Trust
Trust is the model's estimate that this side lands, and it is calibrated against real results. It reads what it says. Nothing is rounded up to look better.
Edge
Edge is what the price is worth: trust x decimal odds - 1. A pick at 55% on 2.00 has an edge of +10%, which means about ten cents expected per dollar staked. A pick at 70% on 1.35 has an edge of -5.5%: it lands far more often and still loses money. That is why the board is not sorted by trust alone.
How the two work together
Edge is the gate, trust is the order. Anything priced worse than the model thinks it is worth never reaches the board, no matter how likely it looks. What is left is sorted by trust, so the most likely of the genuinely well priced picks comes first.
What each number covers
Both numbers describe one pick, not a season. A 60% pick lands about six times out of ten, which means four losses in ten are part of the number working exactly as it reads, and a run of five or six losses is ordinary rather than a sign anything broke.
How AI sports predictions are built, day by day
The models run once every morning and produce every available prop per league. Each row gets a line, the market price at that moment, a trust and an edge. The board is then sorted and published. Lines and prices are frozen at publish time and never rewritten, which is the only way a record means anything later.
How results are measured
Every published pick is graded against the real outcome and kept, which is what the models learn from week to week. That record is an internal instrument, so it stays internal: what the page carries is the model output and the price it was published at, both frozen at publish time.
Reading the charts
Every chart shows the last 15 games. Green bars cleared the line in the direction of the pick, red bars did not. The dashed line is the betting line itself.
What we do not cover
Player props only, for now. Spreads and totals need a different model and it is not built, so those pages are empty rather than filled with guesses.