# Luna deep analysis — slate 20260910

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

# 1. Evidence receipt and gaps

## Sources received

All seven `source_manifest` entries were checked against their declared status:

| Source | Status | Purpose | Receipt finding |
|---|---:|---|---|
| `prediction_review_20260910.json` | Loaded | Official outcomes and scoring errors | Present; 129,033 bytes |
| `PREDICTION_SYSTEM_DEEP_AUDIT_20260910.json` | Loaded | Probability, provenance, starter, and decision contracts | Present; 90,736 bytes |
| `continuous_learning_daily_20260910.json` | Loaded | Stage quality, calibration, lineage, score distribution, starters | Present; 866,957 bytes |
| `detailed_20260910.html` | Loaded | Full game and simulation rendering | Present; 122,840 bytes |
| `20260910_ultra_detailed_player_audit.html` | Loaded | Player-level prediction trace | Present; 11,020,085 bytes |
| `audit_report_20260910.html` | Loaded | Invariant findings | Present; 626,992 bytes |
| `audit_report_20260910_integrity_revision.html` | **Missing** | Post-remediation immutable revision proof | Missing; byte count `null` |

All five expected games were reviewed:

1. `COL@NYY`
2. `HOU@PHI`
3. `PIT@CHW`
4. `TB@ATL`
5. `TEX@SEA`

Coverage is complete for the five-game slate: canonical artifacts loaded `5/5`, continuous-learning games `5/5`, forensic deep-audit games `5/5`, starter comparisons `5/5`, and structural audits `5/5` (`coverage_matrix`).

## Gaps and contradictions

1. **Missing integrity revision audit.** The source manifest explicitly reports `audit_report_20260910_integrity_revision.html` as missing. This prevents confirmation of the post-Luna immutable-revision proof, although the per-game structural audits otherwise pass.

2. **Critical probability-lineage defect at PIT@CHW.** The causal register identifies an unexplained calibrated-to-published delta. The lineage records:
   - calibrated probability: `0.5711`
   - published home probability: `0.528284`
   - delta: `-0.042816`
   - `unexplained_material_delta: true`
   - transformation: `"no recorded transformation explains the publication delta"`
   
   The register correctly treats this as **critical** and says promotion must be blocked until the missing transformation is traced. The data does **not** establish that the governor was bypassed; it establishes only that the final publication transformation is not documented.

3. **Starter-feed inconsistency at PIT@CHW.** The settled per-player feed reports a missing pitching line for CHW starter Hagen Smith: projected ERA `1.93`, reason `"no pitching line for this player in the settled per-player feed"`. However, the game-level `starter_lines` still contains a projected line of `1.0 ER`, `1.93 ERA`, and `14 outs`, while the actual line records `0 ER` and `6 outs`. This is not counted as a structural audit error, but it is a starter-data contract gap requiring reconciliation.

4. **Probability-orientation complexity.** Some fields are expressed as home probability, while the governor and diagnostic fields appear to be pick-side oriented. For example, PIT@CHW has an away pick and published home probability `0.528284`, equivalent to an away pick probability of approximately `0.471716` in `probability_reversions`. The numerical values are reconcilable by orientation, but the lineage labels should explicitly identify side orientation at every step.

5. **Score-stage and winner-stage behavior are not identical.** The stage table shows `historical_ml` winner accuracy `0.6`, `variance_applied` winner accuracy `0.4`, and `enhanced_prediction` recovering to `0.6`. A later recovery does not erase the intermediate regression.

6. **Small-sample calibration.** The slate contains only five settled games. Calibration buckets contain three and two games respectively. The cross-slate persistence evidence is therefore more informative than the current slate alone for parameter decisions.

---

# 2. Executive diagnosis

## Overall result

The system went **3–2**, with winner accuracy `0.600`, versus the walk-forward baseline of `0.559444` and model holdout winner accuracy of `0.598152` (`summary`, `self_assessment`, `model_validation`). That is a modestly positive directional result, not evidence of good score forecasting.

### Separate diagnoses

| Dimension | Diagnosis | Evidence |
|---|---|---|
| Winner accuracy | Acceptable on this slate, but not dominant | `3/5`, `0.600`; two losses were low-confidence or moderate-confidence calls |
| Calibration quality | Mixed and materially imperfect | Weighted absolute calibration error `0.1171`; one bucket underconfident, one overconfident |
| Score accuracy | Weak | Team-run MAE at published stage `2.3`; total-run MAE `4.2`; exact-run holdout R² only `0.011791` home, `0.021038` away, `0.020666` total |
| Starter accuracy | Mixed and a major contributor to HOU@PHI | Nine starter comparisons showed mean ERA error `-3.19`; Cristian Javier was projected at `11.25 ERA` and delivered `1.50 ERA` |
| Artifact integrity | Structural artifacts largely passed, but not clean overall | All five per-game audits passed with zero critical/error/warning findings, but PIT@CHW has a critical undocumented probability transformation and the integrity revision audit is missing |

A correct winner does not imply a good prediction. PIT@CHW was a correct winner but had a team-run error of `5`; COL@NYY was a correct winner but missed the home score by `2` and had a team-run absolute error of `3`.

## Causal issue register reconciliation

### Critical: unexplained published probability transformation

- **ID/title:** `unexplained_probability_publication_delta` — “Published probability differs materially without a recorded transformation.”
- **Severity:** **Critical**
- **Affected game:** `PIT@CHW`
- **Evidence:** one game has an unexplained calibrated-to-published delta of at least two percentage points; actual lineage delta is `-0.042816`.
- **Action:** block promotion until the pick-side math is traced and the missing transformation is recorded.
- **Do-not-infer guardrail:** do not assume governor bypass until the pick-side math is traced.

This is a contract/evidence failure, not proof that the probability governor was bypassed.

### High: intermediate winner-direction regression

- **ID/title:** `pipeline_winner_instability` — “An intermediate stage materially degraded winner direction.”
- **Severity:** **High**
- **Affected game:** `TEX@SEA`
- **Evidence:** `historical_ml → variance_applied` changed winner accuracy from `0.6` to `0.4`; TEX@SEA changed `SEA → TIE`.
- **Action:** preserve before-stage and after-stage picks and list exact flips.
- **Do-not-infer guardrail:** later recovery does not make the unstable step harmless.

This is a material model/process risk, not a critical artifact failure.

### High: weak validated run-model signal

- **ID/title:** `weak_run_model_signal`
- **Severity:** **High**
- **Affected games:** all five games
- **Evidence:** holdout run R² is `0.0118/0.0210/0.0207` for home/away/total.
- **Action:** improve features or target definition through isolated training backtests; do not tune score tails as if exact scores were reliable.
- **Do-not-infer guardrail:** small positive R² does not establish exact-score reliability.

### Informational: complete audit coverage

- **Severity:** **Info**
- **Evidence:** `5/5` games structurally audited.
- **Action:** retain the per-game audit receipt.
- **Guardrail:** complete coverage does not cancel findings inside audited games.

### Informational: simulation inert on this slate

- **Severity:** **Info** on this slate, but escalated across slates
- **Evidence:** zero flip candidates and zero published flips.
- **Action:** measure simulation contribution over rolling slates, separating score realization from winner intervention.
- **Guardrail:** zero winner flips is not itself a defect and does not prove no score reconciliation occurred.

The cross-slate evidence changes the operational status of this finding from single-slate information to an established persistent code/diagnostic problem; see Sections 7 and 8.

---

# 3. Every game

## `COL@NYY` — correct winner, imperfect score

- **Predicted:** COL `4`, NYY `8`; winner NYY.
- **Actual:** COL `3`, NYY `10`; winner NYY.
- **Published pick probability:** NYY `0.580`.
- **System confidence edge:** home-minus-away `+0.16`; diagnostic labels this moderate, not a near coin flip.
- **Conviction:** Moderate.
- **Brier score:** `0.1764`.
- **Team-run absolute error:** `3`.
- **Total-run bias:** `-1` — total forecast `12` versus actual `13`.
- **Correct-winner/wrong-score:** **Yes.** The winner was correct, but the published score was not exact.

### Starter projection versus actual

| Starter | Projected | Actual | Error |
|---|---:|---:|---:|
| Ryan Feltner, COL | 4 ER, 14 outs, 7.71 ERA | 4 ER, 15 outs, 7.20 ERA | ER `0`, outs `+1`, ERA `-0.51` |
| Max Fried, NYY | 2 ER, 18 outs, 3.00 ERA | 1 ER, 10 outs, 2.70 ERA | ER `-1`, outs `-8`, ERA `-0.30` |

Starter projections were broadly accurate in earned runs. The largest score miss came from offense and game realization rather than a major starter ER miss.

### Edges and controls

- Season edge: NYY `+1.720`.
- Recent edge: NYY `+1.254`.
- Pitcher edge: NYY `+0.092`.
- Composite edge: NYY `+2.183`.
- Season quality edge: `+1.7195`; recent quality edge `+0.904`.
- Pitcher adjustment applied: away factor `0.9943`, home factor `1.0867`.
- Starter-experience shrinkage: not applied; both starters had established histories.
- Recency guardrail: applied.
- Score-volatility action: `supported_blowout_margin_expansion`.
- Confidence path reached `0.6858` before publication/recency controls; published probability was `0.580`.

### Score path and player adjustments

The supplied trace does not provide the full score-evolution array for this game, but it records a high-margin pick and a large Savant run-environment swing. Important player deviations included:

- Anthony Volpe: `3-for-4`, HR, 5 RBI versus projected 1 hit and 0 HR; run impact `+1.54` in the player audit and `+1.54`/`+1.54`-class contribution in the forensic records.
- Hunter Goodman: `0-for-5` versus projected hit/HR; run impact `-1.40`.
- Ezequiel Tovar: `2-for-3` versus a projected 0 hits; impact `+1.27`.
- Ben Rice: below projection; impact `-1.26`.
- Jake McCarthy: above projection; impact `+1.09`.

The player-level diagnostic says the saved forecast artifact was clean and was not responsible for a miss; this game was a hit.

### Audit and causes

- Structural audit: PASS; `0 critical`, `0 error`, `0 warning`, `101 good`, `8 info`.
- Primary cause: none assigned.
- Secondary cause: none assigned.
- Interpretation: strong directional call with acceptable starter projections, but exact scoring remained uncertain. The high-margin action helped recognize the seven-run actual margin, though it still underpredicted the margin by three runs.

---

## `HOU@PHI` — wrong winner and large score miss

- **Predicted:** HOU `4`, PHI `5`; winner PHI.
- **Actual:** HOU `2`, PHI `1`; winner HOU.
- **Published pick probability:** PHI `0.566859` (approximately `0.567`).
- **System confidence edge:** home-minus-away `+0.1337`; moderate, not near-coin-flip language.
- **Conviction:** Moderate.
- **Brier score:** `0.321329`.
- **Team-run absolute error:** `6`.
- **Total-run bias:** `+6` — forecast total `9`, actual total `3`.
- **Correct-winner/wrong-score:** No; both winner and score were wrong.

### Starter projection versus actual

| Starter | Projected | Actual | Error |
|---|---:|---:|---:|
| Cristian Javier, HOU | 5 ER, 12 outs, 11.25 ERA | 1 ER, 18 outs, 1.50 ERA | ER `-4`, outs `+6`, ERA `-9.75` |
| Zack Wheeler, PHI | 3 ER, 17 outs, 4.76 ERA | 1 ER, 21 outs, 1.29 ERA | ER `-2`, outs `+4`, ERA `-3.47` |

The starting-pitcher projection was the primary miss driver. Both starters substantially outperformed their projected ER lines, especially Javier.

### Edges and controls

- Season edge: PHI `+0.429`.
- Recent edge: PHI `+0.068`.
- Pitcher edge: PHI `+0.020`.
- Composite edge: PHI `+0.476`.
- Recent-form run differential: HOU `+5`, PHI `+11`; both teams had a `0.600` recent win percentage.
- Pitcher adjustments: HOU-side run factor `0.967`, PHI-side `0.9868`.
- Starter-experience shrinkage: not applied; both histories were treated as established.
- Recency guardrail: applied.
- Score-volatility action: `no_change`.
- Confidence path: governed pick probability `0.6031`; published PHI probability `0.566859`.
- The published commitment was only a `0.13` edge, so this was a moderate commitment rather than a high-certainty call.

### Score path

The forecast evolved as follows:

1. Historical ML: HOU `4.0108`, PHI `5.9834`.
2. Variance applied: HOU `4`, PHI `6`.
3. Enhanced prediction: HOU `4`, PHI `5`.
4. Savant run environment: HOU remained `4`, PHI `5`; adjustments were `-0.487` HOU and `+0.1663` PHI.
5. Fundamental review, cap, and simulation: HOU `4`, PHI `5`.

The model never moved toward the actual low-scoring HOU win. The `no_change` volatility decision left a nine-run forecast against a three-run game.

### Player-level adjustments

Material player changes were recorded for 24 players/metrics. Important deviations included:

- Cam Smith: `2-for-3`, HR, RBI versus projected 1 hit and 0 HR; impact `+1.73` for HOU.
- Yordan Alvarez: `0-for-3` versus projected hit/HR; impact `-1.40`.
- Kyle Schwarber: `0-for-3` versus projected hit/HR; impact `-1.40`.
- Daulton Varsho: `0-for-4` versus projected hit; impact `-0.80`.
- Bryce Harper: `0-for-3` versus projected hit; impact `-0.80`.

The slate player audit found only `11/37` material deviations were supported by the hitter's own season contact quality (`29.7%`), so most player-level deviations should not be overfit.

### Audit and causes

- Structural audit: PASS; `0 critical`, `0 error`, `0 warning`, `102 good`, `7 info`.
- **Primary cause:** `starter_projection`.
- **Secondary cause:** `run_environment_over`.
- Starter synthesis specifically identifies Javier's credible venue split as a possible underweighted signal: away ERA `9.947` versus `5.704` on the other side, despite his actual one-run performance.
- The artifact was structurally clean; this was a model-input/weighting and realization failure, not a presentation failure.

---

## `PIT@CHW` — correct winner, materially wrong score, critical lineage defect

- **Predicted:** PIT `4`, CHW `3`; winner PIT.
- **Actual:** PIT `2`, CHW `0`; winner PIT.
- **Published pick probability:** away PIT approximately `0.471716`; published home probability `0.528284`.
- **System confidence edge:** home-minus-away `-0.0566`, equivalent to a modest away lean; diagnostic rounds this to a `0.06` edge.
- **Conviction:** Lean.
- **Brier score:** `0.222516`.
- **Team-run absolute error:** `5`.
- **Total-run bias:** `+5` — forecast total `7`, actual total `2`.
- **Correct-winner/wrong-score:** **Yes.**

The edge is not a near coin flip in the sense of the report's explicit guardrail: a `0.12` home-away confidence edge is moderate, not a near coin flip. This game was lower conviction than HOU@PHI but still a directional lean.

### Starter projection versus actual

| Starter | Projected | Actual | Error |
|---|---:|---:|---:|
| Jared Jones, PIT | 3 ER, 14 outs, 5.79 ERA | 0 ER, 21 outs, 0.00 ERA | ER `-3`, outs `+7`, ERA `-5.79` |
| Hagen Smith, CHW | 1 ER, 14 outs, 1.93 ERA in game-level line | Actual starter feed line is incomplete; game-level actual records 0 ER and 6 outs | Feed/role reconciliation gap |

The game-level breakdown labels Hagen Smith's starter line as missing from the settled per-player feed. Therefore the CHW starter comparison is not fully contract-clean even though the structural audit passed.

### Edges and controls

- Season edge: PIT `-0.231` home-minus-away.
- Recent edge: PIT `-1.150`.
- Pitcher edge: home `+0.042`.
- Composite edge: PIT `-0.503`.
- Pitcher adjustment: away `1.000`, home `1.0417`.
- Starter-experience shrinkage: applied; probability was shrunk toward 0.5 because of an inexperienced/thin-history starter, with shrink factor `0.8917`; winner unchanged.
- Recency guardrail: not applied.
- Score-volatility action: `no_change`.
- Confidence path: raw model home probability `0.585524`; governed value `0.5711`; diagnostic says the pipeline reached `0.571` before controls, while published home probability was `0.528284`.

### Score path and players

The forecast path is not given as a full score-evolution array, but the published target was PIT `4`, CHW `3`.

Important player outcomes:

- Oneil Cruz: `3-for-4`, HR, 2 RBI versus projected 1 hit, 1 RBI; impact `+1.54`.
- Munetaka Murakami: `0-for-4` versus projected hit/HR; impact `-1.73`.
- Brandon Lowe: `0-for-4` versus projected hit/HR; impact `-1.40`.
- Bryan Reynolds: `0-for-4` versus projected hit; impact `-1.11`.
- Jared Triolo: above projection; impact `+0.80`.

Both lineups scored three runs below their respective projections. The forensic conclusion calls Jones's shutout an outlier start and says no credible venue, opponent, or experience signal explains it.

### Audit and causes

- Structural audit: PASS; `0 critical`, `0 error`, `0 warning`, `100 good`, `8 info`.
- Primary/secondary causal labels: none assigned.
- Forensic interpretation: correct winner, wrong score; starter overperformance and low run environment drove the score miss.
- **Critical artifact/contract issue:** undocumented calibrated-to-published probability transformation. The result being a hit does not reduce the severity of the lineage defect.

---

## `TB@ATL` — wrong winner, largest score miss

- **Predicted:** TB `6`, ATL `5`; winner TB.
- **Actual:** TB `1`, ATL `3`; winner ATL.
- **Published pick probability:** away TB approximately `0.505802`; published home probability `0.494198`.
- **System confidence edge:** home-minus-away `-0.0116`; diagnostic rounds to `0.01`.
- **Conviction:** Toss-up.
- **Brier score:** `0.255836`.
- **Team-run absolute error:** `7`.
- **Total-run bias:** `+7` — forecast total `11`, actual total `4`.
- **Correct-winner/wrong-score:** No.

### Starter projection versus actual

| Starter | Projected | Actual | Error |
|---|---:|---:|---:|
| Nick Martinez, TB | 2 ER, 18 outs, 3.00 ERA | 0 ER, 19 outs, 0.00 ERA | ER `-2`, outs `+1`, ERA `-3.00` |
| Martín Pérez, ATL | 2 ER, 17 outs, 3.18 ERA | 0 ER, 21 outs, 0.00 ERA | ER `-2`, outs `+4`, ERA `-3.18` |

Both starters outperformed their ER projections, but the game was primarily lost through the extreme low-scoring environment and a bullpen failure.

### Edges and controls

- Season edge: ATL `+0.058`.
- Recent edge: TB `-0.766`.
- Pitcher edge: TB `-0.025`.
- Composite edge: TB `-0.215`.
- Recent form: TB had `+5` run differential and 0.600 win percentage; ATL had `-8` and 0.400.
- Head-to-head inputs favored TB: TB won the two previous games by `7–2` and `7–1`.
- Pitcher adjustments: TB `1.0156`, ATL `0.9904`.
- Starter-experience shrinkage: not applied.
- Recency guardrail: applied.
- Score-volatility action: `no_change`, despite a low-score probability of `0.7515`.
- Confidence path: governed home probability `0.508`; published home probability `0.494198`.

### Score path

1. Historical ML: TB `4.217`, ATL `4.677`; slight ATL direction.
2. Variance applied: TB `4`, ATL `5`; ATL direction.
3. Enhanced prediction: TB `5`, ATL `4`; **first wrong-direction flip toward TB**.
4. Savant environment: TB `5`, ATL `5`; tie.
5. Fundamental review: TB `6`, ATL `5`.
6. Capped and simulation stages: TB `6`, ATL `5`.

The enhanced stage moved away from the actual ATL winner. The later stages preserved or reinforced the wrong TB direction. No simulation winner intervention corrected it.

### Player-level adjustments

- Austin Riley: positive projection adjustment, largest RBI change `+2.51`.
- Lane Thomas: material adjustment, largest hit change `+2.48`.
- Junior Caminero: below projection, impact `-1.40`.
- Nick Madrigal: below projection, impact `-0.80`.
- Ronald Acuña Jr.: below projection, impact `-0.80`.
- Ha-Seong Kim produced `2-for-3` with 2 RBI against a zero-hit projection, impact `+0.94`.

The actual TB bullpen allowed `3 ER` in one out through Kevin Kelly, while ATL's bullpen allowed one run. The lineups also underperformed: TB lineup delta `-4`, ATL `-3`.

### Audit and causes

- Structural audit: PASS; `0 critical`, `0 error`, `0 warning`, `101 good`, `7 info`.
- **Primary cause:** `run_environment_over`.
- **Secondary cause:** `toss_up_variance`.
- Starter synthesis says both starter shutouts were outlier performances and does not warrant a starter-model change for this individual game.
- The low-confidence/toss-up classification was appropriate; the loss should not be treated as evidence that a strong directional edge failed.

---

## `TEX@SEA` — correct winner with intermediate winner regression

- **Predicted:** TEX `4`, SEA `5`; winner SEA.
- **Actual:** TEX `3`, SEA `4`; winner SEA.
- **Published pick probability:** SEA `0.527296`.
- **System confidence edge:** home-minus-away `+0.0546`; diagnostic rounds to `0.05`.
- **Conviction:** Lean.
- **Brier score:** `0.223449`.
- **Team-run absolute error:** `2`.
- **Total-run bias:** `+2` — forecast total `9`, actual total `7`.
- **Correct-winner/wrong-score:** No in the strict score-quality sense; it was a solid correct call, with each team one run above or below the forecast.

### Starter projection versus actual

| Starter | Projected | Actual | Error |
|---|---:|---:|---:|
| Jacob deGrom, TEX | 1 ER, 15 outs, 1.80 ERA | 0 ER, 18 outs, 0.00 ERA | ER `-1`, outs `+3`, ERA `-1.80` |
| Logan Gilbert, SEA | 3 ER, 17 outs, 4.76 ERA | 3 ER, 21 outs, 3.86 ERA | ER `0`, outs `+4`, ERA `-0.90` |

Starter ER projections were accurate or conservative.

### Edges and controls

- Season edge: TEX `-0.371`.
- Recent edge: TEX `-0.610`.
- Pitcher edge: TEX `-0.031`.
- Composite edge: TEX `-0.245`; all directional evidence favored TEX, yet published pick favored SEA.
- Pitcher adjustments: TEX-side `0.9648`, SEA-side `0.9336`.
- Starter-experience shrinkage: not applied.
- Recency guardrail: applied.
- Score-volatility action: `high_total_environment_lift`.
- Confidence path: governed SEA probability `0.5378`; published `0.527296`.

### Score path

The stage table explicitly records:

- `historical_ml`: SEA direction.
- `variance_applied`: SEA became TIE; this was a winner-direction regression because it removed the correct SEA direction.
- `enhanced_prediction`: TIE became SEA, recovering the correct direction.
- Later stages retained SEA.
- Published final winner accuracy remained correct.

This game is the affected game in the high-severity pipeline-instability register. The later recovery is not evidence that the variance stage was safe.

### Player-level adjustments

- Julio Rodríguez: HR and 3 RBI against a projected zero-HR line; impact `+0.62`.
- Corey Seager: `0-for-4` versus projected hit; impact `-1.11`.
- Dominic Canzone and J.P. Crawford were below projection.
- TEX bullpen allowed `4 ER`, including Jakob Junis allowing four earned runs without recording an out; SEA bullpen allowed zero.

### Audit and causes

- Structural audit: PASS; `0 critical`, `0 error`, `0 warning`, `102 good`, `7 info`.
- Primary/secondary causal labels: none assigned.
- Interpretation: correct winner and relatively good score forecast, but the `historical_ml → variance_applied` transition temporarily moved the winner toward a tie. This is a process risk independent of the correct final outcome.

---

# 4. Miss reconstruction

## HOU@PHI

### Earliest wrong-direction stage

The earliest wrong result was already present at `historical_ml`: HOU `4.01`, PHI `5.98`, favoring PHI, the eventual wrong winner. `variance_applied` retained that direction at `4–6`; `enhanced_prediction` reduced PHI to `5` but did not correct the side. Savant and later stages retained the wrong PHI direction.

### Error classification

- **Bad input:** No structural artifact error. However, the starter prior/venue information appears insufficiently represented, especially Javier's away split.
- **Bad weighting:** Likely. The model heavily penalized Javier's projected away performance despite actual one-run performance. The forensic recommendation is to validate a higher `context.venue_adjustment_weight`.
- **Calibration:** The published PHI probability `0.566859` was too strong relative to the outcome, but the primary problem was not merely probability calibration; the directional model selected PHI.
- **Starter error:** Major. Javier projected `5 ER` and allowed `1`; Wheeler projected `3 ER` and allowed `1`.
- **Offense/bullpen miss:** Both lineups underperformed, and PHI's offense scored zero. HOU's bullpen allowed zero, PHI's allowed one.
- **Run-environment miss:** Major. Nine projected total runs versus three actual, a `+6` total bias.
- **Irreducible variance:** Present, particularly the extreme starter outcomes, but the large projection miss also exposes model-weighting risk.

## TB@ATL

### Earliest wrong-direction stage

The earliest wrong-direction movement occurred at `enhanced_prediction`, where the score changed from TB `4`, ATL `5` to TB `5`, ATL `4`. The historical and variance stages favored ATL, but enhancement reversed the winner toward TB. This is the earliest model-stage regression toward the wrong result.

The transition evidence records:

- `historical_ml → variance_applied`: no winner flip for TB@ATL.
- `variance_applied → enhanced_prediction`: ATL → TB; regressed game.
- `enhanced_prediction → embedded_savant_run_environment`: TB → TIE.
- `embedded_savant_run_environment → fundamental_edge_target_review`: TIE → TB.
- Final stages retained TB.

### Error classification

- **Bad input:** Recent head-to-head data strongly favored TB, but it was not representative of the actual low-scoring game. No structural input defect is proven.
- **Bad weighting:** Possible. Recent form and head-to-head inputs overrode the modest season/pitcher evidence and the low-score probability.
- **Calibration:** The published TB probability was only approximately `0.506`; this was not a high-confidence calibration failure. The toss-up label was appropriate.
- **Starter error:** Both starters beat ER projections by two runs, but the forensic assessment calls these outliers and recommends no individual starter change.
- **Offense/bullpen miss:** Major. TB scored one run and its bullpen allowed three; the lineup was four runs below projection.
- **Run-environment miss:** Major. Eleven projected runs versus four actual, a `+7` total bias.
- **Irreducible variance:** Substantial. The game was correctly classified as a toss-up and both starters produced shutouts.

## Pipeline-stage audit across all stages and transitions

| Stage | Winner accuracy | Team-run MAE | Total-run MAE | Assessment |
|---|---:|---:|---:|---|
| `historical_ml` | `0.600`, 3 hits | `2.4915` | `4.6756` | Baseline |
| `variance_applied` | `0.400`, 2 hits | `2.5000` | `4.6000` | Winner regression; TEX moved SEA → TIE |
| `enhanced_prediction` | `0.600`, 3 hits | `2.2000` | `3.6000` | Recovered winner accuracy; improved score metrics, but flipped TB@ATL incorrectly |
| `embedded_savant_run_environment` | `0.600`, 3 hits | `2.2000` | `4.0000` | Margin improved, total MAE worsened |
| `fundamental_edge_target_review` | `0.600`, 3 hits | `2.3000` | `4.2000` | Margin remained improved, team-run quality regressed versus prior stage |
| `capped_target` | `0.600`, 3 hits | `2.3000` | `4.2000` | No change |
| `simulation_result` | `0.600`, 3 hits | `2.3000` | `4.2000` | No change |
| `published_final` | `0.600`, 3 hits | `2.3000` | `4.2000` | No change from simulation |

Important transitions:

- **Historical → variance:** winner accuracy fell by `0.200`; TEX@SEA changed `SEA → TIE`; this is the registered high-severity instability.
- **Variance → enhanced:** winner accuracy recovered by `0.200`; PIT@CHW changed CHW → PIT and TEX@SEA TIE → SEA, helping both; TB@ATL changed ATL → TB, hurting that game.
- **Enhanced → Savant:** TB@ATL moved TB → TIE but winner accuracy did not change; margin MAE improved by `0.4`, while team-run MAE was unchanged.
- **Savant → fundamental review:** TB@ATL moved TIE → TB; margin MAE worsened by `0.2`, team-run MAE worsened by `0.1`.
- All later transitions were inert.

The correct final slate result does not erase the intermediate winner-direction regressions.

---

# 5. Cross-game counterfactuals

## Lever tension 1: `score.variance_multiplier / score.variance_floor`

- **Helps:** `HOU@PHI`, where the model lost a moderate-confidence pick and overprojected the total by six runs. Wider distributions could reduce overconfidence and improve Brier/calibration behavior.
- **Hurts:** No hurt game is listed in the supplied tension record.
- **Net evidence:** `1 helped vs 0 hurt`; supports a **small bounded widening**, not a large change.
- **Guardrail:** This lever cannot fix the near-zero run-model signal. It changes dispersion and calibration, not the underlying mean prediction.

## Lever tension 2: `player.pitcher.v4 form weights / player.pitcher_prior_games`

- **Helps:** HOU@PHI, where Javier was four ER better than projected; PIT@CHW, where Jones was three ER better than projected.
- **Hurts:** COL@NYY and TEX@SEA, where starter ER projections were already accurate within one run; reweighting could disturb those calls.
- **Net evidence:** `2 helped vs 2 hurt`; no directional parameter change is supported from this slate.
- **Guardrail:** Do not generalize two extreme starter outcomes into a global pitcher-weight increase or decrease.

## Lever tension 3: `recent.* / season.* run-mean weighting`

- **Helps:** HOU@PHI, team-run error `6`; PIT@CHW, error `5`; TB@ATL, error `7`.
- **Hurts:** TEX@SEA, where team-run error was only `2`; reweighting could damage an already accurate score forecast.
- **Net evidence:** `3 helped vs 1 hurt`; this is the strongest slate-level score lever signal, but it conflicts with the persistent finding that the run model's features explain almost no holdout variance.
- **Guardrail:** Do not make a large nightly mean-weight change. The evidence supports isolated feature/target work, not blind recency or season reweighting.

## Tempting one-game fixes that should not be made

1. **Do not globally increase venue-split weighting solely because of Javier.** HOU@PHI supports it, but COL@NYY and TEX@SEA had accurate starter projections, and the cross-slate starter ERA bias is explicitly nonpersistent: only `6/11` slates share the sign, with `z=1.11`.

2. **Do not treat PIT@CHW's missing probability transformation as evidence to change the governor.** First trace the pick-side math. The register explicitly prohibits inferring a governor bypass.

3. **Do not tune the model to Oneil Cruz, Cam Smith, or individual player deviations.** Only `11/37` material deviations were supported by season contact quality; the remaining deviations are primarily noise.

4. **Do not force simulation winner intervention based on the zero-flip slate alone.** The cross-slate streak makes simulation a code/diagnostic issue, not evidence for a score-mean or confidence parameter move.

5. **Do not reduce confidence caps solely because TB@ATL lost.** It was a classified toss-up with published probability near 0.50; its loss is not a high-confidence calibration failure.

---

# 6. Data and contract defects

## Critical findings

### 1. Unexplained probability publication delta — PIT@CHW

This is the only registered critical finding:

- Calibrated pick probability: `0.5711`
- Published home probability: `0.528284`
- Difference: `-0.042816`
- No recorded transformation
- `unexplained_material_delta: true`

Required action: reconstruct the complete pick-side path and block promotion until the transformation is recorded and reconciled.

### 2. Missing integrity revision audit artifact

The manifest reports the immutable revision audit as missing. This is an evidence/contract gap even though no per-game structural audit failed.

## Error findings

No structural audit reported an error. Each of the five games had `error: 0`; the slate-level structural summary also reports `critical: 0`, `error: 0`, `warning: 0`.

That does **not** negate the critical probability-lineage issue, because the causal register captures a contract defect not surfaced as a per-game structural error.

## Warning findings

No warning findings were reported. This is a coverage result, not proof that the system is defect-free.

## Informational findings

- Per-game informational findings totaled `37` across the five structural audits.
- All five games received structural QA.
- Simulation had zero flip candidates and zero published flips.
- The PIT@CHW settled per-player feed lacked a complete CHW starter pitching line.
- Several game diagnostics identify score volatility, recency, starter, or player-level context that is useful for interpretation but not necessarily a contract failure.

## Prediction-model defects versus presentation/artifact defects

### Prediction/model/process defects

1. **Near-zero run-model signal:** holdout R² of `0.011791` home, `0.021038` away, and `0.020666` total.
2. **Simulation is persistently inert:** no winner intervention for 11 consecutive slates; self-assessment identifies early-exit and selected-draw construction as the root cause.
3. **Intermediate winner instability:** variance application reduced slate winner accuracy from `0.600` to `0.400`.
4. **Starter projection failure in HOU@PHI:** especially Javier's `11.25` projected ERA versus `1.50` actual.
5. **Overdispersed/high run means on this slate:** average lineup delta `-1.7` across ten sides; total overprediction on both losses.

### Presentation/artifact/contract defects

1. **PIT@CHW probability lineage is incomplete.**
2. **Missing integrity revision audit file.**
3. **PIT@CHW settled starter-feed line is incomplete.**
4. Per-game forecast files themselves were reported clean: every game's diagnostic says the saved artifact had zero defects and was not why the prediction missed.

## Complete probability-lineage audit

The required path is raw model → calibrated → governor → score coherence/recency → published.

| Game | Raw | Calibrated/governed | Published | Reconciliation |
|---|---:|---:|---:|---|
| COL@NYY | `0.7103` | `0.6858` | `0.580` | Recorded transformation: score coherence plus matchup-recency guardrail |
| HOU@PHI | `0.6273` | `0.6031` | `0.566859` | Recorded transformation; no unexplained material delta |
| PIT@CHW | `0.5000` | `0.5711` | `0.528284` home / `0.471716` away | **Bypass/undocumented transformation cannot be distinguished; critical gap** |
| TB@ATL | `0.52605` | `0.5080` | `0.505802` away / `0.494198` home | Published value matches calibrated/governed value within rounding/orientation |
| TEX@SEA | `0.5301` | `0.5378` | `0.527296` | Recorded transformation: score coherence plus matchup-recency guardrail |

A bypass should be called only when a material delta has no recorded transformation. That condition is met for PIT@CHW, but the evidence does not prove that the governor itself was bypassed.

---

# 7. Cross-slate persistence

A single slate is noisy, but several findings have explicit persistence evidence.

## Escalated findings

These are established failures, not single-slate noise.

### `simulation_inert_on_slate`

- Consecutive slates: **11**
- Present now: **Yes**
- Incidence: `11/11` in the window
- Target metric: `team_run_maE`, `total_run_maE`
- Target lever listed: `simulation.allow_score`
- Parameter actionable: **No**
- Direction: confirm the simulation module is running and not silently no-oping.
- Root cause from self-assessment: the same selected draws are used for target realization and distribution estimation; the loop early-exits after approximately two draws, making the resulting win rate statistically meaningless.

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

### `weak_run_model_signal`

- Consecutive slates: **11**
- Present now: **Yes**
- Incidence: `11/11`
- Target metrics: `team_run_mae`, `total_run_mae`
- Target levers listed: `training.xgb.n_estimators`, `training.xgb.max_depth`, `training.xgb.learning_rate`
- Parameter actionable: **No**
- Direction: investigate features and target through an isolated training backtest.

The self-assessment correctly states that no bounded nightly knob can reliably fix near-zero holdout R². The recommended action is feature-contract and training-target work.

## Finding streaks

| Finding | Consecutive-slate count | Present now | Interpretation |
|---|---:|---:|---|
| `high_raw_probability_cluster` | `0` | No | Not persistent now; incidence `2/11` |
| `pipeline_winner_instability` | `1` | Yes | Present now, but not escalated; incidence `4/11` |
| `score_margin_compression` | `0` | No | Not persistent now; incidence `3/11` |
| `simulation_inert_on_slate` | **11** | Yes | Established persistent failure |
| `single_slate_calibration_instability` | `0` | No | Not persistent now; incidence `5/11` |
| `weak_run_model_signal` | **11** | Yes | Established persistent failure |

The `escalation_streak` value is `3`, but the individual escalated finding records independently show 11 consecutive slates. The latter is the relevant streak for each finding.

## Metric-trend directions

- **Margin bias: improving.** Mean `-1.6984`; the series improves toward less negative values in aggregate.
- **Mean team-run error on losses: regressing.** Mean `6.1631`; losses remain associated with large team-run misses.
- **Weighted calibration error: improving.** Mean `0.1604`; current slate `0.1171`, better than several earlier slates.
- **Winner accuracy: flat.** Mean `0.5751`; current `0.600` is not evidence of a sustained directional improvement.

## Top-confidence-pick loss streak

- Current top-confidence-pick loss streak: **0**.
- The last three slates, September 8–10, were hits.
- Prior losses occurred on September 1–3 and September 6–7, but they were interrupted by hits.
- Therefore there is no current top-pick loss streak to remediate.

## Other persistent cross-slate evidence

The `cross_slate_bias` section does **not** support changing starter opponent-strength weighting or score variance floor from the available 11-slate tests:

- Starter ERA error: positive mean `0.418`, same-sign share `0.545`, `z=1.11`; status `hold`.
- Lineup run delta: mean `0.0`, same-sign share `0.364`, `z=0.0`; status `hold`.

These nonpersistent findings should not override the two explicitly escalated persistent failures.

---

# 8. Adjustments for today

Recommendations are ranked by persistence and evidence strength. These are bounded recommendations or shadow experiments only; no files or parameters are being changed here.

## 1. Fix the persistent inert simulation path — highest priority

- **Target:** `simulation.allow_score` is registered, but the finding is explicitly non-parameter-actionable.
- **Change:** do **not** toggle the parameter blindly. Implement the registered code/diagnostic action:
  - separate target-realization search from independent win-probability estimation;
  - use an independent unsteered batch of approximately `200–500` draws or a closed-form Poisson/Skellam difference;
  - remove the early exit from the probability-estimation loop;
  - register the batch size;
  - feed the independent win rate into `_distribution_agreement`.
- **Expected metric:** better-informed simulation contribution, improved `team_run_mae` and `total_run_mae`, and measurable nonzero winner-intervention diagnostics.
- **Affected games:** this slate's simulation did not alter any winner, so no individual game can demonstrate benefit; structurally it is relevant to all five.
- **Lever tension:** no direct one-game hurt is established; the principal risk is computational cost or introducing a new simulation calibration error.
- **Rollback monitor:** compare independent-simulation win rates, flip-candidate frequency, Brier score, team-run MAE, and total-run MAE against the current inert baseline.

## 2. Address persistent weak run-model signal — code and training diagnostic

- **Target levers:** registered as `training.xgb.n_estimators`, `training.xgb.max_depth`, and `training.xgb.learning_rate`, but the finding is marked `param_actionable: false`.
- **Change:** do not make a nightly hyperparameter move. Add and validate the prescribed feature/target changes:
  - source-labelled no-vig market probability where available;
  - park/weather run environment;
  - bullpen rest and fatigue;
  - platoon-adjusted lineup quality;
  - isolated same-fingerprint walk-forward backtest;
  - bump model schema version when feature semantics change.
- **Expected metric:** increase holdout home/away/total R² and reduce team-run and total-run MAE.
- **Affected games:** especially HOU@PHI, PIT@CHW, and TB@ATL, which had team-run errors of `6`, `5`, and `7`; TEX@SEA is the principal regression-risk comparison because its error was only `2`.
- **Lever tension:** the slate's recent/season run-mean lever helped three games and hurt one, but this does not justify changing it before feature/target validation.
- **Rollback monitor:** require improvement in holdout R² and out-of-sample run MAE, not merely improvement on this five-game slate.

## 3. Apply one small shadow change for the strongest single-slate driver: widen score variance

- **Registered parameter:** `score.variance_floor`.
- **Direction:** increase slightly.
- **Magnitude:** bounded shadow increase of approximately `+0.10` from the default `2.75` to `2.85`; do not promote without a gate.
- **Expected metric:** lower overconfidence/Brier risk in moderate-confidence low-scoring misses and improve total-run calibration; team-run MAE may improve if the current distribution is too narrow, but the mean-model weakness limits expected gains.
- **Games helped:** HOU@PHI, where the model overprojected total runs by six; potentially TB@ATL, where it overprojected by seven; PIT@CHW, where it overprojected by five.
- **Games hurt:** TEX@SEA is the clearest risk because its team-run error was only two; COL@NYY had a reasonable total and a supported blowout expansion. The lever tension record lists HOU@PHI as helped and no explicit hurt game, but these risks remain operationally plausible.
- **Rollback monitor:** Brier score, weighted calibration error, total-run MAE, team-run MAE, and margin MAE over a rolling multi-slate shadow period.
- **Why not a large change:** the score model has persistent near-zero signal, so widening uncertainty must not be mistaken for fixing the mean.

## 4. Probability-lineage repair before promotion

- **Target:** no parameter move.
- **Change:** reconstruct PIT@CHW's side-oriented raw → calibrated → governor → score coherence/recency → published chain, including the exact transformation from `0.5711` to `0.528284` home probability.
- **Expected metric:** zero unexplained material deltas and a complete audit trail.
- **Affected game:** PIT@CHW directly; all future games through the publication contract.
- **Rollback/gate:** no promotion until the transformation is recorded and the missing integrity revision audit is restored or formally explained.

## 5. Do not change pitcher form weights today

- **Registered lever:** `player.pitcher.v4 form weights / player.pitcher_prior_games`.
- **Direction:** abstain.
- **Reason:** slate tension is exactly balanced, `2 helped vs 2 hurt`; cross-slate starter bias is not persistent.
- **Affected games:** HOU@PHI and PIT@CHW support more adaptive treatment, but COL@NYY and TEX@SEA caution against a global change.
- **Action:** retain shadow diagnostics, particularly venue-split treatment, without changing the global weight.

## 6. Do not change confidence caps solely from this slate

The self-assessment proposes lowering `confidence.base_cap` and `confidence.moderate_cap`, but the actual slate had no high-confidence category and the top-pick loss streak is zero. A small calibration shadow could be reasonable, but it is lower priority than the persistent simulation and run-model defects. The current weighted calibration error is improving cross-slate.

---

# 9. Bottom line

## Known

- The slate finished `3–2` with winner accuracy `0.600`.
- COL@NYY, PIT@CHW, and TEX@SEA were correct winner calls; PIT@CHW was explicitly a correct-winner/wrong-score game.
- HOU@PHI was lost primarily through a severe starter projection error and an overprojected run environment.
- TB@ATL was lost in a toss-up after the enhanced stage flipped from the historically correct ATL direction to TB; both starters and the run environment then produced a very low-scoring game.
- Exact score quality is weak, supported by published total-run MAE `4.2` and holdout run R² near zero.
- Simulation inertness and weak run-model signal are established 11-slate persistent failures.
- The current top-confidence pick loss streak is zero.
- All five games received structural QA, and all per-game structural audits reported zero critical/error/warning findings.

## Uncertain

- The correct interpretation of the PIT@CHW probability publication transformation remains unknown. The evidence proves an undocumented material delta, not a governor bypass.
- The degree to which venue-split weighting would improve starter projections is unresolved; cross-slate evidence does not separate that bias from noise.
- The magnitude of any beneficial score-variance change is uncertain because the run model itself has near-zero validated signal.
- Player-level deviations are mostly not learnable from this slate: only `11/37` material deviations were supported by underlying season contact quality.

## Single safest next action

**Repair and verify the probability-lineage and immutable-audit contract before promotion, while starting a shadow implementation of independent simulation win-probability estimation.** This addresses the only critical evidence defect and the strongest persistent process failure without overfitting the two-game loss pattern or disturbing the balanced pitcher and recent/season weighting tradeoffs.