Nightly self-improvement · slate 2026-09-03

Generated 2026-09-04T02:00:00.004140-04:00 ET from official finals (mlb-api). Every learning stage is listed with what it did. "Blocked" means it wanted to act and a safety rule stopped it; "nothing to learn" means it had no eligible evidence.

Slate

2026-09-03

status completed

Winners

5/9

Brier 0.2598

Learned or changed

of 5 stages

Blocked

wanted to act and could not

Release

v4.17.0

consistent

Audit Detailed Deep Slate Review Luna Post-mortem Daily Learning Adaptive Status

What each learning stage did

Stage	Outcome	Detail
KPI settlement	failed	did not run
Bullpen usage model	nothing to learn	No pregame bullpen forecasts to settle. Refit: not attempted.
Adaptive parameter proposal (Luna)	nothing to learn	candidate_registered:
Auto-apply	nothing to learn	no_promotable_candidate
Code agent	nothing to learn	disabled_by_config: code changes are made on the development host, never in the container.

Games

Game	Predicted	Actual	Home win prob	Result	Main driver	Starters
BOS@BAL	5-4	6-5	0.470	hit		Jake Bennett: 17 proj / 18 real outs; Brandon Young: 17 proj / 20 real outs
CHW@HOU	4-5	2-6	0.559	hit		Luis Castillo: 14 proj / 12 real outs; Hunter Brown: 17 proj / 18 real outs
MIA@KC	5-4	3-7	0.452	miss	Small edge on the losing side	Sandy Alcantara: 19 proj / 18 real outs; Michael Wacha: 19 proj / 18 real outs
MIL@CHC	7-6	1-2	0.466	miss	Starting-pitcher projection was wrong	Logan Henderson: 18 proj / 21 real outs; Kevin Gausman: 17 proj / 21 real outs
OAK@SEA	4-6	7-4	0.637	miss	Backed the wrong side with real conviction	Jack Perkins: 12 proj / 9 real outs; Kade Anderson: 13 proj / 15 real outs
SF@PIT	5-6	2-5	0.530	hit		Blade Tidwell: 15 proj / 15 real outs; Lake Bachar: 15 proj / 7 real outs
STL@LAD	4-5	2-3	0.530	hit		Quinn Mathews: 14 proj / 15 real outs; Tarik Skubal: 18 proj / 17 real outs
TB@TEX	5-4	0-6	0.466	miss	Small edge on the losing side	Shane McClanahan: 16 proj / 15 real outs; Cal Quantrill: 17 proj / 21 real outs
TOR@CLE	4-3	6-3	0.453	hit		José Soriano: 18 proj / 14 real outs; Tanner Bibee: 18 proj / 12 real outs

Day over day

Rolling evidence window against the prior night. Associations, not causal proof.

Metric	Prior	Current	Change	Result
brier_score	0.2535	0.2520	-0.0015	improved
accuracy	0.5250	0.5500	0.0250	improved
team_run_mae	2.7542	2.7667	0.0125	regressed
total_run_mae	3.8750	3.8500	-0.0250	improved
calibration_error	0.1009	0.0989	-0.0020	improved
starter_er_mae	1.9463	1.9192	-0.0271	improved
starter_outs_mae	3.0537	3.0192	-0.0345	improved

Calibration produced for the next slate

Metric	Scale used	Measured	Rows	Status
No calibration produced.

Luna KPI candidates and their promotion gate

Promotion needs 400+ prospective rows over 5+ slates after the proposal, with a gain interval above zero. Only verified pregame forecasts count; local replays never do.

KPI	Status	Target	Proposed	Rows	Slates	Gain interval
luna_pitcher_release_speed_vs_righthanded_starters	shadow	SO	20260923	330/400	5/5	[-0.0007, 0.0007]
luna_pitcher_barrel_rate_vs_rhh_righthanded_starters	shadow	HR	20260923	330/400	5/5	[-0.0004, 0.0049]
luna_batter_contact_rate_vs_breaking_righthanded_starters	shadow	H	20260923	330/400	5/5	[-0.0045, 0.0044]
luna_batter_contact_rate_vs_breaking_left_handed_starters_for_strikeouts	shadow	SO	20260924	122/400	4/5	[-0.0027, 0.0021]
luna_pitcher_contact_xwoba_allowed_vs_righthanded_starters_for_strikeouts	shadow	SO	20260924	122/400	4/5	[-0.0001, 0.0000]
luna_pitcher_whiff_rate_breaking_vs_left_handed_hitters_for_hits	shadow	H	20260924	122/400	4/5	[-0.0056, -0.0011]
luna_pitcher_home_run_rate_vs_righthanded_starters	shadow	HR	20260925	122/400	4/5	[-0.0010, 0.0056]
luna_batter_ground_ball_rate_vs_fastball_vs_lhp_for_home_runs	shadow	HR	20260926	86/400	3/5	[-0.0001, -0.0000]
luna_pitcher_whiff_rate_vs_right_handed_starters_for_batter_strikeouts	shadow	SO	20260929	51/400	2/5	[–, –]
luna_batter_whiff_rate_vs_breaking_right_handed_starters_for_hits	shadow	H	20260929	51/400	2/5	[–, –]
luna_batter_walk_rate_vs_righthanded_pitchers_for_home_runs	retired	HR	20260924	0/400	0/5	[–, –]

Operator actions the cycle asked for (0)

Changes to code or a model recipe are never made by the nightly cycle. These are the measured findings that need one.

Priority	Owner	Finding	Action
No operator actions requested.

Luna post-mortem

Luna (gpt-5.6-luna) post-mortem, status received. Full analysis.

# 1. Evidence receipt and gaps

## Sources received

All six available source-manifest artifacts were marked loaded:

1. **Prediction review** — `/app/output/reports/prediction_review_20260903.json`, 225,604 bytes; official outcomes and scoring errors.
2. **Forensic deep audit** — `/app/output/reports/PREDICTION_SYSTEM_DEEP_AUDIT_20260903.json`, 151,096 bytes; probability, provenance, starter, and decision contracts.
3. **Continuous learning** — `/app/output/learning/continuous_learning_daily_20260903.json`, 1,175,839 bytes; pipeline-stage quality, calibration, probability lineage, score distributions, and official starter comparisons.
4. **Detailed report** — `/app/output/reports/detailed_20260903.html`, 215,805 bytes; game and simulation rendering.
5. **Ultra-detailed player audit** — `/app/output/reports/20260903_ultra_detailed_player_audit.html`, 18,544,069 bytes; player-level prediction trace.
6. **Structural audit** — `/app/output/reports/audit_report_20260903.html`, 1,125,883 bytes; invariant findings.

One manifest entry was missing:

- **Integrity revision audit** — `/app/output/reports/audit_report_20260903_integrity_revision.html`; intended to provide post-remediation proof for immutable revision artifacts. Bytes were null and status was `missing`.

That missing artifact is a receipt gap. It does not establish that the predictions were wrong, but it prevents a complete independent verification of revision immutability and post-Luna remediation claims.

## Slate cov