One Feature Did
All the Work.

On July 10 we locked sixty-five call-up probabilities, the largest cohort this model has published. The window closed October 8. Six of the sixty-five reached the majors, and v0.5 scored a Brier of 0.0746 against a naive baseline of 0.0838.

That is a pass. It is the first time in three checkpoints the model has beaten a constant guess, and the bucket ordering came back monotonic after breaking in September. Both results are real and both are reported below.

They are also both an accident of who was in the cohort. The rest of this page is about why.

The result that matters

All six call-ups came from Triple-A. Six of the eleven Triple-A players were promoted, 54.5 percent. Of the other fifty-four players, across nineteen at Double-A, twenty-three at High-A, eight at Single-A and four at the complex level, zero were promoted.

So we built a one-feature model to compete against ourselves. It knows nothing except whether a player is at Triple-A. It predicts 54.5 percent for Triple-A players and zero for everyone else. On this cohort that model scores a Brier of 0.0420. Thirteen features, four sequential multipliers and a refit probability mapping scored 0.0746.

The one-feature model is 44 percent better than ours.

Inside Triple-A, the ranking is not forwards

The level split is where the comparison above gets its edge, and a fair objection is that the one-feature model was fit on the same sixty-five rows it was scored against. That objection is correct, and it is why the next test is the important one.

Take only the Triple-A players, where call-ups actually happen, and ask whether our numbers rank them correctly. Below are all eighteen Triple-A predictions v0.5 has made across both cohorts, sorted by our own probability, highest first.

The top half of our own ranking went five for nine. The bottom half went five for nine. The two largest numbers the model has ever published, Thomas White at 62 percent and Charlie Condon at 48 percent, both missed. The three smallest Triple-A numbers in the set, Walker Jenkins at 23, Joshua Baez at 22 and Michael Arroyo at 21, all hit.

Formally, the AUC inside Triple-A is 0.433 on this cohort and 0.356 across both v0.5 cohorts. An AUC of 0.500 is a coin flip. Eighteen rows is a small sample and we are not going to claim the ranking is reliably backwards on that evidence. What eighteen rows across two cohorts does support is the weaker and more damaging claim: there is no evidence the ranking is forwards, and that ranking is the entire product.

The headline AUC of 0.874 does not contradict this. It is produced by the model sorting Triple-A players above teenagers in the complex leagues. That requires no model.

Where the calibration landed

Mean prediction was 18.6 percent against an observed base rate of 9.2 percent, so the model ran high by 9.3 points. That is the opposite direction from June and September, both of which ran low, and it is not a sign the recalibration overcorrected. It is a sign that the low bands are badly floored.

Thirty-seven players averaged 12.5 percent and produced zero call-ups. The model expected about 4.6 promotions from that group. Nothing in High-A or below should be priced at twelve percent on a ninety-day horizon, and across both v0.5 cohorts exactly one player out of fifty-seven below Triple-A was promoted.

Every prediction, and what happened

All sixty-five rows, as locked on July 10. The notes column carries sourced detail where we researched the outcome individually; a dash means the player finished the season in the minors with no major league appearance and no further detail was gathered.

The misses, named

Carlos Lagrange, 38 percent, injury-status blind spot. Highest number in the cohort after Genao. Lagrange was placed on the injured list with a right shoulder capsular sprain in early July, days before the July 10 lock, and the model did not know. He returned at the end of the minor league season with no time to be an option for New York. Third consecutive checkpoint in which the top miss was a player whose injury predated the run date.

Ralphy Velazquez, 33 percent, performance features do not predict promotion. Hit .293/.381/.555 with 29 home runs at Triple-A Columbus and was named the Guardians' Minor League Player of the Year. Never called up. The model's performance block read this season as strong evidence of a promotion. Cleveland read it as a reason to protect a 2027 roster spot.

James Tibbs III, 29 percent, performance features do not predict promotion. Thirty home runs, 108 walks, .941 OPS, led the Pacific Coast League in extra-base hits, total bases and walks, and the Dodgers passed on him in September anyway. Los Angeles already carries the lowest org coefficient in the model at 0.88 and it was still not low enough.

Jett Williams, 33 percent, injury-status blind spot. Wrist injury at Triple-A Nashville in late July, hit .234/.341/.385 on the season, never added to the 40-man roster.

Harry Ford, 28 percent, eligibility error. Graded yes on an appearance July 17, but he should not have been in the cohort. Ford debuted in September 2025 and had eight games of major league service before this run. A player with prior service who is optioned to Triple-A is a far better bet to appear than a true first-time candidate, and the model has no feature for it. Second eligibility error in two checkpoints after Owen Caissie in September.

the <20% band, 12 percent, systematic overprediction below Triple-A. Thirty-seven players averaging 12.5 percent, zero call-ups. The model should be predicting close to zero for High-A and below, not twelve percent. This band alone accounts for 0.0058 of the Brier score that a correctly floored model would not have paid.

A row that should not have been there

Harry Ford was scored at 28 percent as a call-up candidate. He appeared for Washington on July 17 and the row grades yes. He also should not have been in the cohort at all. Ford debuted for Seattle in September 2025 and had eight major league games behind him before we scored him; he was optioned to Triple-A Rochester to start 2026 and recalled when Drew Millas needed surgery.

A player with prior major league service who is sitting at Triple-A is a different population from a prospect awaiting a first promotion, and a much better bet to appear. The model has no feature for prior service, so it priced him like a prospect.

We are grading it as written, because closed runs are never edited and that rule does not get suspended when the error flatters us. This row happens to help the Brier score. It is the second eligibility error in two checkpoints, after Owen Caissie in September, and two in a row is a process problem rather than bad luck.

The window was not ninety days

The 2026 regular season ended September 27. A ninety-day window opened on July 10 and closed on October 8 therefore contained about seventy-nine days of actual baseball, and the last eleven days of it had no games in which anyone could have been called up.

That is a flaw in the run design, not in the grading. Every number in this cohort was quoted against a ninety-day promise that the calendar could not keep. The effect is small and it pushes in one direction: it makes the model look worse than a correctly scoped window would, because the denominator of opportunity was shorter than advertised. It is listed as a v0.6 fix below.

Where the record stands

Three cohorts, 107 graded outcomes. Treat that number carefully, because it mixes model versions. Sub-Batch A was scored under v0.4; B and C under v0.5. As an all-time published record the 107 outcomes produce a Brier of 0.1469 against a naive 0.1791. As an evaluation of the model that is actually live, the honest set is the seventy-nine outcomes in B and C: Brier 0.1034 against naive 0.1199, AUC 0.868, and a level-only benchmark of 0.0687 that still beats it.

The training gate for starting work on a trained v1.0 classifier was fifty resolved outcomes. At 107 that gate is cleared. It was a sample-size condition and nothing more, and what this checkpoint found changes what v1.0 ought to be trained to do.

Six rows that should not have been on the board

The checkpoint also swept the live PIX board, and six of the thirteen remaining rows did not belong there. None carried a CUP prediction, so no published score moves, but all six were public as unpromoted prospects.

• JR Ritchie (ATL), unlogged debut. Debuted April 23, 2026. Sat on the prospect board for 172 days as a major leaguer. Logged October 12.
• Ethan Salas (SD), unlogged debut. Debuted September 2, 2026. Logged October 12, 40 days late.
• River Ryan (DET), prior major league service. Debuted July 22, 2024 with Los Angeles. Never a call-up candidate. Retired from the board.
• Logan Henderson (MIL), prior major league service. Debuted April 15, 2025 with Milwaukee. Retired from the board.
• Jonah Tong (NYM), prior major league service. Debuted for New York in 2025, on the roster and at Triple-A Syracuse in 2026. Retired from the board.
• Elmer Rodriguez (NYY), active major leaguer. Four separate stints with New York during the 2026 season. Retired from the board.
• Aidan Miller (PHI), ungraded closed window. Sub-Batch A row at 22 percent whose window closed June 18. The board row stayed open for 116 days after grading. The published Sub-Batch A record already counted it as a miss, so this is a bookkeeping reconciliation and changes no published number. Miller had a back procedure June 5 and finished the season rehabbing at Single-A.

The same gap as the Harry Ford scoring error: nothing in the pipeline checks prior major league service, and nothing checks whether a tracked player has stopped being a prospect. Four of the six rows were players the model should never have been able to see.

The fix is built. A staleness and eligibility audit now runs automatically at the top of every daily and weekly board update. It flags any active row with no logged event in thirty days, escalates Triple-A rows to high severity because a stale Triple-A row is usually a debut nobody logged, flags rows that have never had an event at all, and flags any prediction whose ninety-day window has closed while the row sits open, which is how the Aidan Miller row survived four months past its grading date.

This was proposed in June, again in August and again in September, and written down each time as a thing worth doing. It was not built until it had cost six rows. Its first run flagged one high finding, Seaver King at Triple-A with no event in fifty-three days, which checked out clean against the debut record.

What changes

No refit today. Not because the result stung, but because refitting the probability mapping is the wrong response to this finding. The mapping is not the problem. A model whose features carry no information inside the group that matters does not get fixed by adjusting the numbers it maps those features onto.

The v0.6 list, in priority order:

• Rebuild around level and opportunity, not performance. Three cohorts now say level proximity carries nearly all the signal and minor league performance carries almost none. v0.6 should start from level and roster opportunity and require every other feature to prove incremental value against a level-only baseline before it ships.
• Injury eligibility gate. Carried from June and September, still not built, and it produced the top miss for the third time. Any player on the injured list at run date is either excluded or scored through an explicit injury state.
• Prior major league service feature. From the Ford row. Players with prior service optioned to Triple-A are a different population and need either a feature or their own exclusion rule.
• Hard floor below Triple-A. Zero of fifty-four below Triple-A in this cohort, one of fifty-seven across both v0.5 cohorts. Nothing below Double-A should be scored above about three percent on a 90-day horizon.
• Season-length aware window. A 90-day window that runs past the end of the regular season is not a 90-day window. Either shorten late-season cohorts or count games remaining instead of days.

The first item is a rewrite, not a patch, and it is not a decision to make in a results post. What this checkpoint establishes is the standard any v0.6 has to clear before it ships: beat a level-only baseline out of sample, and show positive discrimination inside Triple-A. v0.5 does neither.

The full run record is published at checkpoint_2026-10-08.json. The running scoreboard is on the record, the September checkpoint is here, and the model is documented in Inside v0.5.

Outcomes verified against Baseball America's record of every 2026 major league debut and against individual reporting for each Triple-A row. Brier score is the mean squared error of the forecast probabilities against binary outcomes, per Brier (1950). The naive baseline is a constant forecast equal to the observed base rate of the cohort. The level-only benchmark predicts each level's own observed rate within the cohort. AUC is the probability that a randomly chosen player who reached the majors carried a higher CUP than a randomly chosen player who did not. All headline figures were computed twice by independent code paths and in closed form before publication.