Ops

Ops / Pipeline Status

Real components mapped to the spec's software-agent org chart roles -- see docs/22-agent-org-chart.md for exactly what this is and isn't. There is now an autonomous orchestrator (scripts/orchestrator/run.mjs): it runs unattended in CI on every push and on schedule, executes the pipeline in dependency order, gates each stage on a health check of the data that stage actually wrote, retries what it can, and records the outcome below. It is deliberately rules-based rather than an LLM -- the same standard the playoff engine is held to. Everything on this page is read from that pipeline's own database at build time; nothing here is written by hand.

Recent runs: degraded (local-schedule) · degraded (local-schedule) · degraded (local-schedule) · degraded (local-schedule) · degraded (local-schedule)

2026 week 5 Week 5 of 2026, score-ridge/v1: 15 of 15 games have a projection from before kickoff. Projected margin MAE is 10.67 points and RMSE is 14.27; the closing spread on the same games has MAE 10.40 and RMSE 13.04. Projected total MAE is 11.91 points and RMSE is 12.77; the closing total on the same games has MAE 10.17 and RMSE 10.87. The projection and the closing spread disagreed on 14 games, and 8 landed. 14 is below the 30-game floor used before a rate is read against 52.4%, so this is not that test. The projection and the closing total disagreed on 15 games, and 5 landed. 15 is below the 30-game floor used before a rate is read against 52.4%, so this is not that test. On 1 game the spread stored with the projection differs from the closing spread by at least 0.25 points. The grade uses the closing spread. The walk-forward measurement on 2199 games, 2018-2026, remains the bar: margin disagreement 50.07%, total disagreement 49.67%, projected margin MAE 10.23 against the closing line's 9.79. Both disagreement rates sit below the 52.4% break-even at -110. This grade does not replace that measurement.

2026 through week 5 Through week 5 of 2026, score-ridge/v1: 46 of 79 games have a projection from before kickoff. Projected margin MAE is 8.11 points and RMSE is 11.13; the closing spread on the same games has MAE 8.20 and RMSE 10.99. Projected total MAE is 12.00 points and RMSE is 13.34; the closing total on the same games has MAE 10.35 and RMSE 11.60. The projection and the closing spread disagreed on 41 games, and 28 landed (68.29%). 52.4% is the -110 break-even. This live count does not replace the walk-forward result. The projection and the closing total disagreed on 45 games, and 18 landed (40.00%). 52.4% is the -110 break-even. This live count does not replace the walk-forward result. On 4 games the spread stored with the projection differs from the closing spread by at least 0.25 points. The grade uses the closing spread. 1 game had only a post-kickoff snapshot, left out. 32 games had no snapshot, left out. Weeks 1 and 2 are final with no pre-kick snapshot, so those weeks are not in these averages. The walk-forward measurement on 2199 games, 2018-2026, remains the bar: margin disagreement 50.07%, total disagreement 49.67%, projected margin MAE 10.23 against the closing line's 9.79. Both disagreement rates sit below the 52.4% break-even at -110. This grade does not replace that measurement.

The qa-watchdog stage counts these and points here. Each row is one open problem in qa_defects: a missed pre-kickoff run, a cron that will not fire, a session that has expired. It stays until a later pass no longer sees it.

Cross-cutting checks run after ingestion, over the whole database. Each one stores the number it actually measuredalongside the threshold it was compared against, so this is evidence rather than a green tick. A blocking failure stops the deploy; a warning ships and is shown. Last run 2026-10-19 06:17:12.096+00.

The teams table holds 36 codes for 32 clubs, and different providers spell some of them differently. Every row below was derived from real data, not typed in: relocations from the schedule's own season spans, duplicate codes from clubs sharing a name with no games played, and external spellings by asking ESPN's public teams endpoint what it calls each club. The live scoreboard reads this table instead of the hand-written map it used to carry.

15,096 rows — one per team per game across 1999–2026 (7,548 games) — built by scripts/features/build-features.mjs. Nothing is imputed: a value that can't be computed from real data is left null, and the reason is stated below. Coverage is measured at build time, not asserted. Form, Elo and season-to-date splits are computed strictly from games played before each row's own game, so a backtest can't score itself with results it wouldn't have had.

This table is deliberately not a claim that these features predict anything. Measured on the completed games in it, the de-vigged market probability alone picks winners at 66.3% and the Elo baseline at 62.4% — a model built on the rest of these columns still has to beat the market number before it's worth trusting, and so far none has.

These are real messages passed between pipeline components during actual runs — not a simulated conversation. The rule they are held to: a message only counts if something behaves differently because of it. So each one records whether it was acted on and what was actually done. A message nobody acted on shows as exactly that, because a bus where nothing is consumed is decoration. 40 of the last 40 were acted on.

Per the spec's own framing: "hiring" means activating a real template. These roles have no real implementation behind them yet, so they're honestly listed as not hired rather than shown with invented activity -- and each one now states what is actually missing, so the list is something you can act on rather than a shrug.

Snapshotted via scripts/ops/snapshot-ci-status.mjs -- run manually, not live on every page load (the app doesn't hold a standing GitHub token).