Every NFL game on this site gets a projected score for both teams, produced by a model that reads play-by-play data and nothing else — no tipsters, no consensus picks, no hand-adjusted lean. The projection is written down before kickoff and is never edited afterwards, whether it turned out right or wrong.
This page explains the method at the level anyone can check, and links to the raw record so you do not have to take it on trust.
For each matchup the model builds a projected score for both sides from team-level efficiency in recent play-by-play, adjusted for who is starting at quarterback, the opposing defence, home advantage and the forecast at the venue. The two projected scores give the winner and the margin. That is the whole output: a score for each team.
It does not produce a written case for a pick, and this site does not print one. What you see on a game page is what the model computed — the two projected scores, the winner, and when the pick locked. If a paragraph explaining why appeared there, something other than the model would have written it.
A model that reads recent play-by-play has nothing to read in September. For
the first three weeks of a season this one runs on the previous season's
play-by-play instead, adjusted for the quarterbacks actually starting now,
and it says so on every pick it produces: those entries are marked
prior_season_2025 in the public record. From week 4 it runs on
the current season.
Every pick carries the timestamp at which it was frozen, and that timestamp is before the game started. Nothing is added to a game after it kicks off, and nothing is changed after it finishes.
This is the only thing that makes a published record worth reading. A pick written down afterwards is not a prediction, and a record that can be tidied up later is not a record. The lock time is published on each game's own page so the claim can be checked rather than accepted.
No pick is ever removed from the record and no result is ever rewritten. A losing pick stays on the site exactly as it was published. If a correction is ever needed — a scoring change, a feed error — it is added beside the original, never applied on top of it.
The full history is downloadable as JSON and CSV, so you can count the record yourself instead of reading a number somebody typed. The browsable version is at the pick history.
Any player named anywhere on this site is checked against the current season's roster feed before it renders. A name that cannot be matched to a current-season entry for the team it would appear under is dropped rather than guessed at.
Jack Zakaria
https://www.tiktok.com/@picksbyjackzak
A model reads public play-by-play data and projects a score for both teams in every game. The two projected scores give the winner and the margin. There is no tipster, no consensus and no hand-adjusted lean, and the model produces no written rationale -- so this site does not print one.
Before kickoff, every time. Each pick carries the timestamp at which it was frozen, published on the game's own page, and nothing is added to a game after it starts.
No. A losing pick stays on the site exactly as published. Corrections, if one is ever needed, are added beside the original and never applied on top of it, and the full history is downloadable as JSON and CSV so the record can be counted rather than taken on trust.
Straight up. A pick counts as correct when the team it named won the game; a tie is a push and counts as neither. Where a market line appears on this site it is shown for comparison and is never what the record is measured against.
Play-by-play, schedules, depth charts, injury reports and rosters come from the public nflverse datasets. Venue weather comes from Open-Meteo's forecast API. Market lines, where shown, come from The Odds API. Any player named anywhere on the site is checked against the current season's roster feed first, and a name that cannot be matched is dropped rather than guessed at.