Luna deep analysis — slate 20261009

Coverage receipt: COMPLETE · 9/8 sections; 0/0 games; 2/2 causes; 3/3 sources

1. Evidence receipt and gaps

The supplied diagnosis identifies a slate date of "" and reports games: [], games_reviewed: 0, hits: 0, and misses: 0 (date, top level). Thus no individual game or game outcome is present to review. The zero counts do not establish that the actual slate had no games; the expected-game count is itself zero, and the audit reports that coverage is incomplete.

Source manifest, entry by entry:

Coverage and internal consistency gaps:

• coverage_matrix reports expected_games: 0, canonical_artifacts_loaded: 0, structural_audits_completed: 0, forensic_deep_audit_games: 0, continuous_learning_games: 0, and games_with_starter_comparisons: 0; all_games_structurally_audited is false.
• The causal register nevertheless describes the coverage failure as “0/0 games structurally audited.” That accurately states no audit rows exist, but does not establish that there were zero games to audit.
• learning_eligibility.eligible is true because “all game artifacts passed critical/error structural gates,” despite zero canonical artifacts loaded and zero structural audits completed. That eligibility claim is not supported by substantive per-game QA; it appears compatible with a vacuous pass, not proof of a clean slate.
• structural_audit_summary reports zero critical, error, info, and warning findings, while the causal register explicitly records a critical coverage issue and an info simulation issue. These summaries have different apparent scopes or are inconsistent; the missing structural-audit artifact prevents reconciliation.
• The top-level date is blank, while manifest paths and source descriptions refer to 20261009. The supplied JSON does not resolve that discrepancy.

Bottom line on evidence receipt: the manifest records three sources as loaded, but this payload contains no game records and the coverage matrix says no canonical game artifacts were loaded. The actual slate’s game count, game identities, picks, outcomes, probabilities, scores, starters, and player adjustments cannot be independently confirmed from the supplied evidence.

2. Executive diagnosis

Causal issue register reconciliation:

1. incomplete_slate_audit_coverage — severity: critical; domain: coverage.
2. Evidence: “0/0 games structurally audited” (causal_issue_register[0].evidence).
3. Affected games: listed as []; with no game manifest or game rows, the affected games cannot be named.
4. Recommended action: “Do not call the slate clean; run per-game QA on every canonical artifact first.”
5. Guardrail: “A clean subset of games cannot certify the games that were never audited.”
6. Diagnosis: the critical failure is that coverage cannot be demonstrated. It does not prove that any particular prediction was wrong.

1. simulation_inert_on_slate — severity: info; domain: simulation.
2. Evidence: “flip candidates 0; published flips 0.”
3. Affected games: []; no game-specific simulation record is provided.
4. Recommended action: measure the simulation’s contribution over rolling slates, separating score realization from winner flipping.
5. Guardrail: “Zero winner flips is not a defect and does not mean the simulation did no score-reconciliation work.”
6. Diagnosis: zero flips alone is informational, not a demonstrated slate-level prediction failure. The separate cross-slate evidence escalates a persistent simulation-inert pattern; that is addressed below.

No game-level outcome, probability, or scoring claim can be made from these summaries.

3. Every game

No games are present in the supplied games array. Consequently, there are no identifiable games to enumerate, and no per-game predicted/actual score or winner, probability, confidence edge, conviction bucket, Brier score, run error, starter line, season/recent/pitcher/composite edge, control, score path, player adjustment, audit finding, or cause can be reported.

Correct-winner/wrong-score games: cannot be identified. This is not evidence that there were none; no game results or predictions are supplied.

4. Miss reconstruction

There are no observable loss records to reconstruct: games is empty, misses is zero as a count over zero reviewed games, and loss_primary_drivers is {}. Therefore it would be incorrect to say there were no losses on the actual slate. No earliest wrong-direction model stage, bad input, bad weighting, calibration effect, starter error, offense/bullpen miss, run-environment miss, or irreducible-variance contribution can be assigned to a specific game.

The instruction that a 0.12 home-away confidence edge is moderate rather than near a coin flip is noted, but this payload contains no confidence edges to assess.

Pipeline-stage and transition audit:

• pipeline_stage_quality.status is not_measurable.
• stages and transitions are both empty; the largest team-run regression versus baseline is null.
• Therefore there is no evidence here to compare raw, intermediate, or later score stages with official results, and no evidence of an intermediate winner-direction regression or later recovery.
• A later-stage recovery cannot be credited or used to erase an intermediate regression when no transitions are supplied.

Persistent simulation evidence, distinguished from game-level reconstruction: the cross-slate escalation says simulation_inert_on_slate occurred for 11 consecutive slates, with zero flip candidates and zero published flips in the current register entry. The self-assessment provides a specific hypothesized mechanism: the score-realization search and win-probability distribution use the same draws, and the loop may stop once it finds a target-realizing path, leaving an estimated win rate based on roughly two selected draws. That mechanism is a supplied diagnosis and merits investigation; it is not independently verifiable from the missing per-game traces here.

5. Cross-game counterfactuals

All supplied lever_tension entries have empty helps and hurts lists. With no games or game-level outcomes, no affected-game counterfactual can be named.

Tempting one-game fixes not justified: changing pitcher-form weights, recent-versus-season weighting, or run-variance parameters based on a purported game-level miss would be unsupported because no game-level miss is provided. These are not shown to damage a particular game here; rather, their cross-game effects are unobserved, so no such one-game fix is evidence-based.

6. Data and contract defects

Findings and severity

• Critical — incomplete_slate_audit_coverage (causal register). No canonical game artifacts were loaded and no games were structurally audited; the register records “0/0 games structurally audited.” The required action is per-game QA on every canonical artifact. The game IDs and actual coverage denominator are missing.
• Info — simulation_inert_on_slate (causal register). The register reports zero flip candidates and published flips. Its own guardrail says this does not prove the simulation did no score-reconciliation work. Cross-slate evidence separately establishes a persistent issue, discussed in Sections 7–8.
• Error/warning findings: no individual error or warning finding records are supplied. The structural_audit_summary counts are all zero, but because the structural audit HTML is missing and coverage is zero, those counts cannot certify that all games had zero errors or warnings.
• Contract inconsistency: learning_eligibility.eligible: true (“all game artifacts passed critical/error structural gates”) conflicts with the zero artifacts and zero structural audits in coverage_matrix. Treat eligibility as unsubstantiated until complete coverage is demonstrated.
• Slate identity gap: top-level date is blank, while the manifest paths identify the 20261009 reports.

Coverage reconciliation

The slate is not cleanly covered: all_games_structurally_audited is false, and all artifact/game audit counts are zero. The manifest’s missing detailed, player-audit, and structural-audit reports further limit verification. A structural summary of zero findings does not reconcile that absence.

Pick-side probability lineage

probability_lineage is an empty array. There is no trace for any game through:

1. raw model probability,
2. calibration,
3. confidence governor,
4. score-coherence or recency adjustment, and
5. published probability.

Therefore it is not possible to audit the full lineage or quantify any stage delta. No material delta is demonstrated, so a probability-lineage bypass cannot responsibly be asserted; it is also not possible to rule one out. probability_reversions is explicitly unavailable because no game artifacts were supplied. Calibration monitoring has zero settled games, empty buckets, and null weighted error.

Model defects versus presentation/artifact bugs

• Demonstrated evidence/contract concern: incomplete audit coverage and the unsupported learning-eligibility claim.
• Persistent code/diagnostic concern: simulation distribution may be ineffective as a win-probability signal, according to the escalated trend and supplied root-cause diagnosis.
• Presentation/artifact availability: three HTML artifacts are missing, which blocks review of rendering, player traces, and structural findings. The evidence does not establish a specific rendering bug.
• Not established: a per-game prediction-model defect, a particular probability bypass, a starter error, or an erroneous player adjustment. The game evidence required to establish these is absent.

7. Cross-slate persistence

Raw structural evidence

• margin_distribution: no rows, predicted margins, or actual margins; counts for actual margins of at least 4 and predicted margins of at most 2 are both zero. It cannot be evaluated from this slate’s supplied structural rows.
• pipeline_stage_quality.transitions: empty; no transition-level regression or recovery can be checked.
• calibration_monitoring: zero settled games, empty buckets, and null weighted absolute calibration error.
• high_raw_cluster: threshold 0.65, with zero games, zero hits, and no rows. No high-raw-confidence cluster can be assessed on this slate.
• simulation_control: key simulation_flip_control; zero observed games, zero simulation flip candidates, zero published flips, priority P2, status diagnostic_only_without_independent_support. This current-slate record is consistent with the informational causal-register entry; the persistent diagnosis comes from the cross-slate streak below.

Every finding streak

Escalated finding

• simulation_inert_on_slate — 11 consecutive slates.
• Target lever listed: simulation.allow_score; target metrics: team_run_mae, total_run_mae.
• param_actionable is false. The recommended direction is to confirm that the simulation module is running and not silently no-oping.
• The associated self-assessment identifies an operator/code action, not a parameter retune: separate the target-realization search from win-probability estimation; compute probability from an independent unsteered batch (suggested 200–500 draws, or a closed-form distribution), avoid early exit for the probability batch, register its size, and feed the resulting win rate into _distribution_agreement. The stated suspected issue is reuse of target-seeking draws that may stop after roughly two selected draws.
• Since this is escalated across 11 consecutive slates, it should be treated as an established process failure requiring the named code/diagnostic action—not dismissed as single-slate noise. It does not, by itself, prove that score MAE worsened on any specific game.

Cross-slate trend directions

• margin_bias: regressing, mean -1.7778. The series includes null entries; the supplied direction should be reported, but the underlying margin rows are absent from this slate payload.
• mean_team_run_error_on_losses: improving, mean 2.375. The series is sparse, so this summary is not a per-game explanation.
• weighted_calibration_error: regressing, mean 0.429. This is a cross-slate trend; current-slate calibration itself is unmeasured.
• winner_accuracy: improving, mean 0.7778. This trend does not override the missing current-slate accuracy or establish score quality.

Top-confidence-pick loss streak: loss_streak: 0. The listed results include misses on 20261005 and 20261007, but hits on 20261006 and 20261008, with two entries marked unknown and blank dates. Thus the reported consecutive loss streak is zero; the list is not a complete, cleanly dated record for every window slot.

Cross-slate bias holds, not escalations: context.opponent_strength_weight has negative starter ERA error over 10 slates (mean -0.385, same sign 6/10, z=-0.89) and is explicitly held as not separable from noise. score.variance_floor has negative lineup run delta over 10 slates (mean -0.562, same sign 7/10, z=-1.71) and is also explicitly held. These do not warrant overriding their stated no-change actions based on this evidence.

The trend window says 11 slates, while cross_slate_bias.lookback_slates is 12; the per-slate listing includes a zero-game date, and window_dates contains blank entries. These are additional limits on interpreting the rollups.

8. Adjustments for today

Recommendations are ranked by evidence strength and scope. None is a promotion or a claim that a change has been validated.

1. Highest priority — resolve the critical audit-coverage failure before using slate-level learning eligibility.
2. Action: run per-game structural QA against every canonical game artifact and reconcile the canonical game count with the expected slate count; do not treat zero artifacts as a successful all-clear.
3. Parameter: none. This is a coverage/contract action, not a model lever.
4. Expected metric and direction: structural-audit coverage should increase from 0 audited games to the full canonical game count; eligibility should only be claimed after that coverage is demonstrated.
5. Affected games / tension: none can be named from this payload; all game IDs are absent. It should not directly alter predictions.
6. Risk: low model-regression risk; high importance for preventing unsupported claims about slate quality. The missing game artifacts mean the target count must first be established.

1. Persistent escalated pattern — shadow diagnostic/code experiment for simulation probability estimation.
2. Finding: simulation_inert_on_slate, present for 11 consecutive slates; target metrics are team_run_mae and total_run_mae; param_actionable: false.
3. Action: implement the supplied bounded diagnostic in an isolated shadow path: retain target-realization search for score construction, but estimate win probability from an independent, unsteered fixed-size batch. Start with 200 draws, within the supplied 200–500 suggestion; do not early-exit that batch for finding a target score. Register the batch size, calculate the independent win rate, and compare it with the existing distribution-agreement signal before any production use.
4. Registered parameter: the finding lists simulation.allow_score, but the supplied root-cause/action says this is a code-path issue, not a parameter-tuning action. Do not blindly toggle simulation.allow_score; verify its behavior and instrument it. The registered draw-count control described by the action plan should be exposed for the shadow test.
5. Expected metric and direction: the first required outcome is a non-vacuous, independently measured simulation contribution. The intended downstream direction is lower team_run_mae and total_run_mae; improvement is a hypothesis to test, not a guaranteed effect.
6. Affected games / tension: no slate games are supplied, and all lever-tension lists are empty. The game-level helps-versus-hurts balance is unknown.
7. Risk: moderate code/compute risk; keep it shadow-only, compare rolling outcomes, and roll back if run-error metrics worsen or the independent distribution is unstable. This follows the escalated operator action rather than forcing an unsupported parameter change.

1. Strongest single-slate process driver — restore auditable game-level evidence before making model adjustments.
2. Action: reconcile the blank slate date and zero expected-game denominator with the 20261009 manifest paths; load the canonical game artifacts and produce the missing per-game audit rows.
3. Parameter: none. No registered model lever can repair missing evidence.
4. Expected metric and direction: canonical_artifacts_loaded, structural_audits_completed, and forensic_deep_audit_games should move from zero to the verified game count; calibration and prediction metrics should become measurable rather than null.
5. Affected games / tension: currently unidentified. No model-side help/hurt tradeoff can be assessed.
6. Risk: low prediction-regression risk; this is the prerequisite for reliable diagnosis.

No confidence, pitcher, run-mean, or variance parameter retune is recommended from the absent game-level records. Persistent but nonconsecutive findings and cross-slate bias holds do not justify a speculative change here. A bounded score-margin shadow test could be considered in a later, fully covered review, but this payload cannot identify which games it helps or hurts.

9. Bottom line

Known: there are no game rows or canonical game artifacts in the supplied evidence; the slate is not structurally covered; calibration, score, starter, and winner accuracy are not measurable from this payload. The cross-slate record reports improving winner-accuracy trend, regressing margin-bias and weighted-calibration-error trends, and an escalated simulation-inert pattern persisting for 11 consecutive slates.

Uncertain: the number of actual games reviewed, every game’s prediction and result, whether any winner or score prediction was correct, the game-level probability lineage, and whether the simulation issue caused any specific miss or score error.

Safest next action for the very next slate: ensure every canonical game artifact is loaded and structurally audited before interpreting eligibility or making model changes. In parallel, develop the independent simulation win-probability calculation as a shadow diagnostic; do not promote it or tune unrelated model levers on the basis of this incomplete slate.