The rules are frozen so the record means something

An AI model projects NFL player performance and compares its projections to DraftKings player-prop lines in three markets: passing yards, rushing yards, and receptions. The picks with the largest gap between projection and line — at most four per slate day, from that day's earliest kickoff window — are posted publicly. Ranking is mechanical. There is no human cherry-picking which names go out.

Everything is frozen roughly 90 minutes before kickoff: the model's inputs, the DraftKings lines being played, and the picks themselves. Picks post to X before kickoff, or they are voided. A post made after kickoff never counts, and nothing is ever backfilled.

Every pick is flat one unit risked, at the posted price. At −110 a win pays 0.909u. Unders settle at the under's posted price. Push is zero. If the player doesn't play — inactive, out, or the game is lost to circumstances — the pick is a void, not a loss and not a win. Results come from official box-score statistics the next morning.

Losers are never deleted, reworded, or hidden. Every pick on this site links its original X post so the timestamp can be checked against kickoff by anyone.

Win rate over a small sample is mostly noise. The sharper question is whether a pick beat the number the market closed at — closing line value. If we take Over 249.5 at −110 and the market closes Over 249.5 at −125, the close implies the Over was more likely than the price we paid: positive CLV.

We compute it by removing the bookmaker's margin (“de-vigging”) from both the entry and closing prices, then taking the difference in implied probability. When the closing line lands on a different number than the entry, the comparison is modeled to the entry number using a standard distribution for that market. Every graded pick shows its CLV; the sitewide average and the share of picks that beat the close are on the ledger.

Anything produced outside the regular season — tooling rehearsals, sample boards — is labeled PRACTICE and is never mixed into the record's numbers. The record starts with NFL Week 1.

A season of daily cards is roughly a hundred-something graded picks. That's enough to show our work; it is not statistical proof of an edge, and we won't pretend otherwise. The point of this site is that you never have to take our word for anything — the receipts are the product.

Pipeline version: recipe-v1|prompt-v1|gpt-5.6-luna|t0|effort-none