specific one-game weighting mistake. - **Calibration:** the confidence probability was nearly even and the winner happened to hit. Calibration is a recurring cross-slate issue, but it is not the demonstrated cause of the run-total shortfall. - **Starter error:** Messick allowed fewer ER than projected but had a higher ERA and substantially fewer outs than projected. Smith’s settled per-player line is missing. The evidence does not isolate starter error as the main cause of CLE’s four-run excess. - **Offense/bullpen miss:** CLE’s actual nine runs exceeded its projected five by four. The supplied diagnostics identify hitter deviations, but side-level player-forensics assigns the game closer to variance, not an offense verdict. Bullpen detail is insufficient to apportion the score error: the feed reports CLE bullpen ER 1 and CHW bullpen ER 0, but does not establish that bullpen performance caused the miss. - **Run-environment miss:** total runs were under-projected by five; this is directly established. It is consistent with weak run-model signal and the supplied player deviations, but not enough to identify a unique causal lever. ### Pipeline-stage quality and transitions All listed stages have one game of measurement; interpret the metrics as one-game observations. | Stage | Winner hits / accuracy | Margin MAE | Team-run MAE | Total-run MAE | |---|---:|---:|---:|---:| | historical_ml | 0 / 0.0 | 5.429 | 2.7145 | 5.1472 | | variance_applied | 0 / 0.0 | 5.000 | 2.500 | 5.000 | | enhanced_prediction | 1 / 1.0 | 3.000 | 2.500 | 5.000 | | kpi_layer | 1 / 1.0 | 3.000 | 2.500 | 5.000 | | embedded_savant_run_environment | 1 / 1.0 | 3.000 | 2.500 | 5.000 | | fundamental_edge_target_review | 1 / 1.0 | 3.000 | 2.500 | 5.000 | | capped_target | 1 / 1.0 | 3.000 | 2.500 | 5.000 | | simulation_result | 1 / 1.0 | 3.000 | 2.500 | 5.000 | | published_final | 1 / 1.0 | 3.000 | 2.500 | 5.000 | - Historical ML → variance improved margin MAE by 0.429 and team-run MAE by 0.2145, with no winner-accuracy change. - Variance → enhanced improved margin MAE by 2.0 and changed CLE from CHW to CLE, improving this game’s winner direction. The transition also contains the unprovided TB@NYY direction change. - Enhanced → KPI → embedded run environment → fundamental-edge review → capped target → simulation → published final shows no listed winner-direction or score-metric change. - The table therefore records an early wrong winner direction that was **recovered at enhanced prediction**. A later recovery does not erase the initial wrong direction. The transition’s extra TB@NYY reference remains an audit gap. - `largest_team_run_regression_vs_base` is labeled variance-applied, but its deltas versus historical ML are negative (improvements), not regressions. The label and numeric deltas are internally inconsistent; the numeric values show improvement on this one game. ## 5. Cross-game counterfactuals The supplied `lever_tension` contains three entries; with only one game, its listed benefits and costs are limited to that game. 1. **`score.variance_multiplier / score.variance_floor` — widen run distributions** - Helps: none listed. - Hurts: none listed. - Net listed evidence: 0 helped, 0 hurt. The target field lists Brier score and calibration error, although the causal margin-compression finding is about score-distribution tails. - **Decision:** no production-wide widening is supported by this one result. The actual margin exceeded the predicted margin, but the register explicitly cautions against widening every score after one blowout. A bounded shadow test can be considered because margin compression recurs in the cross-slate streaks. 2. **`player.pitcher.v4` form weights / `player.pitcher_prior_games`** - Helps: none. - Hurts: CLE@CHW—listed as a hit where the starter projection was “accurate (ER off 1),” so reweighting could disturb it. - Net: 0 helped versus 1 hurt. - **Decision:** do not change pitcher weights on this slate. The projection was not wholly accurate: ERA and outs differ materially, and Smith’s per-player line is missing. That limits the strength of the “accurate” characterization, but still gives no evidence for a broad pitcher-weight change. 3. **`recent.* / season.*` run-mean weighting** - Helps: CLE@CHW, whose team runs were off by 5. - Hurts: none listed. - Net: 1 helped, 0 hurt. - **Decision:** this is a plausible diagnostic lead, not a validated adjustment. The supplied lever name is a wildcard rather than an exact parameter, and the present slate is one game. Do not pick an arbitrary underlying parameter value. **Tempting one-game fixes to avoid:** widening every score after the 9–5 result; changing pitcher-form weights based on this single starter comparison; fitting individual hitter weights to the deviations of Lowe or Bazzana without supporting cross-slate evidence; and treating the 0.505 winner probability as proof that confidence should be lowered or raised. Each would overfit sparse evidence or conflict with the supplied lever-tension/guardrail evidence. ## 6. Data and contract defects ### Findings by reported audit severity - **Critical:** none reported in the supplied structural audit summaries. - **Error:** none reported. - **Warning:** none reported. - **Info:** one informational finding in the CLE@CHW structural audit (`good: 108, info: 1`), and one informational note in the forecast-file integrity summary. Their detailed text is not supplied, so they cannot be named more specifically. The causal register separately records two informational findings: - `complete_slate_audit_coverage`, with evidence “2/1 games structurally audited,” action to retain per-game receipts, and guardrail that coverage does not cancel game-level findings. - `simulation_inert_on_slate`, with 0 flip candidates and 0 published flips; its guardrail is that zero winner flips is not, by itself, a defect. - No per-game audit finding is listed as critical, error, or warning. The count inconsistency and probability mismatch below are identifiable data/provenance defects, but the JSON does not assign them one of those formal severity labels. ### Coverage reconciliation The matrix says the expected game count is 1 while several loaded/audited counts are 2. The supplied game array and `games_reviewed` are both 1. The “2/1” causal-issue evidence therefore cannot be reconciled from this payload. The extra TB@NYY pipeline transition reinforces, but does not resolve, the discrepancy. Do not call coverage clean solely because `all_games_structurally_audited` is true. ### Probability-lineage audit for the pick side The published side was away/CLE at **0.505**. 1. **Raw model:** `model_trace` reports raw pick probability **0.526027**. `probability_lineage` instead reports raw model pick probability **0.500**. These values conflict; the payload does not explain whether they have different definitions. 2. **Calibrated:** probability lineage reports calibrated pick probability **0.505**, calibration sample size 571, empirical hit rate 0.5499. 3. **Governor:** lineage reports governor input **0.526027** and output **0.505**. The model trace also says governor applied and governed pick probability 0.505. 4. **Score coherence:** reported as **0.505**. 5. **Recency:** guardrail applied, but discount factor is null. No numerical recency transformation can be verified or inferred. 6. **Published:** pick probability **0.505**, equal to calibrated, governed and score-coherence values; `published_minus_calibrated` is 0 and the lineage marks the published probability transformation as matching the calibrated/governed value. The lineage does **not** label an unexplained material delta between calibrated/governed probability and published probability. However, the 0.500 versus 0.526027 raw-value conflict is unresolved, and the displayed raw-to-governor-input path has no explicit transformation description in the lineage. This is an observability/contract gap; the JSON does not establish that the final published probability bypassed a recorded calibration or governor transformation. A separate `probability_reversions` record reports trained home-win probability **0.653438**, blended home probability **0.575938**, a blend-stage side change to home, and a published away side at home-win probability 0.495; it says the blend side did not survive publication (`overwritten: 1`). That material event is recorded in the separate reversions block, but it is not represented as a step in the pick-side `probability_lineage`. Treat this as an incomplete cross-block trace, not proof of an unrecorded bypass. **Model/process defects versus presentation/artifact concerns** - The score and starter misses are prediction-quality issues, not reported file corruption. - Near-zero run-model R² and the hitter projection diagnostics are model-validation concerns. - The missing Hagen Smith settled-feed line is a data completeness limitation. - Conflicting raw probability values and unresolved game coverage are contract/provenance issues. - The forecast file is expressly reported clean; no presentation bug is established as the reason for this game’s pick or score. ## 7. Cross-slate persistence ### Required structural blocks - **`margin_distribution`:** one row, CLE@CHW; predicted margin 1, actual margin 4, signed margin error −3. One observation only. - **`pipeline_stage_quality.transitions`:** winner direction first remains wrong through variance-applied, then changes from CHW to CLE at enhanced prediction; later stages show no change. The TB@NYY references are not corroborated by the game list. - **`calibration_monitoring`:** one settled observation, confidence 0.505, actual hit rate 1.0, signed error −0.495 (underconfident for the observed bucket), weighted absolute error 0.495. - **`high_raw_cluster`:** 0 games, 0 hits, threshold 0.65. There are no high-raw picks in this block. - **`simulation_control`:** 2 observed games, 0 flip candidates, 0 published flips; status `diagnostic_only_without_independent_support`. This is not evidence that a winner flip should have occurred. The cross-slate streak, however, makes the repeated lack of useful simulation signal an established diagnostic concern. ### Finding streaks — every entry - `pipeline_winner_instability`: **0 consecutive slates**, not present now; 3 incidences in the 11-slate window; no target lever. - `score_margin_compression`: **1 consecutive slate**, present now; 4 incidences in the window; target levers `score.variance_multiplier`, `score.variance_floor`. - `simulation_inert_on_slate`: **11 consecutive slates**, present now; 11 incidences; target list includes `simulation.allow_score`, but `param_actionable` is false. - `single_slate_calibration_instability`: **6 consecutive slates**, present now; 9 incidences; target levers `confidence.empirical_max_weight`, `confidence.base_cap`, `confidence.moderate_cap`. - `weak_run_model_signal`: **6 consecutive slates**, present now; 9 incidences; target list includes training XGBoost parameters, but `param_actionable` is false. ### Escalated findings — established failures, not single-slate noise 1. **Simulation inertness, 11 consecutive slates.** Target lever listed: `simulation.allow_score`; target metrics: team-run MAE and total-run MAE; stated direction: confirm the simulation module is running rather than silently no-oping. This has a concrete code/diagnostic action, not a supported parameter adjustment. The self-assessment identifies a likely structural cause: the same draws are used both to search for a target-realizing play-by-play and estimate the win distribution; early exit can leave only about two selected draws. It recommends an independent, unsteered distribution batch that does not early-exit, feeding the resulting win rate to the distribution-agreement policy. This root-cause description is supplied as a self-assessment, not independently demonstrated by the game receipt. 2. **Calibration instability, 6 consecutive slates.** Target levers: `confidence.empirical_max_weight`, `confidence.base_cap`, `confidence.moderate_cap`; metrics: Brier and calibration error; recommended direction: increase empirical shrinkage. Its persistence warrants a bounded shadow experiment. The current one-game bucket is underconfident, while the self-assessment proposes lower caps; this directional conflict argues for shadow comparison and monitoring rather than automatic application. 3. **Weak run-model signal, 6 consecutive slates.** Target levers are listed as `training.xgb.n_estimators`, `training.xgb.max_depth`, and `training.xgb.learning_rate`, but `param_actionable` is false. Target metrics: team-run and total-run MAE. The stated action is to investigate features/target by isolated training backtest, not to force a nightly parameter change. Holdout R² is 0.002415 total, 0.002143 home, 0.030936 away. ### Trend directions, stated explicitly - `margin_bias`: **regressing**, mean −1.7778. - `mean_team_run_error_on_losses`: **improving**, mean 2.375. - `weighted_calibration_error`: **regressing**, mean 0.429. - `winner_accuracy`: **improving**, mean 0.7778. ### Top-confidence-pick loss streak - Reported `loss_streak`: **0**. - Listed outcomes: 20260929 hit; 20260930 hit; 20261001 hit; 20261003 hit; 20261004 hit; 20261005 miss; 20261006 hit; 20261007 miss; 20261008 hit. Two entries have blank dates and `unknown`; those cannot be counted as hits or misses from this record. - Current top pick on 20261008 was a hit, but this short sequence includes two recent misses and does not imply robust high-confidence calibration. ### Cross-slate bias context The separate 10-slate bias record does **not** establish an actionable starter or lineup-run bias: - Starter ERA error mean −0.561, 6/10 same sign, z −1.67; status `hold`, not persistent. - Lineup-run delta mean −0.593, 7/10 same sign, z −1.89; status `hold`, not persistent. These are not grounds for changing `context.opponent_strength_weight` or `score.variance_floor` on their own. ## 8. Adjustments for today Recommendations below are bounded proposals or diagnostics; none is promoted or applied here. Persistent patterns receive priority over the one-game signal. 1. **Persistent calibration instability — shadow experiment.** - **Parameter:** `confidence.empirical_max_weight`. - **Direction/magnitude:** increase by **0.02 absolute from its current registered value** in a shadow arm only; do not change the base or moderate caps in this same experiment. - **Expected metrics:** test for lower rolling weighted calibration error and Brier score. These are hypotheses to measure, not guaranteed improvements; the current one-game bucket is underconfident, so also check for deterioration on correctly calibrated low-confidence picks. - **Games helped/hurt:** no specific game-level helps/hurts are recorded for this lever in `lever_tension`; none can be named from this slate. CLE@CHW was a 0.505-probability hit, so blindly reducing its confidence could hurt calibration for this one observation. - **Why this ranks highly:** six consecutive slates and nine window incidences, with calibration-error regression in the trend summary. Keep it shadow-only because present-slate direction and self-assessment direction conflict. 2. **Persistent simulation inertness — code/diagnostic action, not a parameter move.** - **Target named in the finding:** `simulation.allow_score`; **do not force a change to it**, because the finding is explicitly `param_actionable: false`. - **Action:** operator/code review of the simulation draw loop and `_build_simulation_distribution`; verify whether early exit and target-realization selection are contaminating the win-rate sample. Compare against an independent unsteered, non-early-exiting probability batch as proposed in the supplied self-assessment. - **Expected metrics:** a non-vacuous simulation contribution, evaluated against team-run and total-run MAE and winner-direction intervention over subsequent slates. No game-level levers are identified in the tension list for this code path. - **Why this ranks highly:** 11 consecutive slates. The next step is verification and measurement, not asserting that a winner flip should have happened. 3. **Persistent near-zero run-model signal — isolated feature/target backtest, not parameter tuning.** - **Target list:** `training.xgb.n_estimators`, `training.xgb.max_depth`, `training.xgb.learning_rate`; the finding marks this non-actionable by parameter. - **Action:** follow the supplied operator recommendation: investigate the shared feature contract and run an isolated same-fingerprint training/backtest; consider source-labelled no-vig market prices when available, park/weather, bullpen rest/fatigue and platoon-adjusted lineup quality. Bump the model schema version if feature semantics change. - **Expected metrics:** improved holdout run-model R² and out-of-sample team-run/total-run MAE, measured before any deployment. No specific `lever_tension` games map to training changes. - **Why this ranks highly:** six consecutive slates plus the supplied holdout R² values. Nightly variance tuning cannot repair missing predictive signal. 4. **Strongest single-slate driver: compressed margin and under-projected run total — bounded shadow only.** - **Parameter:** `score.variance_multiplier` (a named target of `score_margin_compression`; do not simultaneously alter `score.variance_floor`). - **Direction/magnitude:** test a **+0.05 absolute** change from the current registered multiplier in a shadow arm only. - **Expected metrics:** improve rolling margin tail coverage and margin MAE; monitor team-run and total-run MAE for regression. One game cannot establish the expected direction or size of a real improvement. - **Games helped/hurt:** current `lever_tension` lists no helped or hurt games for score variance. The only directly relevant game is CLE@CHW, where predicted margin 1 and actual margin 4 favor a wider tail; the causal register explicitly warns not to widen every score from one blowout. No other game is supplied to assess collateral effects. - **Why lower-ranked:** the single-slate compression observation is real, and margin bias is trending regressing, but the single game is too weak for a production change. Use a shadow comparison with a rollback monitor; do not promote on this result. No production change to pitcher weights or a wildcard `recent.* / season.*` parameter is recommended: the supplied tension gives no net cross-game evidence, and the latter does not name an exact parameter. The self-assessment’s general “lower caps and widen run variance” recommendation is not taken as a direct production instruction because it conflicts with the current observed calibration direction and the one-game score guardrail. ## 9. Bottom line **Known:** CLE@CHW was a correct-winner, wrong-score prediction. The away win probability was 0.505; the pick hit, but CLE was under-projected by four runs and the total by five. Exact run models have near-zero reported holdout R². The game artifact passed its supplied structural checks, while the starter-feed line for Hagen Smith is missing. **Uncertain or defective:** only one supplied game can be fully reviewed despite multiple reported coverage counts and a TB@NYY transition. Raw pick probability is reported as both 0.500 and 0.526027. Calibration’s current single observation is underconfident, despite a persistent calibration-instability finding and a conflicting self-assessment recommendation. The magnitude and causes of the score miss cannot be cleanly separated among run-model signal, offense, starter usage and variance from this game alone. **Safest next action:** have the operator verify the simulation’s early-exit/draw-distribution path and measure an independent, unsteered win-probability distribution in shadow. This addresses the longest persistent failure without forcing a parameter move or claiming that the single game proves a particular score or winner adjustment.