An Applied AI engineering portfolio demonstration: real upcoming probabilities, immutable prediction history, current team/player context, postgame scoring, deterministic investigation and separately versioned challenger research. Python produces checksummed static JSON; React verifies and displays it. The frozen schedule model remains operational. Nothing has been deployed or promoted. The approved destination is https://thomashoward.dev/fourth-down/. Start with publication steps and the $0 hosting plan. Model Intelligence includes an interactive seeded simulation and a cited, deterministic research walkthrough; Architecture explains how the design would scale commercially.
Node 22.14.0/npm 10.9.2 were used. From this project:
cd web
npm ci --ignore-scripts
npm run dev -- --port 4175 --strictPortOpen http://127.0.0.1:4175/fourth-down/. The six product views are Upcoming Games, Matchup Intelligence, Forecast History, Model Intelligence, Postgame Analysis and Architecture. Existing 2023 retrospective matchups/explanations/performance remain separately accessible. At kickoff, verified forecasts remain visible as archived pregame predictions. Verified final results and scores appear after reconciliation; this is not an always-live service. Included static artifacts and notices make a clean checkout browsable without Python, raw data or training.
Use Python 3.14.2 with the unchanged requirements and constraints. This machine's existing interpreter is /Users/thomas/Documents/ChatGPT/AI/nfl-predictor/.venv/bin/python; do not alter the original repository/environment.
# Exact current-season acquisition, frozen inference, history and official-final reconciliation:
python -m ml.refresh
# Advisory only; substitute an actual retained normalized snapshot:
python -m ml.refresh --plan --snapshot data/current/2026/snapshot-<sha>.json
# Independent context-only current annotations; no probability adjustments:
python -m ml.context --season 2026 --snapshot data/current/2026/snapshot-<sha>.jsonThe schedule source is season-isolated SportsDataverse NFL NGS; current injury, QB depth-chart and basic team-stat context comes from season-isolated nflverse assets. Each capture records actual update/retrieval times and hashes. No current acquisition opens 2024–2025. Preserve earlier observations: a first capture after the cutoff cannot prove pregame availability. OPERATIONS.md explains failure bounds, freshness, local evidence bundles, final-forecast selection, result corrections, cost and disabled scheduling. CURRENT_FORECASTS.md details the frozen feature contract.
Forecast inputs remain season win rate, scoring margin, recent margin and rest, with exact frozen smoothing/scaling/coefficient definitions. Current context has a separate cutoff and does not change these inputs. Missing injury report times remain unknown; missing reports do not mean healthy; a depth-chart slot does not confirm a starter. Weather, verified news, travel distance and individual QB performance remain unavailable. Real prospective performance is unknown until recorded games finish; the existing schedule's earlier results are not assigned retroactive predictions.
python scripts/test_release.py
# Read-only byte replay; requires retained current source observations and matching code:
python scripts/replay_current.py
# Independent raw-CSV probabilities/factors:
python scripts/verify_current.py
# Separate retained 2010–2023 source only; no historical acquisition or fitting:
python scripts/recompute_validation.py
cd web
npm test
npm run test:release
VITE_BASE_PATH=/fourth-down/ npm run build
VITE_BASE_PATH=/fourth-down/ npm run preview -- --port 4174 --strictPort
# In another terminal, using installed Google Chrome:
TEST_BASE_URL=http://127.0.0.1:4174/fourth-down/ npm run test:e2eThe release runner forbids estimator/scaler fitting and skips the original fit-repeatability test. Raw-data integration tests skip in a clean checkout; synthetic leakage/integrity tests remain active. Frontend tests verify artifact arithmetic and chronology independently. Browser tests cover desktop/mobile, deep links, errors, responsive layout and selected axe checks; they are not exhaustive accessibility certification. Production builds verify every public artifact, original freezes, registry, protocols, dependency notices and Action SHA pins.
Saved challenger experiments remain browsable under CHALLENGER_PROTOCOL.md. Do not rerun training or tune/promote models without explicit approval. The release workflow performs frozen inference and tests only; saved research results need no fitting.
The actual challenger has ten features across seven schedule-derived domains, six chronological 2018–2023 diagnostic folds, independent prior-season calibration, seven domain ablations, baselines and season-week bootstrap intervals. Its 2023 Brier is 0.23228 versus the frozen incumbent 0.23704, with paired 95% difference interval crossing zero. Pooled historical results compare a contemporaneously refitted incumbent recipe, not the frozen model on its training years; calibration and training-window differences prevent a feature-only interpretation. 2023 was already used for model selection. All results remain research diagnostics.
No promotion: a per-season calibration criterion fails, prospective shadow evidence and historical publication vintages are missing, and current 2026 prior strength would require locked 2025 data. MODEL_REGISTRY.json keeps schedule-v1-741cf41752 operational. Further changes require a new versioned protocol, measured evidence and explicit owner promotion approval.
The original five scripts remain unchanged in legacy/, with hashes matching ~/Documents/ChatGPT/AI/nfl-predictor. The owner-selected working project is this separate ~/Documents/ChatGPT/nfl predictor directory. Do not execute legacy baseline/training scripts, ml.data, ml.train or the historical all-season loader: they are preserved provenance, not authorized operational entry points.
The candidate coefficients/preprocessing/features, eight frozen files, 2023 replay, validation evidence and holdout protocol/exposure remain unchanged. The locked 2024–2025 period is partially exposed: previously observed legacy aggregate accuracy 0.5359, Brier 0.2492 and log loss 0.6915 remain disclosure only. No additional protected results were acquired or evaluated.
The latest Simplify and Ship instruction governs this portfolio release. Public deployment, recurring scheduling, personal-site changes, training, protected holdout use and model promotion still require explicit approval. The portfolio artifact template remains inactive. It retains original source evidence in Git and exports a checksummed static snapshot for import into the existing website; no database, paid service or LLM call is required. Earlier RC1 findings remain historical evidence rather than a new release process.
Current schedule attribution: SportsDataverse, repository MIT notice in web/public/CURRENT_DATA_NOTICE.txt. Historical/current context: nflverse, historical schedule contribution by Lee Sharpe, CC BY 4.0. Transformations and predictions are derived modifications. Project source is MIT; the cited public datasets and bundled dependencies retain their own terms and attribution. The local logo candidate uses 32 recognizable team logos from documented Wikipedia/Wikimedia file pages, with accessible text fallbacks. Third-party artwork is excluded from MIT; see asset sources and limited-use assessment. No blanket license or legal clearance is claimed. Full dependency notices ship with the app; see THIRD_PARTY.md. No affiliation with NFL or nflverse. Educational research, not betting advice.