How our picks work.

The data we pull.

Every NFL game on the slate, with the inputs that move lines: official NFL schedule and game time, team injury reports as they post, market movement across the board, and live price from 30+ sportsbooks — FanDuel, DraftKings, BetMGM, Caesars, BetRivers, the books that drive market consensus, and the long tail of regional apps that catch line drift first.

That pricing flows in continuously through our odds endpoint and lands in the odds cache: each sport, each market (moneyline, spread, total), each commence time, the best available price we found, the full set of book prices behind it, a 30-minute rolling TTL so we're never quoting stale numbers.

The analysis.

Raw odds are not a pick. For each game we hand the inputs to an OpenAI-assisted analysis pass — the model weighs injury context, market consensus, and best-available line against a structured prompt tuned for NFL, and returns a side, a confidence score (1–5 bucketed from the decimal confidence score), a unit sizing recommendation, and a short rationale.

We capture the modeled pick against the actual sportsbook line available right at that moment, not the line two hours later. That snapshot — the numbers, the market context, the book — gets stored alongside the pick so what we posted is what we later claim we posted.

The lock.

Picks lock at a fixed time before kickoff. Once a pick is locked we don't revise it, don't cherry-pick a better line, don't move units around — what was posted is what was posted. Every pick is auditable back to the odds snapshot and the analysis it was generated from, so the record is the record.

That same immutability is what surfaces the verified track record: every result, win, loss, or push, gets posted back to the picks feed against the original pick, never overwritten, never re-scored.

The cadence.

During the NFL season we post new picks daily — a full slate on each game day, and any mid-week injury-or-line-driven updates flagged clearly. Outside the season the slate is empty and we say so rather than manufacture volume.

Full result history stays live: every prior pick, the line it was posted against, what we hit, what we missed. The record is the proof — and it never gets edited after the fact.

What you see in the app.

Every pick surfaces the same five fields: matchup, side / total / prop, confidence (1–5 with the underlying decimal score), the locked odds, and the unit sizing. That is the verified pick surface — no editorial overlay, no rewrite on the fly.

Below the surface: a full history view, a record view, and the ability to bring the same data into a commissioner league so your friends can either match the model or fade it.

Sample picks in action.

The methodology above is the same one that produced the pick below — locked against real odds from across the sportsbook market, scored for confidence, and settled against the actual game result. This is what a single TorchPicks output looks like end to end.