AI and Data Disclosure

A statistical model makes the numbers. An AI language model writes those numbers up in plain English in Willie's voice. Automated checks compare the writing against the numbers before anything is published. Every call gets a permanent receipt and is graded in public after the game.

Licensed and public sports data: schedules, final scores, play-by-play stats, rosters, injury reports and past closing lines.

Betting-market prices: spreads, totals and moneylines from multiple licensed sportsbooks. We keep our own time-stamped record of how they move through the week.

Official and public weather forecasts for outdoor games.

Every Deep Dive records exactly which data it used and when it was pulled. We don't use news feeds, social media, or data scraped from sites that don't allow it. None of our data sources receive any information about you.

Want to know more about where Willie's data comes from? Ask us in writing at hello@wagerwillie.com.

The model rates every team in points based on past results, and adds how efficient each offence and defence has been play by play, the starting quarterback, rest, travel, division games and recent form. For totals it also looks at conditions like the roof and the weather.

It then blends its view with the betting market. In testing, the market's line was more accurate than the model on its own, so Willie's fair line is the market line moved only a small part of the way toward the model.

Cover chances are calibrated. In testing, the raw model was overconfident, so its chances are scaled back based on how earlier seasons actually turned out. That is why most of Willie's cover chances sit close to 50%.

Injuries: the current model version puts a point estimate on missing linebackers and cornerbacks who have been playing regularly and are ruled out or doubtful, because those were the only positions where past data showed a clear effect. Other injuries are described in the report but not given a point value. Starting quarterbacks are handled separately.

Willie only leans on a side when two things are both true: his fair line differs from the market line by at least 1 point, and his calibrated chance of covering beats the break-even for the price (52.4% at standard -110 pricing). If either test fails, he calls it a Fair Line, and where he can, he names Willie's number, the spread at which he would lean.

Most games are Fair Lines. Totals are always Fair Lines in the current version, because the model showed no skill on them in testing. The threshold is fixed in advance and is never lowered to produce more leans. A Fair Line is recorded as a Fair Line, never as a win.

Every Deep Dive includes a game read: who the model thinks will win and a projected score. Willie's Top Picks are the week's strongest of those, chosen by a fixed rule from the model's win chances. A pick to win says who Willie thinks will win the game. It is not a bet recommendation and does not mean a bet on that team is good value.

Hunches are spread calls made by a fixed set of written rules (for example a big rest or travel gap, or extreme weather). They are clearly labelled as lower-confidence calls and kept on their own record, separate from Willie's leans.

The language model receives the facts for one game: the numbers, injury statuses, forecasts and times, already calculated and rounded. It writes the explanation you read in a Deep Dive, in Willie's voice, for fans who are still learning. It is told to use only those facts and to say plainly when something is missing ("The injury report for Sunday isn't out yet") instead of guessing. The name of the AI model that wrote each report is stored with it.

Ask Willie answers are written by the AI when you ask, from a fixed set of Willie's data for the games in your question. The same automated checks run before you see an answer: every number must come from that data, and betting instructions, guarantees and money talk are rejected. If an answer fails twice, you get no answer and your credit comes back. A person does not review each answer before you see it. Answers can be wrong.

Before a Deep Dive is published, software checks it against the facts it was written from. Every number in the text must trace back to a number in the facts. It also rejects banned wording, including guarantees, "locks", "risk-free", chasing losses, profit or bankroll talk and betting units, rejects sportsbook or offer talk inside the analysis, and rejects lean language on a Fair Line game.

If a draft fails, the AI gets one chance to fix the listed problems. If it fails again, it is not published. The same wording checks run again when the page is shown.

It does not choose leans, Fair Line calls, Top Picks or hunches. Those come from the model and fixed rules.

It does not invent stats, records, quotes, reviews or testimonials, and it does not produce Willie's record, which is calculated from graded receipts.

It never sees your account details. No account details, email address, location or browsing activity is sent to the AI. When you use Ask Willie (Willie Pro), the text of your question is sent to the AI, with email addresses and phone numbers removed first, together with Willie's data for the games it's about. Nothing else about you is sent: no name, email address, account ID, plan or location. Please don't put personal details in a question.

Willie's model does not beat the betting market, and he says so.

We tested it on 3,028 NFL games from 2015 to 2025. Willie's fair line missed the final margin by 9.81 points on average, the same as the market's closing line. The model on its own missed by 10.06.

We also tested it against six seasons of real sportsbook line history (2020 to 2025), keeping the last three seasons locked away until the end. On those 2023 to 2025 games, Willie's early-week leans won 49.0% against the spread (283-294-13), below the 52.4% needed to break even at standard pricing.

These are backtests on past seasons, not live results. Live results, wins and losses, are on Willie's Record.

Injury news breaks at any time. Our injury data doesn't update in real time, and game-day inactives are not captured, so a late scratch may not be in a Deep Dive.

The starting quarterback, weather and lines are projections or snapshots at the time the report was built. They can change before kickoff.

The model only puts a point value on some injuries. Everything else about a missing player is description, not a number.

Data can be late or wrong, and sports have a lot of randomness. Even a real edge loses often, and Willie has not shown one.

The checks confirm the numbers and the wording rules. They can't confirm every judgment in the writing, so the AI can still emphasize the wrong thing or phrase something awkwardly.

Deep Dives are written and published automatically once they pass the checks above. A person does not approve each one before it goes live.

Every week, a review is generated from the database: the record, the biggest misses and the facts behind them, and any data problems. The owner reads it and decides what to fix. Data fixes can ship any time. A new model version must pass tests written down in advance, on games it was never fitted to, and needs the owner's sign-off. Its version number is recorded on every receipt it produces, so you can always see which version made a call.

Receipts are never edited or deleted. If Willie's call changes before kickoff, for example a lean appears, disappears or switches sides, a new receipt is added that points back to the old one, and the old one stays visible marked as updated. The last receipt before kickoff is the one that gets graded.

Deep Dives are not overwritten either. New data produces a new version tied to its own data snapshot. If you spot a mistake, email hello@wagerwillie.com or use "Report a data error" on the Contact page.