This is the scoreboard. Every number the model publishes is locked before the outcome is known, graded when its 90-day window closes, and never edited afterward. The misses stay up next to the hits. If you want to know whether to trust this site, this page is the answer, and it will say so either way.
Fourteen of 28 reached the majors and the bands came out in the right order, with all six top-band players debuting. The probabilities ran low, which produced a 0.269 raw Brier against a 0.250 naive baseline, so the model lost to a constant guess. That result triggered the v0.5 recalibration. Full results ›
Five of fourteen. Brier 0.2375 against a 0.2296 baseline, so v0.5 lost too, and the band ordering inverted for the first time. Thomas White at 62 percent, the largest number the model has published, missed after being shut down with a shoulder injury he already had on the day we scored him. Full results ›
Six of sixty-five. Brier 0.0746 against a 0.0838 baseline, the first checkpoint the model has won, with monotonic band ordering restored. It also found that all six call-ups came from Triple-A and that a one-feature model knowing only that fact scores 0.0420, which is better than the full model. Full results ›
Sub-Batch C closed on October 8 and all 65 predictions were graded. Six reached the majors. The Brier score was 0.0746 against a naive base-rate baseline of 0.0838, so the model beat a constant guess for the first time in three checkpoints, and the band ordering came back monotonic.
The same checkpoint found the more important result. All six call-ups came from Triple-A; zero came from the other 54 players at Double-A and below. A one-feature model that knows only whether a player is at Triple-A scores 0.0420 on the same cohort, which is 44 percent better than the thirteen-feature model. Inside Triple-A, where call-ups actually happen, our ranking carries no information: across all eighteen Triple-A predictions v0.5 has made, the top half of our own ranking went five for nine and so did the bottom half. The full checkpoint, with every outcome and every miss named.
All 79 outcomes scored under v0.5, across Sub-Batch B and Sub-Batch C. These are closed windows, not projections.
| Band | n | Reached MLB | Predicted | Observed |
|---|---|---|---|---|
| 45%+ | 4 | 2 | 50.0% | 50.0% |
| 30-44% | 9 | 4 | 34.7% | 44.4% |
| 20-29% | 27 | 5 | 23.0% | 18.5% |
| <20% | 39 | 0 | 12.3% | 0.0% |
Ordering holds at the top and the bottom band is the clear failure: 39 players priced at an average of 12.3 percent produced zero call-ups. Nothing below Triple-A should carry a double-digit number on a ninety-day horizon, and exactly one player out of 57 below Triple-A was promoted across both cohorts.
Twenty-five of 107 predictions resolved yes, a 23.4 percent base rate across the year. As an all-time published record the 107 outcomes give a Brier of 0.1469 against a naive 0.1791. Read that number carefully: Sub-Batch A was scored under v0.4 and the other two under v0.5, so it is a record of what was published rather than an evaluation of one model. For the model that is actually live, the honest set is the 79 outcomes in B and C, where the Brier is 0.1034 against a naive 0.1199 and the level-only benchmark of 0.0687 still beats it.
Six rows also came off the live board during the October sweep that should never have been on it, four of them players with prior major league service and two unlogged debuts, one of them five months old. None carried a CUP prediction so no score moved, but it was a process failure and the audit that should have caught it is now built and runs on every board update.
The PIX board keeps its own record. When a tracked player debuts, his card is frozen as a receipt: the live number at debut, the number 30 days prior, and the sourced events that moved it. Recent receipts, including Luis Lara, the call-up the daily scan missed and owned in public, live on the board, and every point move ever applied sits in the append-only event ledger.
Nothing is open. All three 2026 cohorts are resolved and the next batch locks when the 2027 season starts. Before it does, any revision has to clear a bar this season set: beat a level-only baseline out of sample, and show positive discrimination among Triple-A players. v0.5 does neither, which is why the next version is a rewrite rather than a patch. The 107-outcome training set for that work is complete and published.
The deeper material lives one click away: the calibration log holds every graded prediction, the June 18 technical report shows the recalibration math, and Inside v0.5 documents the whole model, including what we know is still wrong with it.