How a prediction is built
- Quarterback first. A starter over his replacement is worth 8–12 points. EPA per dropback + CPOE grades the quarterback; yards do not. An unconfirmed quarterback caps the pick's confidence until he is confirmed.
- Efficiency over volume. EPA per play and success rate, opponent-adjusted (DVOA-style). EPA is the strongest single predictor; success rate corrects for one-off explosive plays.
- Trenches and pressure. Line matchups, pressure rate vs protection, and cumulative injuries — the market prices star injuries fast but underreacts when the losses pile up at non-star spots.
- Situation. Rest, byes, travel, short weeks. Modern home-field is ~1.5–2 points, and less in divisional games. Key numbers 3 and 7 shape projected margins.
- Weather. Wind first — it is the factor that actually moves passing games and totals. Rain and cold alone rarely do.
- Market as a check, not a crutch. The model's number is blended against the line; a sharp disagreement without an identified reason downgrades confidence instead of inflating it.
The full-game audit protocol
After every game, the pick is re-graded from the play-by-play and the full stat line, not the memory of the score:
- Down-by-down efficiency: EPA/play and success rate for both offenses.
- Explosive plays (20+ yards) and who produced them.
- Turnovers — forced or gifted — and the points they actually created.
- Red-zone and third-down conversion, the two spots where games are actually decided.
- Pressure and line play versus what the pre-game read expected.
- Inactives and in-game injuries versus what was priced into the pick.
Each audit ends in one of four verdicts, published on the week's page: Right (read matched the game), Variance (read was sound, one-score bounces decided it), Model miss (the read was wrong on ordinary downs), or Process failure (the information existed and did not reach the pick in time). Variance is not allowed to launder a model miss — the Texans game, outgained 420–237, is filed as a miss.
Challenge log — what changed because of results
| Date | What the games proved | What changed |
|---|---|---|
| Oct 25 | High tier went 1–2: Texans outgained 420–237 by the Giants; Panthers lost 43–13 with 4 turnovers. | High now requires two things: a real quarterback/efficiency edge AND verified ball security on the picked team. One-legged High picks are banned. |
| Oct 25 | 49ers @ Falcons: the "danger game" flag was exactly right — and the pick stayed on San Francisco and lost. | Danger flags flip picks. A flag that does not move the pick is decoration, and decoration is banned. |
| Oct 25 | Breece Hall was ruled out Sunday morning; the Jets pick never got the downgrade and lost. | The morning check runs on time and reaches the pick before kickoff — inactive-driven changes are posted the same morning, and late runs are logged as gaps, not hidden. |
| Oct 25 | Nacua active → Rams scored 28 by halftime; the Raiders upset pick's own condition had fired. | Fired conditions downgrade the recorded pick immediately when inactives post, instead of riding as a footnote. |
| Oct 25 | Denver covered Surtain's absence easily; Broncos depth underrated. | Star-absence discounts get a depth check before they move a number. |
| Oct 25 | Future games were only being picked the day before — too late to track how news actually moves a matchup. | Predictions now open a week out and are adjusted daily, with every change and its reason logged (see Week 8). |
The daily adjustment rule
Every morning and evening, every future game on the board is re-checked: quarterback news, practice and injury reports, inactives leanings, line movement, weather. If the probability moves, the old number, the new number, and the reason are written into that game's log the same day. A pick is only allowed to flip on quarterback news, a star inactive, or a fired pre-game condition — and a flip is announced as a flip, never quietly rewritten. The pick that was live when a game kicked off is the pick that gets graded.