There is no NFL pick to publish from this source set.
The market rows show where captured spread endpoints differed. They do not contain the model number, fair probability, offered price, or publication record required to turn that difference into a bet.
The Jets-Steelers row is a question, not a side
Jets at Steelers spans -3.5 to 2.5, a 6-point range . Provenance tier: live_pregame from game_odds. That makes the matchup worth a quote-level audit. It does not support the old article's Jets model number, trigger price, or positive-value claim because none of those values has a source in the module.
Panthers at Jaguars spans -2.5 to 1.5, a 4-point range . Provenance tier: live_pregame. The smaller range is still only a range. Comparing the two does not reveal which team covers.
The surrounding board
| Game | Observed home-spread range | Derived range | Provenance |
|---|---|---|---|
| Bills at Browns | -3 | 0 points | live_pregame / game_odds |
| Ravens at Vikings | -3.5 to -3 | 0.5 points | live_pregame / game_odds |
| Chiefs at Buccaneers | -5.5 | 0 points | live_pregame / game_odds |
| Commanders at Lions | -4.5 to -1.5 | 3 points | live_pregame / game_odds |
| 49ers at Chargers | -1.5 to 1.5 | 3 points | live_pregame / game_odds |
| Seahawks at Titans | -4.5 to -3 | 1.5 points | live_pregame / game_odds |
Flat rows are not automatically harder to beat, and wider rows are not automatically model opportunities. The source set has no backtest with an ATS record, window, and sample showing that either rule works.
Market disagreement is not model disagreement
A market range compares observed market quotes with one another. A model gap compares a model probability with a market price. Those are different objects. The old article saw the former, declared the latter, and then named a side. That missing bridge is the entire betting decision.
Even a well-built model should not use dispersion as permission to ignore uncertainty. A wide range may reflect stale snapshots, different timestamps, copied feeds, or genuine disagreement. The response is to resolve the data shape, not to reward the model for being decisive while the inputs are messy.
The model itself must be frozen before kickoff and calibrated on comparable markets. If the article claims a side based on a spread threshold, the threshold should be derived from the offered price and the model's probability distribution. A naked point gap does not account for juice, push probability, or estimation error.
A pick needs a receipt before it needs a result
The publication record should capture the event, market, side, line, price, book, and timestamp before the game begins. Later grading should compare the ticket with the settlement result and, separately, with the closing market. Without the pregame record, a correct-looking article can always move its threshold after the fact.
No such receipt exists in this module. That means there is no ATS record to report and no claim about historical accuracy to preserve. The absence is not a reason to hide behind generic risk language. It is a reason to keep the pick off the page.
What a publishable pick requires
A real card needs a market identity, identified book, quote timestamp, offered price, model probability, decision threshold, and publication timestamp. The NFL picks page can show that comparison when those fields exist. This article cannot synthesize them from a spread range. Later grading should compare the ticket against the close, the standard closing-line value check.
That contract also makes a future update easy to audit. If a current quote and model output arrive, the post can add the decision without rewriting history. If either input remains absent, the no-pick state names the cure. The reader gets a reproducible boundary instead of a retrofitted story about why the market range “meant” Jets.
The cure
Attach the current Jets-Steelers quote at an identified book and a calibrated model distribution generated strictly before kickoff. If the resulting probability clears the real price after uncertainty, publish the side and log it. Without that evidence, pass.
Model calibration from graded predictions
Calibration points render only when a verified source binds prediction probabilities to settled outcomes for the same observations.
Expected value from graded outcomes
Expected-value cells render only when a verified source binds observed win outcomes to the price paid for the same bets.
