On the last Friday in August, a high school team in Fort Worth shut out Aledo, twenty six to nothing. Aledo had gone seventeen and two the season before, scoring forty seven points a game. It had not been shut out since 2003.
On the same night, a team in San Antonio won eighty to nothing.
Rank those two by the scoreboard and the eighty point night wins by a distance. Rank them by who was on the other side and it is not close in the other direction. The team Harlan beat eighty to nothing went three and seven last year. North Crowley held the best offense in Texas to zero.
That gap between what the scoreboard says and what actually happened is the whole reason Recruit Intel built the Team Performance Board. This piece is the case for it. It is also the second version of this piece. The first version passed our validation gate and then failed our own adversarial re-read, and three of the things it claimed are corrected below rather than quietly deleted.
The adjustment is one line of arithmetic
Every team in our database carries two ratings for each season, built only from game scores. An offense rating is what a team would score against an average defense. A defense rating is what it would allow against an average offense. They come out of a standard iterative estimator, the kind that has been used to rate teams for decades, with a shrinkage term so that a program with three games does not rocket to the top.
off_i = mu + sum over rated games (points scored on j, minus def_j) / (n_i + 1) def_i = mu + sum over rated games (points j scored on i, minus off_j) / (n_i + 1)
To score a single game we credit a team for the quality of what it faced:
adjusted points for = points scored, plus (league average, minus opponent defense rating) adjusted points allowed = points allowed, minus (opponent offense rating, minus league average)
Subtract those two and the league average cancels out, leaving something a reader can check with a calculator:
adjusted margin = raw margin + opponent rating
That is the entire model. Take the margin, add the opponent's rating. There is no black box, no weighting committee, and nothing we can quietly tune after the fact. It also means the whole thing rests on a single assumption, that one point of opponent quality is worth one point of margin. We tested that assumption. It did not come back as cleanly as we first reported, and the section below is the corrected version.
Testing it on the only thing that matters
A team rating that cannot predict anything is decoration. So we asked the plainest question available: does an opponent adjusted performance tell you more about a team's next game than the raw margin does.
We took every team game in our database from 2015 through 2025 that had a decided next game and an opponent carrying a rating from the previous season. That is 167,745 games on the registered gate run. We sorted opponents into thirds by strength, within each season so that changes in scoring across eras cannot drive the result, and then compared teams that posted the same margin against different quality of opposition.
| Result | Games | Next game win rate, weak opponent | Next game win rate, strong opponent | Difference |
|---|---|---|---|---|
| Lost by 22 or more | 41,197 | 26.5% | 38.6% | +12.1 points |
| Lost by 8 to 21 | 22,432 | 36.1% | 53.9% | +17.8 points |
| Lost by 1 to 7 | 16,490 | 39.9% | 58.5% | +18.7 points |
| Won by 1 to 7 | 18,782 | 46.5% | 59.5% | +13.0 points |
| Won by 8 to 13 | 9,001 | 48.0% | 63.8% | +15.7 points |
| Won by 14 to 20 | 12,464 | 53.5% | 65.7% | +12.2 points |
| Won by 21 to 27 | 12,322 | 55.7% | 67.6% | +12.0 points |
| Won by 28 to 34 | 11,223 | 58.9% | 72.0% | +13.2 points |
| Won by 35 or more | 23,527 | 67.1% | 76.0% | +8.9 points |
Nine bands, nine times the same answer, never by less than 8.9 points and twice by more than 17. Two teams that won by a field goal last Friday are not the same team. One of them is roughly thirteen points of win rate better than the other, and the scoreboard cannot tell you which.
Note what this table is not. It is not a model. It is a count of what happened, sorted two ways, with every cell carrying thousands of games behind it.
How much better, and what out of sample does and does not buy here
Counting is persuasive but it is not a validation gate. So we fit the model version: a score built only from the raw margin against a score built from the adjusted margin, trained on 2015 through 2024 and then scored on 2025, a season the fit had never seen. The numbers below are areas under the curve, where 0.5 is a coin flip and 1.0 is perfect. Intervals come from resampling whole team seasons rather than individual games, because a team's games are not independent of each other.
| Season held out | Games tested | Raw margin | Adjusted margin | Gain | 95% interval |
|---|---|---|---|---|---|
| 2025 | 44,532 | 0.6420 | 0.6641 | +0.0221 | +0.0194 to +0.0251 |
| 2024 | 44,292 | 0.6479 | 0.6647 | +0.0166 | +0.0136 to +0.0196 |
| 2022 | 12,875 | 0.6592 | 0.6781 | +0.0189 | +0.0139 to +0.0243 |
Three separate holdout seasons, three intervals, none of them touching zero. Re pulled from scratch on the evening of 2026-09-01 for this correction pass, the 2025 holdout reproduces at 0.6423 raw against 0.6645 adjusted on 44,289 test games, so the panel moving underneath us by a few hundred rows in a day does not move the answer. Add a control for who the team plays next, which is the single biggest thing driving whether it wins next, and the adjustment does better rather than worse: 0.7663 against 0.7364 on the 2025 holdout, a gain of +0.0300 with an interval of +0.0276 to +0.0323, on 44,532 test games.
Correction, and it is ours. The first version of this piece leaned on the phrase out of sample for those headline numbers, and reported that a flexible baseline with seven bend points bought the raw margin exactly nothing, offered as proof that the adjusted model's edge was not a curve sneaking in through the back door. Both claims are empty. Each of those two headline scores ranks the test season by one number, and ranking by one number does not change when you fit a coefficient to it first. We checked it directly on the 2025 holdout: ranking by raw margin with no model at all gives 0.6423, the trained straight line gives 0.6423, and the trained seven bend curve gives 0.6423, identical to four decimal places. That is not a passed robustness check. It is arithmetic, and we should not have presented it as evidence.
The comparison that really is out of sample is the one with more than one number in it. Fit the raw margin flexibly, with the same seven bend points, and let the model choose its own coefficient on opponent strength instead of forcing it to one. Train on the same seasons, score the same 2025 holdout. That model lands at 0.6639, against 0.6641 for the fixed weight we ship, a difference of -0.0002 with an interval of -0.0009 to +0.0005 on 44,532 test games. Letting the data pick the weight, out of sample, on a genuinely fitted model, buys nothing over the weight we already publish. That is the robustness result worth keeping, and it is the one we should have led with.
Is one point of opponent worth one point of margin
The board asserts that it is. We can check by letting the data pick its own weight instead, reading the ratio of two coefficients in a model that carries raw margin and opponent strength separately. We first reported that the weight of one survives. It does not survive. It is not refuted either. The check simply does not settle it, and here is every specification we ran, all eight, including the one we almost left out.
| Specification | Games | Implied weight | 95% interval | Excludes one |
|---|---|---|---|---|
| Pooled 2015 to 2025, no controls | 125,935 | 0.8165 | 0.7836 to 0.8481 | Yes |
| 2024 alone, no controls | 44,292 | 0.7754 | 0.7231 to 0.8294 | Yes |
| 2025 alone, no controls | 44,532 | 0.8905 | 0.8394 to 0.9516 | Yes |
| Pooled, next opponent controlled | 125,935 | 1.0495 | 1.0229 to 1.0744 | Yes |
| Pooled, next opponent and the team's own previous rating controlled | 105,166 | 0.8600 | 0.8118 to 0.9072 | Yes |
| 2025 alone, next opponent controlled | 44,289 | 1.1100 | 1.0691 to 1.1539 | Yes |
| 2024 alone, next opponent controlled | 44,050 | 1.0333 | 0.9930 to 1.0762 | No |
| Pooled, own previous season rating controlled but not schedule | 105,166 | 0.4734 | 0.4097 to 0.5327 | Yes |
The first four rows are the registered gate run, taken at 13:12. The last four are the evening correction pass, on a fresh pull of the same panel a few hundred rows later, which is why the games column does not always move for the reason a control does. We are printing that split so a reader comparing two adjacent rows knows when the count changed because we added a control and when it changed only because the two rows come from different pulls.
The eighth row is the one we almost left out. It controls for a team's own previous season rating without also controlling for the schedule that team played, which removes most of the variation opponent strength is supposed to explain, because a program's own rating is strongly correlated with the quality of schedule it faces. It is not a serious estimate of the weight, and we are printing it anyway, with that reason attached, because hiding an unflattering specification is the exact error this correction exists to fix.
Set the eighth row aside and seven specifications remain, every one of them a serious fit. Seven of seven intervals exclude one, from both sides. Read without a schedule control and the data wants about 0.8, so the board looks like it over credits the opponent by a fifth. Add the control and the data wants 1.05, or 1.11 on 2025 alone, so the board looks like it under credits. Add the team's own previous season rating on top of the schedule control and it comes back to 0.86. Across all eight specifications, outlier included, the implied weight runs 0.4734 to 1.1100. Across the seven serious fits it runs 0.7754 to 1.1100, and the only one of the seven that contains one is 2024 alone with the schedule control, at 1.0333 with an interval of 0.9930 to 1.0762 on 44,050 games.
What we wrote first was that the shipped weight of one is confirmed at 1.05 once schedule is controlled. That was choosing the specification that agreed with us and calling it the answer. The defensible statement is narrower and duller. Across the seven specifications worth taking seriously, the implied weight is specification dependent, roughly 0.8 to 1.1, and the weight of one we ship sits inside that range without any single fit confirming it. We ship one because it is the transparent choice, the one a reader can check with a calculator, not because the data picked it.
The place a reader actually feels it
There is a sharper test than a ratio of coefficients, and it is the one that maps onto what a board does to a reader. Take every band of equal adjusted margin, the games the board itself is calling the same. Inside each band there are two ways to have got there: a big raw margin against a weak opponent, or a small raw margin against a strong one. If the weight of one were exactly right, both routes would predict the same next game.
They do not. Across eight bands covering 125,124 games, the teams that arrived on the bigger raw margin won their next game more often in every single band, by between 0.7 and 10.9 points of win rate, averaging 5.1. Eight of eight, same direction.
| What the board scored the game | Games in band | Won next, arrived on the big margin | Won next, arrived on the strong opponent | Residual |
|---|---|---|---|---|
| Worse than minus 20 | 25,500 | 31.8% (n=6,532) | 20.9% (n=6,625) | +10.9 points |
| Minus 20 to minus 8 | 16,311 | 39.5% (n=4,394) | 35.0% (n=4,291) | +4.5 points |
| Minus 8 to 0 | 12,411 | 44.6% (n=3,289) | 43.9% (n=3,254) | +0.7 points |
| 0 to 8 | 13,267 | 51.8% (n=3,348) | 47.7% (n=3,488) | +4.1 points |
| 8 to 16 | 13,516 | 54.7% (n=3,599) | 52.5% (n=3,503) | +2.1 points |
| 16 to 24 | 12,533 | 60.8% (n=3,422) | 56.9% (n=3,137) | +3.9 points |
| 24 to 32 | 10,907 | 64.6% (n=2,872) | 59.3% (n=2,848) | +5.3 points |
| Better than 32 | 20,679 | 74.3% (n=5,246) | 65.0% (n=5,414) | +9.4 points |
Each band carries two arms, and the win rates beside the games-in-band count are computed on those two arms, not on the full band. Every cell above carries hundreds to thousands of games behind it.
This is the number we should have printed first and did not print at all. It says that at the level a reader experiences the board, two games it scores as equal are not equal, and the one that got there by beating somebody good is the weaker of the two. The board over credits opponent strength. The coefficient table above is the same fact in a harder to read form.
That is an open calibration gap and we have not closed it. It does not undo the result in the section before it: the adjusted board still ranks games better than the scoreboard does, by a measured and repeated margin, and that is the claim the validation supports. What the validation does not support is the claim that two games the board scores equal are equal. Anyone reading a board should read it as a better ordering, not as a verdict on a tie.
The correction we owe the reader
Our own pitch for this product was that a forty two to nothing win over a doormat should score below a narrow win over a top defense. That is a good line. It is also not what our data says, and we would rather print the correction than the line.
| Kind of night | Games | Average margin | Average opponent rating | What the board scores it | Next game win rate |
|---|---|---|---|---|---|
| Won by 28 or more over a bottom third opponent | 18,610 | +41.5 | -17.4 | 24.10 | 64.9% |
| Won by 1 to 7 over a top third opponent | 6,265 | +4.1 | +20.1 | 24.26 | 59.5% |
The board scores those two nights as equal, to within about a sixth of a point. Reality gives the blowout team 5.4 points more next game win rate, 64.86% against 59.49%. So the blowout does not score below the close win against a good team. It scores level with it, and it performs slightly better afterward.
Here is why that is still the story. On the scoreboard, those two nights are thirty seven points apart. That is the difference between a rout and a nail biter, the difference that fills a highlight package and a local write up. In terms of what the team does next, the gap is five points of win rate. The scoreboard's version of the gap is about seven times too big. The first version of this piece then went one step further and told you the leftover five points was mostly a schedule effect rather than a flaw in our adjustment. We are withdrawing that sentence. The calibration table above says the leftover is exactly the direction and roughly the size of the over credit the board is carrying, so part of it is our model and we cannot currently say how much.
What this actually changes, week to week
A correction nobody notices is not worth publishing. The opponent adjustment has a standard deviation of 16.6 points across 125,935 games. It moves 56.8% of games by more than ten points and 24.7% by more than twenty.
Rebuilt across 128 weeks of history, a top five offensive board built on adjusted scoring shares an average of 2.44 of its five places with the raw scoreboard's top five. On defense it shares 0.59. The defensive board is almost entirely a different board every single week.
The weekend of August 28 shows why. These are the five best defensive performances in the country by adjusted points allowed, with what each opponent did last season:
| School | Final | Opponent | Opponent last season |
|---|---|---|---|
| North Crowley, Fort Worth TX | 26 to 0 | Aledo | 17 and 2, 47.0 points a game |
| Selma, Selma AL | 58 to 0 | Southside, Selma | 13 and 1, 50.1 points a game |
| Christian Academy of Louisville KY | 52 to 0 | Owensboro | 13 and 2, 42.9 points a game |
| St. Thomas Aquinas, Fort Lauderdale FL | 38 to 14 | DeSoto, DeSoto TX | 13 and 3, 51.9 points a game |
| Holy Innocents Episcopal, Atlanta GA | 47 to 0 | Athens Academy | 10 and 2, 41.9 points a game |
And these are the five the raw scoreboard picks, ranked on fewest points allowed:
| School | Final | Opponent | Opponent last season |
|---|---|---|---|
| Harlan, San Antonio TX | 80 to 0 | East Central, San Antonio | 3 and 7 |
| Martinsburg WV | 74 to 0 | Hedgesville | 1 and 6 |
| Duchesne, Duchesne UT | 73 to 0 | Ignacio | 2 and 7 |
| Liberty County, Bristol FL | 70 to 0 | Vernon | 2 and 11 |
| Lawndale CA | 69 to 0 | Roosevelt, Fresno | 12 and 8 |
Not one school appears on both lists. The five opponents on the adjusted board went a combined 66 and 10 last season. The five on the raw board went 20 and 39. Both lists are shutouts or near shutouts. Only one of them is about defense.
On offense the two boards overlap on two of five, and the opponent records run 39 and 23 against 17 and 33.
The view that describes a finished season is not the one that predicts
One more correction, and this one matters to anyone who queries our data directly. The database view that attaches an adjusted score to a game joins the opponent's rating from the same season as that game. That rating was computed using the game being scored, and using every result the opponent went on to post afterward. It is a reasonable way to describe a finished season and it is not a predictor of anything.
We sized it. On the 2025 holdout the same season version reaches 0.6847 against 0.6645 for the previous season version, and that entire gap of about +0.020 is knowledge of the future, on 42,949 test games. The registered gate run puts the same look ahead at +0.0201 with an interval of +0.0178 to +0.0224. So roughly half of that view's apparent edge over the raw scoreboard is not real.
That view is now labelled descriptive only. Every board we publish, and every predictive number in this piece except that one comparison, is built on the previous season rating, which is also the only version that can exist for a season still in progress.
Does it predict anything besides football
Yes, and it is the reason this matters for recruiting. We took every school season in the database, matched it to that program's next graduating class, and asked which measure better identifies the classes that put players on a verified Division One roster. Fit on classes through 2021, scored on 2022 through 2025, covering 31,210 tracked players across 5,215 schools.
Every number is computed inside a single graduating class, because the measured Division One rate falls from 72.7% for the class of 2016 to 14.9% for the class of 2025, almost entirely because attainment takes years to resolve. We do not carry a separate n for each of those ten classes in this finding; the tracked population behind the decline is the same one behind the table above, 6,949 players in the training classes of 2016 through 2021 and 31,210 in the test classes of 2022 through 2025. Pooling across classes would let that decline masquerade as signal.
| Measure | Discrimination | 95% interval |
|---|---|---|
| Raw point differential, the scoreboard | 0.6192 | 0.6089 to 0.6304 |
| Adjusted overall rating | 0.6674 | 0.6551 to 0.6785 |
| School enrollment alone | 0.4652 | 0.4504 to 0.4792 |
The adjusted rating beats the scoreboard by 0.0483, with an interval of 0.0397 to 0.0575, and it wins in all four test classes. Re derived on a fresh pull for this correction pass it reads 0.6673 against 0.6190 on 31,192 tracked players, the same answer to three places and the same gap of 0.0483. Two details are worth more than the headline. Put both measures in the same model and the scoreboard's coefficient turns negative: once you know who a program played, its raw points carry essentially nothing extra. And enrollment on its own sits below a coin flip, so this is not big schools wearing a disguise.
This one is confirmatory rather than new. It re derives, on a longer panel and a different statistic, a result our own methodology work had already established in August. We report it because the re derivation is independent, not because it is fresh news.
What the board cannot do
A model page that lists no limits is marketing. Ours has real ones, and two of them are new since the first version of this piece.
The weight of one is not confirmed. Across all eight specifications we ran the implied weight runs 0.47 to 1.11. One of the eight controls for a team's own previous rating without controlling for schedule, which is not a serious estimate and is printed only so nothing gets hidden; excluding it, the seven serious fits run 0.78 to 1.11 and all seven intervals exclude one. We ship one for transparency. A reader is entitled to know that the data does not single it out.
The board over credits opponent strength where a reader will notice. Among games the board scores as equal, the one with the bigger raw margin wins its next game more often, in eight bands out of eight, by 5.1 points on average across 125,124 games. Treat a board as a better ordering, not as a statement that two of its rows are equal.
Two in five games are missing. For the weekend of August 28 we could adjust 3,653 of 6,106 games, or 59.8%, because the rest had an opponent we do not carry a previous season rating for. Twenty three states cleared our coverage rule. A school absent from a board has not been judged and found wanting. It was not eligible, and every board we publish says so.
The ratings are a year old. In the first weeks of a season there is nothing else to use, so a 2026 game is adjusted by its opponent's 2025 rating. Ratings carry over at a correlation of 0.722 year to year across 6,324 schools, which is most of the way but not all of it. A program that turned over its roster or its coaching staff is mis adjusted, and early season boards are the most exposed to this.
It rewards running up the score. That is what a margin based model does. Harlan's eighty to nothing still reaches the national offensive board even after a ten point penalty for the opposition. The adjustment is a correction, not an eraser.
Where somebody else already rates a team, we no longer claim what we used to. An earlier version of this piece said outside power ratings cover about 55% of team seasons and that we uniquely supply the other 45%. That number does not hold up, and its only source was an earlier, audit-failed methodology page we should not have quoted. Checked live against the outside ratings we ingest: of the 29,404 team seasons in our adjusted-ratings panel, 28,164 (95.8%) have a matching row in that outside table, and 19,019 (64.7%) carry an actual rating value, never below 94.6% in any single year from 2015 to 2025. Outside coverage is not thin. What we have not measured, and are not claiming here, is how much our adjustment adds once an outside rating already exists.
It ranks, it does not forecast. Nothing here supports a win probability, and nothing here says anything about an individual player. A team rating is a statement about a team.
The row that did not survive checking
An earlier version of the board above had a sixth name on it. Southwest DeKalb of Decatur, Georgia beat a school called Carver thirty four to nothing, and our database resolved that opponent to Carver of Columbus, which had gone seventeen and zero the season before. That would have been one of the best defensive nights in the country.
It was the wrong Carver. The opponent was Carver of Atlanta, a different school with the same name in the same state. We carry no previous season scores for that program at all, which means the game cannot be adjusted and does not belong on the board in either direction.
We found this because we check board rows against public sources before we publish them, which is a standing rule here for exactly this reason. Two schools sharing a name in one state is the most ordinary failure in high school data, and it is invisible from the inside. We are printing it because a model page that only shows you the rows that worked is not telling you how it behaves.
One more thing about that Aledo game
Our model ranked North Crowley the best defensive performance in the country that weekend using nothing but a box score and a table of last season's ratings. It did not know that Aledo is one of the most decorated programs in Texas. It did not know the streak.
After the game, reporters covering it noted that this was the first time Aledo had been shut out since 2003.
The model had no idea. It just knew who was on the other side.
Correction, 2026-09-04. One number has been removed from this piece. The sentence describing the iterative estimator used to date the method to a decade. Our research record carries nothing about when that family of estimators was first used, so the date is cut and the sentence now says only that the method is long established. Every other number here was re derived live on 2026-09-04 against the same panel and registered in RI-RF403 as this article addendum, including the nine band table, the eight band calibration table, the adjustment magnitude shares and both weekend boards. All of them reproduce: the two tables to the decimal place printed, the boards row for row and in the same order.
Validation and every number in this piece: RI-RF403, Recruit Intel research registry, gate run 2026-09-01 with a correction pass the same evening. Where the two pulls of the panel differ, both readings are printed. Ratings from our adjusted team ratings model over 2015 through 2025. Board window: games of 2026-08-27 through 2026-08-29.