No black box theatrics. Here is exactly what happens between raw NFL data and the pick on your screen.
Every night the pipeline gathers the current week's inputs: the schedule, player-prop lines, injury reports and official depth charts, plus contextual signals. Everything is rebuilt for the live week so picks reflect the roster and situation as they stand now.
A production stack of fourteen models scores the full universe of qualifying props across quarterback, running back, wide receiver and tight end markets. Some markets use a canonical frozen architecture; four are refit weekly as new outcomes accumulate, so the model keeps learning from the season in progress.
This is the part most pick sites skip. A raw model score is not shown to you. Once a pick qualifies, its confidence is remapped through historical accuracy curves, so the percentage you see corresponds to how often comparable picks have actually hit. The scale is consistent in one direction: a higher number always means a stronger historical record, whether the pick is More or Less.
Before each game window, the relevant picks are locked into an immutable pre-kickoff record. If lines move or data refreshes afterward, the update overlays the frozen row rather than replacing it. That means the pick we grade is the exact pick you could have seen before the game — no hindsight rewrites.
Only rows that clear the qualification threshold are published to your board. You are never shown the internal scoring mechanics — just the final, calibrated, frozen prediction, with the player, market, line, side and confidence.
After games finish, results are tied back to the exact frozen pre-kickoff prediction and realized stats. That is what powers the results page — no cherry-picking a friendlier version after the fact.
