How a number becomes a pick

What markets does it price?

Player props first — receiving yards, rushing yards, receptions, passing yards, passing touchdowns: exactly the markets whose prices this site archives, no more. The reasoning is structural: sportsbooks post thousands of these with less attention per line than a spread gets, so if a model has any edge anywhere, this is where it would live. Stated at its true confidence: that is industry consensus and structural reasoning, not an academic finding — no peer-reviewed study of NFL prop efficiency exists, and most sources asserting props are soft have a commercial interest in saying so. Game totals and sides get a published model number every week as reference output, and sides are essentially never played: they are the most efficient market in US sports and the one where we have no informational advantage. Anytime-touchdown prices are archived alongside the rest, but that market sits on the never-play list — the archive exists so the record can eventually prove it should not be played, not so a pick can be published in it.

How does a number become a pick?

Seven steps, per candidate market, in this order:

1. A model probability. The probability the side hits, from an explicit conditional distribution — negative binomial for counting stats, gamma or lognormal for yardage. Never a mean-versus-line comparison: yardage is right-skewed and the mean lies.
2. Calibration. The raw probability is shrunk through a per-market calibration fit out-of-sample on prior seasons and frozen before Week 1. There is no in-season refit on results — that is required statistically, and it is why this site never claims the model learns or recalibrates.
3. The market's fair probability. A consensus of at least three books, with the vig removed multiplicatively from both sides. The comparison is always against a de-vigged number, never a raw price.
4. The edge. Our calibrated probability minus the market's fair probability, in probability points. This gap is the entire product.
5. Uncertainty. A standard deviation combining model variance and cross-book market variance. It must be validated empirically against realized outcomes, not asserted — and that check will be published next to the calibration chart.
6. The real hurdle. Break-even is computed at the best available price, not a flat 52.4%: −110 needs 52.38%, −115 needs 53.49%, −120 needs 54.55%. Props usually trade −115 to −125, so the honest hurdle is 53.5–55.6%.
7. Probability the edge is real. The probability the calibrated edge clears the actual price, and the expected value at the best price. These two numbers, with the gap, are what every published card shows.

What stops a number from becoming a pick?

Universal gates — fail any one and the market is not published, whatever the edge says:

• The market bucket has at least 200 graded historical observations behind its calibration.
• The calibration slope clears a floor — a bucket the model cannot forecast is not published.
• At least three books are quoting the line.
• The player passes a mechanical eligibility rule (snap share or target share over the trailing three games, a minimum of graded observations this season) — "key players, not the whole team", expressed as a rule so it can be audited rather than judged.
• The player's status is resolved at release, or the pick is routed to a Held queue that shows the market and the trigger — never a side — until it resolves.
• A raw edge above 10 probability points is quarantined, not promoted. This is the winner's-curse gate: a huge apparent edge is far more often our bug than the market's mispricing, and published research on betting models found realized profit peaks at moderate expected value rather than growing with it. The biggest number the model ever produces is the one it trusts least.

How are picks ranked?

Published picks land in one of three conviction tiers, assigned mechanically from the size of the gap, its uncertainty, and the probability it clears the price. The tier names are still an open decision; whatever they end up being, they will never imply certainty or an instruction to act — no name from the tout vocabulary. Tiers exist to rank conviction and to break the record down, and they will never become a paywall: the moment revenue depends on how many picks land in the top tier, the boundaries acquire a commercial incentive, which is why this product is free.

How is every pick graded?

• Graded at the worst price at or better than the published floor observed in the capture window — never at our own released price, because grading at your own number is the standard way picks services manufacture a record.
• If no price at or better than the floor ever traded, the pick is graded no-action.
• Our own error voids the pick with a correction row that leaves the original standing. Nothing is ever graded at a number that never existed, and nothing is ever silently edited.
• Pushes, voids, DNPs and postponements are explicit statuses, never dropped rows.
• Every market priced and rejected is published too, with its reason code. Publishing the rejects is what proves the selection was not retrospective.

What does the language model do — and not do?

It never produces the probability, never chooses the side, and never ranks the card. Those come from the fixed pipeline above, and any of them coming from a language model would be theater. What it does do: turn unstructured text — beat reports, practice participation — into typed, timestamped notes, where every note carries a source link or is dropped, and no note may contain a number that is not already in the structured payload. A short prose line on a card passes a machine gate that extracts every numeral and checks it against the payload; if the gate fails, the card ships without prose. No pick depends on a model-extracted fact until that extraction pipeline has a measured accuracy record of its own — news is color, not an input.

Where does the data come from?

Play-by-play, usage, snap counts, schedules and official injury designations: data from nflverse (CC BY 4.0), modified. Forecast data from NOAA/NWS; weather data for internationally hosted games from MET Norway. The odds and prices behind these pages come from a licensed odds aggregator whose terms permit displaying derived values; its raw feed is never republished. Book lines shown anywhere on this site — including the draft tool's sportsbook columns, which come from a separate third-party aggregator — are periodic snapshots, never live odds. None of these organizations endorse this site or its output.

The full sample-size argument — what a public record of this length can and cannot demonstrate — has its own page, published before the first pick.