NFL Models — Proof

Verification gates (every cloud run)

• GATE 1 — pytest tests/ incl. the leakage guard (shift(1) contract).
• GATE 2 — scripts/math_checks.py: Phi numerics, cumsum identity (max-diff 0.00e+00 expected), spread-sign on freshly fetched data.
• GATE 3 — scripts/validate_results.py: file presence, row counts, metric ranges, JSON schema. Fail-closed: any gate red = no deploy.
• GATE 4 — post-deploy HTTP checks on the published URLs.

No gate outputs committed yet for the latest run.

How to reproduce (any machine, cloud or local)

git clone <this repo>
pip install -r requirements.txt
python scripts/fetch_nflverse.py
python -m pytest tests -q
python scripts/math_checks.py
python scripts/backtest_markets.py
python scripts/validate_results.py
python scripts/build_site.py --out dist

Embed contract

Any dashboard can embed these views without modifying itself:

<iframe src="https://abet-nfl-models.pages.dev/embeds/models.html"></iframe>
<iframe src="https://abet-nfl-models.pages.dev/embeds/math.html"></iframe>

Machine-readable status: /data/latest.json. No auth, no cookies, no external runtime dependencies (KaTeX is served same-origin from vendor/).

Honest limitations

• Pure models do NOT beat closing lines: spread MAE margin ≈10.6 vs market ≈9.8; totals ≈10.9 vs 10.4. Hybrid variants (closing line as a feature) match the line by construction.
• Prop probabilities vs real book lines are untested (no historical prop lines in nflverse); evaluation is vs the player's own trailing mean.
• Interception-like stats are near-noise (corr ≈0.15) — sigma floors flag them unreliable.