Instruction file imported from Risky-Scout/NCAAF-Predictive-Model (.cursor/rules/00-charter.mdc). Copyright stays with the author.

Charter

Objective: a reproducible NCAAF prediction pipeline that, for every game, outputs the six probabilities (moneyline, both spread sides, game total O/U, both team total O/U) from a negative-binomial joint PMF built on ridge opponent-adjusted ratings. Target: verified model with LIVE Week-1 2026 predictions within 4 days, per WEEK1_LIVE_PROTOCOL.md. Reference modules in reference_impl/ are the validated critical path — assemble around them, never rewrite them.

Repo: github.com/Risky-Scout/NCAAF-Predictive-Model Owner: Joey — credentialed actuary and odds compiler. Communicate like a peer: direct, no cheerleading, surface problems immediately.

Non-negotiables

1. If a step's validation fails, STOP. Report the failure and the suspected cause. Never continue to the next step on top of a failed validation, and never weaken a validation threshold to make it pass.
2. Never fabricate data, columns, endpoints, or results. If a file or field is not in DATA_MANIFEST.md, it does not exist locally.
3. The Model Training Pack (.pkl models) is NOT on disk. Any spec reference to it is optional/deferred. Do not block on it, do not stub it, do not pretend to load it.
4. Ship first, optimize later. No premature refactors, no OOP where functions suffice, no new dependencies without asking.
5. Every module runnable as python -m src.<module> must print what it did: rows in, rows out, key stats.

Definition of done (Day 3)

• python -m src.pipeline.predict_week emits a frozen, versioned predictions CSV (model_version, config hash, priors mode, generated_utc embedded) with all six probabilities for every FBS game with both teams rated.
• Walk-forward evaluation on 2022–2024 (per VALIDATION_PROTOCOL.md — the full-season-ratings shortcut is forbidden): win accuracy ≥ 60%, spread MAE < 15, total MAE < 20, margin-residual SD 15–16, calibration slope 0.85–1.15 on the FBS spread market, log-loss beating the coin-flip baseline.
• 2025 holdout run ONCE with frozen parameters; report generated.
• bet_log.csv schema and weekly_report.py in place (paper mode).
• pytest tests/ green incl. determinism test. README documents one-command reproduction. requirements.lock and starter-pack SHA256 manifest committed.
• v1 status: LIVE for pockets passing WEEK1_LIVE_PROTOCOL §1 historical gates (both windows), PAPER for all others; §2 Week-1 modifications active in config; pocket qualification table signed by the owner in PROGRESS.md.