Every February it happens somewhere in California: a lightly recruited senior signs a Division 1 letter, and the story writes itself. He played next to the five-star. We published that story on 2026-08-13, built around a single number: a 3.5 point halo, drawn from what we described as 5,022 prospects. On 2026-09-02 we pulled the piece, rebuilt the entire battery from raw data under our RF pipeline, and registered the result as RI-RF408. The original number was wrong in the direction nobody expected. It was not too big. It was about a third of what is actually in the data, and almost none of what is actually in the data comes from a teammate.

What we got wrong

Re-derived from the same tables, the cohort of California 2-star and 3-star prospects who graduated between 2014 and 2024 comes to 4,698 players across 700 high schools, not 5,022. Of those, 1,858 had a 4-star or 5-star teammate at the same school within one graduating class; 2,840 did not. The halo group signed a Division 1 offer at 73.2 percent, 1,360 of 1,858. The non-halo group signed at 63.1 percent, 1,793 of 2,840. That is a 10.1 point gap, 95 percent confidence interval 7.4 to 12.7, roughly three times the 3.5 points we printed. The window does not rescue the old number either. Widen or narrow the definition of a nearby graduating class and the raw gap runs 9.2 points at the same class only, 10.1 at plus or minus one class, 11.0 at plus or minus two, 11.2 at plus or minus three, which is the window the original piece used, and 10.8 at plus or minus four. The 70.8 percent non-halo signing rate we published does not reproduce at all: the non-halo rate in this cohort is the 63.1 percent above.

The headline was not off by a few tenths. It understated the raw association by a factor of three. That alone should have been the first sign something was wrong with the method, not just the arithmetic.

The real question, and why the raw number is a trap

A pooled gap that large, between two players who carry the same star rating, has one obvious explanation that has nothing to do with either player: they may not be going to the same kind of school. We tested that directly, in ascending strictness. Stratifying by graduating class alone does not close the gap, adjusted to 9.4 points. Stratifying by CIF section does not close it, adjusted to 9.7 points. Stratifying by rating decile inside the star tier does not close it, adjusted to 5.6 points. Holding the school itself constant does. Among the 159 California high schools that produced at least one halo player and at least one non-halo player in this window, 2,431 players, the gap goes to -0.1 points, 95 percent confidence interval -3.8 to +3.6. That interval sits on zero, and the point estimate is on the wrong side of it.

Every stricter version of the test agrees. Exact matching on class year, star tier, position group and school blue-chip band together comes to +0.5 points, confidence interval -2.5 to +3.5, on 2,035 players. Five different definitions of a Division 1 outcome (any commitment or signing, dated events only, signings only, dated events in a plausible year for the class, and Power 4 commitments or signings specifically) produce five raw gaps between 5.6 and 10.1 points, and five within-school gaps between -1.3 and +3.1 points. Not one of the five clears zero once the school is held constant.

The placebo that settles it

A California high school career runs four years. The widest gap between two graduating classes that could ever have shared a season, a locker room or a single practice is three. So we asked what happens to a player's signing rate when the nearest blue-chip at his school graduated four or more classes away, meaning the two of them could not have overlapped at all. Among 629 such players the signing rate is 69.3 percent, 436 of them. Among 2,211 players at schools with no blue-chip within three classes in either direction, it is 61.4 percent, 1,357 of them. That is a 7.9 point gap, 95 percent confidence interval 3.7 to 12.0, between a player and a teammate he could not possibly have met. A meaningful share of the raw 10.1 point gap is reproduced by a player who was never on the roster with him. That is not a teammate effect. That is the school.

The claims we tested one at a time

The original piece made several specific claims beyond the headline. We re-ran each one against the same 4,698-player cohort.

Three of eleven classes flip negative, and 2024 does most of the lifting. Not reproduced. One class of eleven flips negative, and it is 2014, at -5.9 points, not three. 2024 is not the outlier: 2021 (+18.6) and 2023 (+18.5) are as large or larger, and every class from 2018 forward runs between 9.8 and 18.6 points. What actually moves across that decade is our own rating coverage, which fell from 43.7 percent of the rostered senior class in 2015 to 6.2 percent in 2023. The step change in the year table is a step change in who we hold a rating for, not a step change in recruiting.

The gap is largest for interior offensive linemen and athletes. Not publishable, in either direction. The finding re-derived the position breakdown and then ruled it not reportable in any form, so we are not printing those cells and we are not printing a replacement position table either. The reason is sample size: split 4,698 players eleven ways, then hold the school constant inside each position, and most cells run into the dozens of players, with intervals wide enough to contain almost any answer. Those are not findings. They are noise with a plausible looking number attached. Two things can still be said without the table. The athlete cell shows no halo we can measure. And the interior offensive line figure we originally published as a single range was never a range: it was two separate measurements of the interior line, taken on two different definitions of the group, that disagree with each other and were printed as the two ends of one spread.

In Los Angeles the halo is worth 17.3 points; in the Central section it reverses to -9.5. Not reproduced, and the original source row disagreed with itself here. Los Angeles re-derives at +12.8 points, not +17.3. The Central section re-derives at -0.2 points, not -9.5. The other sections land between those two: Southern at +11.5, North Coast at +10.7, Central Coast at +8.6, San Diego at +7.6 and Sac-Joaquin at +2.5. Stratify by section and the adjusted gap is 9.7 points across nine strata on 4,634 players, so region explains essentially none of the raw association.

It disappears at higher ratings. Refuted, and backwards. The raw gap grows with rating: 5.5 points for 2-stars (n=1,042), 10.1 for 3-stars (n=3,656), 14.3 for 4-stars and 27.8 for 5-stars. The finding stores no cohort size for the top two tiers, so read those two as direction only, and their intervals say the same thing: 7.9 to 21.4 points for 4-stars, 5.9 to 52.3 for 5-stars. Once the school is held constant, only the 4-star tier survives as a real signal, at +7.4 points, confidence interval 1.7 to 13.1, on 149 players. That is the opposite of what we published: whatever is left of the halo lives at the top of the rating scale, not the bottom, and it is a single lead for a future pass, not a finding we are reporting today.

What actually survives

Two things, and both are about the school.

Players at schools that have produced blue-chip recruits sign a Division 1 offer at 73.2 percent. Players at schools that have not sign at 63.1 percent. That gap is real and it is durable: the halo group's schools produced an average of 7.52 blue-chip recruits over the decade we studied, against 0.90 for the other group's schools. Their other 2-star and 3-star players, not just the one player you would point to, sign at 73.1 percent against 64.2 percent. Being at the right California high school is worth roughly ten points of Division 1 probability. Standing next to the right teammate, once you separate the two, is worth nothing we can distinguish from zero.

The honest ceiling on a teammate effect, if any is left at all, is 3.6 points: the upper bound of the interval around the within-school estimate on those same 159 schools. The 3.5 points we originally printed as the effect itself is, almost exactly, that ceiling. What we published as the finding was the largest value the data can still permit, not the value the data actually shows, which is -0.1.

Why the bigger raw number is not the story

The raw gap in the current data is three times larger than what we printed in August. The easy version of this piece points at that bigger number and calls it confirmation. It is not confirmation. It is the same mistake the original piece made, at a larger scale, for the same reason: a real, large, well measured difference between two groups of California high schools, misread as a difference between two players. A school that produces blue-chip recruits recruits well for reasons that have nothing to do with any one teammate, and every player at that school, rated or not, inherits some of that advantage in exposure and evaluation. That advantage is real. It was never the teammate's to give.

This rebuild went through a three reviewer check before publication. Two of three reviewers independently reproduced the raw gap, the near zero result once the school is held constant, and the placebo above. One reviewer flagged, and did not resolve, two open questions: the underlying data mixes several recruiting feeds with different signing rate baselines, and roughly a third of the cohort's school assignments are not independently verified, with a higher unverified share among halo players than non-halo players. Neither concern changes the near zero result once the school is held constant; every reviewer, including the one who raised these questions, reproduced that result under at least one version of the test. But neither concern is closed yet, and a version of this finding beyond California should wait until they are.

Coverage, stated plainly

This is a California-wide claim, so the coverage has to be on the page. Only 12.4 percent of the 42,058 California high school players on file in these graduating classes carry any composite star rating at all, and that share fell from 43.7 percent in the 2015 class to 6.2 percent in the 2023 class. No sentence in this piece describes "California 2-star and 3-star prospects" as a population. Every number describes the roughly one in eight California high school players we hold a rating for, and the composition of that one in eight has shifted every year for a decade. It is also not a random one in eight: a player with a blue-chip teammate is 3.74 times more likely to carry a rating at all than a player without one, 32.1 percent against 8.6 percent. Being near a blue-chip is, before anything else, part of what puts a player into a dataset like this one.

What we did not re-check

This pass covers California only. The original piece also claimed a much smaller halo in Texas, a ratings inflation effect worth roughly a tenth of a star, and gaps in three-year college retention between halo and non-halo signees. None of those numbers came out of this rebuild, and none of them are cleared to run again until each one gets the same treatment this piece got.

Limitations

The cohort here is the rated set, not the population of California recruits; see the coverage note above. The within-school test uses 159 of 700 California high schools, the ones that happen to have produced both a halo player and a non-halo player in this window, so the near zero result is measured on a set of schools that lean larger and better known than the median program. A graduating class within one year of another is a proxy for two players overlapping on a roster, not a direct measurement of it; transfers, redshirts and reclassifications can break that proxy in either direction. Position and section tags come from source data and were not independently re-verified. This rebuild's cohort combines several recruiting data feeds with different historical signing rate baselines, and about a third of the cohort's school assignments have not been independently verified, with a higher unverified share among halo players; both were flagged in review and neither is resolved. Full methodology, every intermediate control and the reviewer disagreement live in the underlying finding, RI-RF408.