What this page is
A shortstop who can actually play shortstop is worth more than a first baseman who hits the same. Every scout knows this; the site did not. This page is about four years’ worth of failed attempts to fix that — condensed into a few weeks — and the one structural change that made the signal appear immediately, along with an honest account of why it had been hiding.
Update · 16 August 2026. Two things changed since this page was written, and the ranks quoted below are today’s. First, the model no longer counts minor-league seasons a player played after he had already established himself in the majors as part of his own prospect record — those seasons exist only because he reached, so using them is a look at the answer. The refit that required moved every rank on the site; the two illustrations below moved up (Fernández to 127th, Molina to 1,436th), and the complaint this page opens with is now clearly separated on the live board: Jesus Made sits 8th among all hitters, Luke Adams 138th. Second, and unrelated to the model: a display bug was found and fixed that left the career-defensive panel blank on every player page for first, second and third basemen. The underlying position values were correct throughout and no score was affected — the column names were being silently rewritten when the database was assembled, so the page looked up a position that no longer existed under that name. If you looked at a first baseman’s page before 16 August and saw no defensive games, that is what you were seeing.
Update · 17 August 2026. A further, unrelated correction: minor-league records that began before this database’s coverage does were being scored as complete ones and are now excluded from the coefficient fit and reference pool — see Where the Data Begins. None of the four names below is a 1977-floor record, so none is directly touched by that fix, but the refit it required moves the board again: Fernández 90th, Molina 976th, Jesus Made 8th (unchanged), Luke Adams 165th, out of 39,413 total. The positional mechanism this page is about is untouched. (As of 21 August: Fernández 87th, Molina 976th, out of 39,440.)
Update · 21 August 2026 — the glove now counts designated-hitter time, and reads from one shared position record. Two changes, and the second is the one that matters. First, where the position term gets its games: it used to read the raw minor-league fielding file directly, which has no designated-hitter rows at all — DH is not a fielding position, so a fielding table simply doesn’t contain it. It now reads the same per-season position record the rest of the site uses, which derives DH by subtraction (a player’s batting games minus his fielding games) and fills two gaps the fielding file has: the Dominican Summer League, missing from roughly 45% of Rookie-level seasons in 2010–13, 2015–17 and 2019, and every season after 2025, which the file does not cover at all — a player who debuted this year previously had no glove on record whatsoever. Second, DH itself became a rated position, at −11.4 runs per 162 games.
Held out across the era split, this made the model more accurate, not less: rank correlation with real career WAR rose from +0.233/+0.361 to +0.236/+0.361, both bust-calibration scores improved, and the refit raised its own weight on the position term from +0.70 to +0.82 — the model’s way of saying the re-sourced number carries more signal than the one it replaced. Ranks quoted below are today’s.
Two things were tried and rejected on the evidence, which is worth recording because both sounded obviously right. Removing the minimum-games floor — 50 career defensive games, below which a neutral value is used — cost accuracy on both era arms and cut the model’s weight on the term to +0.24: a position estimated from one or two games is noise. And giving the generic “OF” tag a value of its own turned out to be measuring an era, not a position: before 2009 this database records 1.86 million outfield games with no left/center/right split, so pricing them all as right field stamps one number across three decades. Generic outfield games are shown on player pages but carry no value until they can be split properly.
Update · 11 August 2026. The contact statistics in the companion paper were rebuilt on a sounder footing, and because every term in the model is fitted jointly, the position term was re-measured alongside them. It still clears its placebo cleanly in both directions, but its measured contribution is smaller than first published — the figures below are the corrected ones. The reading has not changed, only sharpened: position is real, and it is the lesser of the two fixes.
The complaint
Jesus Made is a switch-hitting shortstop and a consensus top prospect. Luke Adams is a right-handed third baseman who most evaluators expect to end up at first base. On the old index the two sat side by side near the top of the current-prospect list — the corner bat, if anything, a step ahead of the shortstop.
Nothing about that is defensible, and it isn’t a rounding artifact. It was the model doing precisely what it was built to do: read two minor-league batting lines, notice they were comparable, and rank the players accordingly. Position never entered the calculation.
It was worse than blind — it was backwards
The natural assumption is that a model with no defensive input is simply neutral about defense. Measured against the real careers of the players who actually reached the majors, that turned out to be wrong. Among big-leaguers, the site’s old index ran backwards against the defensive half of their careers — a negative correlation of about −0.15 to −0.23, depending on whether you measure it straight or by rank. The better a player’s glove and position, the lower, on average, the index placed him. (Across the whole prospect pool the correlation is roughly zero — but that is only because the four-in-five who never reach the majors have no big-league defensive record to be measured against. Among the players who do reach, the bias is real.)
The mechanism is not mysterious once seen. Premium defensive positions — shortstop, catcher, second base — are physically demanding, and the players who hold them tend to be smaller and less powerful. The index rewards power. So it was quietly penalizing exactly the players whose defensive value was highest.
The illustration is stark, and it is not one player. Ozzie Smith — the glove that defines the type, a career built almost entirely on defense — is in this database, but barely: he spent one season in the minor leagues and came to bat 333 times before the Padres decided they had seen enough. That season was 1977, which is the year this database’s fielding coverage begins after, so he has no fielding record here at all and the position term cannot see him. He is scored on his bat alone, and he ranks 6,149th because of it. There is no fix for this at the model level; the data does not exist. His descendants filled the basement instead. Tony Fernández, a shortstop with a 45-win career, ranked 11,371st of 39,365 on the old index; Yadier Molina, a 42-win catcher, 20,684th. These were among the largest misses the model made, on exactly the positions that are hardest to play — and the rebuilt index lifts them, Fernández to 127th and Molina to 1,436th, though a defense-first catcher with a modest bat still tests its limits.
Four things that didn’t work
Attempt one: measure fielding skill directly. Errors, assists, putouts, double plays, range — all of it, from newly acquired minor-league fielding records covering half a million player-seasons. Nothing survived. Minor-league fielding statistics are dominated by how the official scorer felt and how many balls happened to be hit at someone; the skill signal inside them is faint enough to be unusable at this scale.
Attempt two: add a positional bonus to the score. Straightforward, and wrong in a way that took a while to see. The site’s index is a percentile, not a quantity of runs. Adding a positional run value to a percentile is like adding someone’s height to their weight — the arithmetic runs, the result means nothing. This version did technically improve total-value accuracy, but only by badly degrading the model’s accuracy about hitting, which should have been the giveaway. A change that buys one thing by breaking another usually indicates a units error rather than a trade-off.
Attempt three: keep two numbers instead of one. Fit the combination properly rather than stacking it, and let a bat score and a value score coexist. This genuinely worked, and it removed the false trade-off from attempt two entirely. But it only worked among players who had already reached the majors — which is not a prospect list. A prospect list has to be right about players before anyone knows whether they will play a game.
Attempt four: fit total career value across every prospect at once. The honest version of the question, and it made the defensive signal vanish. Scrambled position assignments scored as well as real ones. On that evidence, defense was dead.
Why it kept vanishing
It was the shape of the equation, not the data.
Roughly four out of five minor leaguers never appear in a major-league game. When you fit one equation for career value across everybody, that enormous population of zeros dominates it — and a factor that only matters after a player arrives gets averaged into nothing by the overwhelming majority for whom it never gets to matter at all.
Defense wasn’t absent. It was being diluted by the four-fifths of the sample — the players who never reached the majors — that it had no business being measured against.
Career value splits naturally into two questions that are not the same question:
How likely is he to make it? × How good is he if he does?
This site is genuinely strong at the first — that is what it was built for and what it has always done well. It is weak at the second. And defense belongs only to the second. A shortstop’s glove does not much change whether he reaches the majors. It changes a great deal how valuable he is once there.
Split the equation along that seam and the signal appears at once. It has been there the whole time.
What that does
Position values were not imported from a published table. They were derived from the real career records themselves — regressing actual career positional value against games played at each position, over roughly six thousand major leaguers. What came back, per full season:
| Position | Value | Position | Value | |
|---|---|---|---|---|
| Shortstop | +9.6 | Center field | +0.5 | |
| Catcher | +8.6 | Right field | −7.6 | |
| Second base | +4.2 | Left field | −7.9 | |
| Third base | +2.7 | First base | −8.2 | |
| Designated hitter | −11.4 |
Nobody told the model that shortstop and catcher are the hard positions. It worked that out from what happened to the players.
The designated-hitter row is the one exception, added 21 August 2026: it is not fitted, because a designated hitter plays no defense and so has no defensive record to regress against. It is placed below first base in the same proportion FanGraphs’ published adjustment places it below first base, which is the only defensible way to price a non-position. Two generic tags this database also carries — a bare “OF” or “IF” with no specific spot — have no value at all and are weighted zero; see the 21 August update above for why assuming “OF” means right field made the model measurably worse.
Applied to the original complaint, among the rookie-eligible hitters in the database at the time: Jesus Made, the shortstop, ranked 15th; Luke Adams, the corner bat, ranked 105th. The two the old index had sitting together — corner bat ahead — were ninety places apart, the shortstop decisively in front.
Update · 16 August 2026. Several unrelated corrections have landed since this page published — most recently, removing a reliability discount that was reducing trust in a player’s own batting-average and walk-to-strikeout numbers for prospects with a short minor-league record (see Punished for Making Contact). None of them touch the positional-value mechanism this page is about — the position run-rates in the table above are unchanged — but they move the overall ranking the three anchors below sit in, re-verified against today’s data (20,958 rookie-eligible hitters, up from 20,916): Jesus Made 3rd, Josué De Paula 16th, Luke Adams 29th. The qualitative point stands — the shortstop still sits decisively ahead of the pure bat, and the bat-only corner outfielder still ranks in the top tenth of one percent despite grading near the bottom on the glove term this page is about — but the specific gaps have closed: Adams is no longer ninety places behind Made, and De Paula is no longer “a hair behind” him. Read the rank numbers elsewhere on this page as the 7 August 2026 snapshot they were measured on.
Update · 17 August 2026. One more unrelated correction, covered in Where the Data Begins: minor-league records beginning before this database’s coverage does are now excluded from the coefficient fit and reference pool. None of the three anchors below is such a record, so none is directly touched, but the refit it required moves the ranking again (20,969 rookie-eligible hitters, up from 20,958): Jesus Made 2nd, Josué De Paula 15th, Luke Adams 25th. Same qualitative point, same three names, ranks tightened slightly further.
Update · 21 August 2026. After the position term was re-sourced and DH added (see the top of this page), the same three anchors, now out of 20,996 rookie-eligible hitters: Jesus Made 2nd, Josué De Paula 13th, Luke Adams 26th. The shortstop still leads, and the corner bat still finishes inside the top tenth of one percent — the point this section makes is unchanged.
The part that matters more than the promotion
A defensive adjustment that simply punishes corner bats would be a different bug, not a fix. It doesn’t.
Josué De Paula grades near the bottom of the entire population on positional value — a corner-outfield profile, the exact type this change is supposed to cost. He still ranks in the top tenth of one percent of rookie-eligible hitters (13th of 20,996 today — see the update above). His probability of reaching the majors is about 97%, and that is the half of the equation the glove doesn’t touch.
This is the behavior the split makes possible and a single blended number cannot. The glove lifts premium-position players. It does not sink a bat that is good enough regardless. Those are two different operations, and only a two-part equation can perform one without performing the other.
The whole-pool measurement nearly erased it
Fit across every prospect at once, the glove almost vanished. In the forward era split it added just +0.006 to the ranking — roughly a seventh of what it would prove to be worth — because a term that only matters after a player arrives was being averaged across the four-in-five who never do. It cleared its shuffle test even there, but at a size no one would build a feature around. This is the same dilution that had sunk the first four attempts: a real, post-arrival effect measured over a population that is mostly players it can never touch.
Re-tested where position actually acts — among the players who reached the majors — the signal held clean. Across the era split it added +0.014 and +0.007 to the held-out ranking, and none of the eighty scrambled versions beat it in either direction (p = 0.012). Small, but real, and this time measured where it lives.
Is it real, or is it circular?
A fair objection remains: career value contains a positional adjustment by construction, so a model that predicts a player’s future position and is then graded against career value is partly grading itself. Position predicting career value is arithmetic, not a hypothesis.
The real hypothesis is narrower — whether a player’s future position is knowable from his minor-league one. So the impact model was refit using players’ actual major-league positions: perfect knowledge of where they ended up. On the identical 2,988-player sample, that perfect knowledge added +0.015 and +0.009 to the ranking — and the minor-league estimate, which added +0.027 and +0.018, did better in both directions. Minor-league position predicts major-league position at 0.83. A shortstop’s pedigree travels with him even when he ends up somewhere else — which is why the signal is recoverable at all, and why it is not merely the built-in positional adjustment reflected back at the model.
The honest limits
This is position, not fielding skill. The model now knows that shortstop is harder than first base. It still has no idea whether a particular shortstop is any good at it. Attempt one — measuring the glove directly — failed, and nothing here rescues it. What the model has is the value of the position, not the quality of the player standing in it.
It is the smaller of the two fixes. The contact-hitting statistics covered in the companion paper are worth more and reach more of the list. Defense earns its place in the model; it is not the main event, and the vividness of the Made-versus-Adams result should not be mistaken for the size of the measurement.
Hitters only. Everything above concerns position players. Pitchers are unaffected.
The general standard
Every scoring change here faces the same three checks, and this one is the reason the three exist. Named real players, start to finish. Fit on history it can see, tested on history it cannot, both directions reported separately. Eighty randomized versions of itself, because beating nothing proves nothing.
The lesson worth carrying past this particular change: a signal that only operates under certain conditions has to be measured under those conditions. Four attempts failed here, and only one of them failed because of the data. The other three failed because of the equation they were poured into — and each time, the empty result looked exactly like evidence that defense didn’t matter.