I built the model, the application around it, and the pipeline that feeds it. Three people pay for it. What I would show another engineer is not on the pricing page: it is the harness that scores a shadow model against the live one, game by game.
A screenshot of a winning slip says nothing about the ones that lost. All 1,882 selections are on the board. Complete, or it is not a record.
A service that needs you playing every day cannot afford a skip tier. This one has one. It flags the games it did not want played, and 618 of them sit on the board, graded like everything else.
A pick made before the starting pitcher is confirmed is a guess with a number attached to it. Every selection here waits for confirmed lineups and starters. Early is worth less than right.
A good fortnight looks exactly like a better model. Ask v8. Every pick it makes carries a version number and a date, and without those two fields nobody can say whether v9 improved anything or simply ran hot for a week.
The Strategy Lab filters settled picks by market, tier, confidence, odds, side, result, date and team. The engine filter carries four positions: v8 live, v9 shadow, the games where the two agree, and the games where they split. A new version runs against real slates before it is allowed to pick anything. Then it earns the slot.
Every pick lands in one of four tiers and the model is allowed to say do not play. High runs 64.2% across 299 games. Skip runs 47.6% across 618. Both sit on the front page with their game counts attached. Check the ladder rather than believe it.
Accuracy by confidence runs 74.2% on unit plays, then 64.2%, 62.3%, 55.5% and 47.6% down through skip. The ladder is monotonic. That is the part that makes a confidence number worth printing, and the game counts next to each row are 62, 299, 477, 488 and 618.
The season page carries one row per day, split into high, mid, low and skip, going back to May 1. Season totals read 53.5% across 1,418 games with a best streak of 12. Every selection is written the moment it is made. None are edited afterward.
Team Performance ranks every club by how the model did on the games where it backed them. Philadelphia sits at 49-18 across 87 games for +17.6% ROI. Arizona is 34-20 for +21.4%. Ten backings is the floor.
Every market carries a target of 67% on its highest confidence picks, and any market sitting under that target is queued for a rebuild. The changelog dates each one: a v9 moneyline engine in preview, the World Cup model rebuilt to v2, over/under rebuilt to v4. Closing line tracking is in there too. It did not survive. Built, run, and pulled back out once it turned out to say nothing useful about baseball.
Reports are long pieces on what the record actually says, published whether or not the answer flatters the model. One is titled Why we lose to the closing line. The average loss is 1.66%. Another measures where the edge shows up across the first five innings.
Happy to walk through this one screen by screen, in the order it got built.