What "shutting out a top offense" was worth: exactly nothing

Recruit Intel's weekly performance score, RPM, has a schedule term built into it from the start. The formula's opponent bonus, worth as much as seven and a half index points in either direction, is supposed to reward a player who produced against a strong opponent and discount one who padded a line against a weak one. From August 12 through August 22, every board Recruit Intel published ran that term at exactly zero, on every row, because the number it needs, an adjusted team rating for the current season, cannot exist until the season the rating describes has finished. The rating is an iterative solve over a completed schedule, and 2026 does not have a completed schedule yet.

That is not a bug in the sense of broken code. It is a structural gap: the model asks a question the data cannot answer this early. This piece is about the fix Recruit Intel shipped on September 1, what it actually measures, how much of the board it moved, and the honest size of what it improved, published as research finding RI-RF402.

Two ladders: who did you play, and how good were they

The fix needed two separate answers for every game in the window. First: which school did this player actually play. Second: how strong was that school, using the only rating available, one from a season that already finished.

The first question turned out to be harder than it sounds. A player's own box score row carries an opponent field, but on its own that field only resolves 9.2 percent of the scored window (1,763 of 19,063 rows). Cross checking against the opponent's own scoreboard entry for the same date adds another 76.6 percent (14,611 rows), for a combined 85.9 percent of the 19,063 scored rows resolved. A further 0.7 percent (137 rows) turned up more than one plausible opponent on the same date and were left unresolved rather than guessed at; 13.4 percent (2,552 rows) had no resolvable opponent at all and scored neutral.

The second question uses the opponent's own most recently finished season, looking back at most two years: 2025 first, then 2024. Across the offensive rows the prior applies to (n=9,778), 65.1 percent (6,368 rows) scored off a 2025 rating, 1.5 percent (142 rows) off 2024, and 33.4 percent (3,268 rows) had no rated season in that two year lookback and stayed neutral. A third rung reaching back to 2023 was measured and left unshipped: it would have added another 1.6 percent of the raw window (559 of 36,030 rows) at a weaker year over year correlation than the two year lookback already gets, so the limit is a measured choice, not an arbitrary one.

Does a year old rating still mean anything

Before trusting a stale number, Recruit Intel checked whether it carries forward at all. Adjusted overall ratings persist from 2024 to 2025 at a Pearson correlation of 0.707 across 8,376 schools rated in both seasons; reaching back one season further, 2023 to 2025, that drops to 0.576 across 8,205 schools. The ratings are not noise from one year to the next, but they do decay.

The more direct test is whether last season's rating predicts anything about this season's football. Across every 2025 scoreboard game where the opponent carried a 2024 rating (n=72,662), points scored correlates with the opponent's prior season defensive rating at r=0.302, against r=0.481 for the same season rating, so the prior recovers roughly 63 percent of the live signal. Restricted to the early season window the weekly board actually runs in (n=14,717), that falls to r=0.250. A prior season rating is a real but weakened signal, and the finding does not claim otherwise.

Does it make the board better

Signal is not the same as improvement, so the model was tested the way a ranking should be tested: against something the prior itself cannot influence. The target is each player's own production for the rest of that season, a separate index built from volume, efficiency and touchdowns with no opponent term in it anywhere, so it cannot be contaminated by the thing being measured. The season opening board, built on games from August 28 to September 5 in each season, with and without the opponent prior, from a single database pull so the only difference between the two versions is the prior itself, was ranked against that later production using Spearman rank correlation, on two separate seasons.

Across the offensive families the prior applies to, rank correlation with rest of season production rose by 0.0076 in 2025 (0.4637 to 0.4713, 95 percent confidence interval 0.0042 to 0.0111, n=13,262) and by 0.0103 in 2024 (0.4245 to 0.4348, confidence interval 0.0064 to 0.0140, n=9,526). Broken out by position, quarterbacks clear zero comfortably in both seasons (n=2,847 in 2025, n=2,141 in 2024); running backs also clear zero in both seasons but by a narrower margin in 2025, where the confidence interval's lower bound sits at 0.0022 (n=4,882 in 2025, n=3,715 in 2024); receivers and tight ends move in the same direction in both seasons but their confidence interval touches zero in both, so that particular gain counts as directional, not established. The honest description of the whole effect: the board gets a little better at finding players who keep producing. It does not get dramatically better, and it should not be sold as if it did.

What failed: crediting a defender for a good offense

The obvious mirror of this idea is to credit a defender for the offense he faced, the same way an offensive player gets credit for the defense he faced. Recruit Intel built that version and measured it the identical way. It failed in both seasons: rank correlation with rest of season production fell by 0.0128 in 2025 (n=3,121, confidence interval negative 0.0205 to negative 0.0055, clear of zero) and by 0.0061 in 2024 (n=1,377, confidence interval negative 0.0184 to 0.0052).

The explanation came after the measurement, not before it. Mean tackles per game stay close to flat across deciles of opponent offensive strength in 2025: 6.1 against the weakest offenses and 6.0 against the strongest (n=3,121). The 2024 gap is wider, 6.2 against 6.8 (n=1,377), but the direction is the same: no reliable signed relationship for a plus or minus correction to fix. A great offense scores fast and runs few plays; a bad offense goes three and out and also runs few plays. That is why the defensive family keeps its opponent bonus at zero by design, not by omission. The rejected calculation is still stored on every row so the decision stays checkable without rebuilding anything.

How many boards actually changed

The prior first went live on the window opening August 28 (18,583 gated players across 38 states). Measured against the identical pool scored without it, state level top five cross family boards changed membership in 25 of the 38 states, 65.8 percent, with 34 of 190 total slots changing hands, 17.9 percent. Four more boards kept the same five names in a different order; 9 were untouched. State top three boards changed in 19 of 38 states; state top ten boards changed in 26 of 38. Every national cut, top five through top one hundred, changed.

Two swaps from that window show what the correction is built to catch. In Arkansas, a receiver's 122 yard night against a defense rated well above average took a board slot from a longer receiving night against a dead average defense. In Florida, a 209 yard receiving night against a well above average defense displaced a 210 rushing yard, 20 receiving yard combination against a well below average defense. Neither swap is close on raw yardage alone; both read as the right call once the opponent is priced in.

What is still missing

The prior is not full coverage. 6,510 of 19,063 scored rows in the launch window carry it, 34.1 percent of everything scored and 66.6 percent of the offensive rows it applies to. The rest score neutral, at the midpoint of the bonus range, not zero credit. That missing third leans very slightly toward the honest side of the gap: among schools rated in 2025, the 742 with no usable 2024 prior average an overall rating of 0.46, against 1.35 for the 8,376 that had one, a gap of about 0.05 standard deviations. A neutral fill is mildly generous to a player who happened to face an unrated opponent, not a hidden penalty against him. The prior also does not appear to be quietly reshuffling the board toward players it can measure: the share of state top five slots held by a player carrying a usable prior moved from 32.1 percent to 36.8 percent, close to that group's 34.1 percent share of the full pool.

Two structural limits stay in place regardless of coverage. The rating behind the prior describes last year's team, not this year's, so it is blind to a coaching change, a wave of transfers, or a jump between classifications. And roughly half of every cross family board is defensive rows that carry no schedule adjustment at all, sitting next to offensive rows that do, because the defensive version of this idea failed its own test rather than because no one tried it.

Sourcing

This piece publishes Recruit Intel research finding RI-RF402, derived from player_game_stats, school_game_scores and team_adjusted_ratings in Recruit Intel's canonical database as of September 1, 2026. Every figure above traces to that finding's stored fields or to its banked source file, social-assets/ri-research/RF-20260901-weekly-rpm-opponent-prior.md, both independently reproducible against the same tables. The underlying scoring model is Recruit Intel's Performance Score, RPM; this finding changes only how the opponent term is priced, and does not touch the weighting of volume, efficiency or touchdowns within it.

Correction, September 4, 2026

Two figures from the first version of this piece have been cut. Both were re-derived on September 4 against RI-RF402's own queries, neither reproduced, and the standing rule is that a number which will not reproduce comes out of the text rather than stays in it. The first was a raw count of resolved rows in the opponent resolution paragraph. That count was a sum of the two rungs the finding records rather than a figure the finding itself stores, and a fresh run of the finding's resolution query now returns a different split between those rungs, because the opponent column on the box score table has been rewritten in place by later stats sweeps. The share the count described, 85.9 percent of the 19,063 scored rows, is the finding's own recorded figure and stands unchanged. The second was the receiving yardage of the displaced player in the Arkansas board swap. The only Arkansas row carrying that yardage in this window was created on September 2, after the finding's September 1 pull, and its opponent carries a different rating from the one the finding names, so it cannot be the row the finding describes. The yardage has been removed rather than replaced, because naming the displaced row exactly needs a rebuild of both arms of the September 1 board comparison, which this pass did not run. Everything else in that passage, including the 122 yard night and both Florida figures, reproduces exactly and is now registered in RI-RF402 under the block named for this article.