Why we tested our own board
Recruit Intel builds two versions of its Performance Score model: a season board, scored once a player has enough games on the year, and a weekly board, scored off only the games played since the last one and refreshed as new box scores land. Before shipping another week of the weekly board, we tested whether it actually predicts anything, specifically whether the players it favors go on to land a new offer at a higher rate than chance, and how that compares to the season board's own track record. This piece reports what two full seasons of reconstructed weekly boards found.
Two numbers do the measuring throughout. AUC runs from 0.5, a ranking no better than a coin flip, to 1.0, a perfect one, for how well an ordering of players separates the ones who go on to land a new offer from the ones who do not. Lift compares the offer rate inside a ranking's top decile, the top tenth of players by that ranking, against the offer rate for everyone tested; a lift of 2 means the top decile lands offers at twice the base rate.
The weekly board beats chance, comfortably
Reconstructing every offensive weekly board window (quarterback, running back, wide receiver and tight end) from the 2024 and 2025 seasons produces 284,023 player-weeks across 49,500 distinct players; 1.999 percent of those player-weeks land a new Division I offer within 90 days of that week's games. The weekly board's top decile, 28,412 of those player-weeks, lands a new offer for 4.716 percent of them, 2.359 times the base rate (95 percent confidence interval 2.224 to 2.500). Every decile below the top one earns a new offer at a lower rate than the decile above it, a clean, monotone climb from bottom to top. The weekly board is measuring something real.
But it is a much weaker read than the season board
The season-long version of the model was put through the same kind of check, though not an identical test: 23,853 player-seasons, a different unit than player-weeks, scored on a cohort that excludes seniors and against a longer offer-capture window than the weekly test's 90 days. On that basis, where the base rate for a new offer is 6.78 percent, it scores an AUC of 0.7677 and a top-decile lift of 3.584. The weekly board's AUC, on its own 284,023 player-weeks, is 0.667 (95 percent CI 0.6554 to 0.6781), and its lift is 2.359 against the season board's 3.584. Even allowing for the different test design, a single week of games, scored by the same underlying formula, carries much less information about a player's trajectory than a full season does. Neither number is a ranking claim about any one player; both describe how well an ordering of many players lines up with which of them actually land a new offer.
Almost all of what it sees is volume
Most of the weekly board's ability to separate future offer-getters from everyone else is not coming from the model's own weighting. Ranking the same 284,023 player-weeks purely by raw production, ignoring the model's efficiency and touchdown terms entirely, scores an AUC of 0.657 and a lift of 2.252 on its own. The full model adds only 0.0100 AUC (95 percent CI 0.0072 to 0.0133) on top of that. On the 6,821 player-weeks where the model's ranking and a pure-volume ranking disagree about who belongs at the top, the model's pick lands a new offer at a 3.958 percent rate against a 3.064 percent rate for the volume pick, a rate ratio of 1.292 (95 percent CI 1.087 to 1.565). The structural edge from the model's efficiency and touchdown terms is real and holds up where it is tested directly, but it is small next to the size of the raw volume signal underneath it.
Two changes that would help more than the model itself
Two changes to how the weekly board is built, tested directly against the shipped version, each move the needle further than any adjustment to the model's own weights would. First, scoring players on season-to-date production, at the same calendar point a one-week board would be built, rather than only the games since the last board, raises AUC by 0.0621 (95 percent CI 0.0527 to 0.0705) and lift from 2.275 to 2.990 (CI 2.714 to 3.257) on a matched set of 266,027 player-weeks, reaching 3.200 (CI 2.724 to 3.709) once players carrying a prior D1 offer are set aside. Second, every one of the 284,023 offensive weekly rows in this test ran with its opponent-strength term at exactly zero, because no in-season opponent rating exists yet for the year being scored. RI-RF405 also tested substituting the player's own school's prior-season strength of schedule rating, the same field the scorer already joins on, rather than a rating tied to the specific opponent faced that week. That test scores an AUC of 0.6852 (95 percent CI 0.6733 to 0.695) and a lift of 2.516 (CI 2.364 to 2.661), with a usable prior-season rating on 88.4 percent of its rows (300,200 of 339,607). That row count is larger than the 284,023-row offensive set the figures above are built on, and the finding does not state what accounts for the difference, so this piece does not characterize it either. A same-season rating, which could only be computed after those very games are already known and can never be used to build a live board, tops out at an AUC of 0.6915 on that same test. Both changes are identified directly from this test.
It has nothing to say about a player already carrying an offer
For the 15,997 player-weeks in this test belonging to players carrying a prior D1 offer (1,894 players), the weekly board carries no signal at all: AUC 0.5119 (95 percent CI 0.4949 to 0.5277), a lift of 1.052 (CI 0.982 to 1.117) that does not clear 1.0 with confidence, even against that group's own much higher base rate of 25.949 percent for a follow-up offer. A strong week from a player already carrying a prior offer does not predict a follow-up offer any better than chance would. The board's only demonstrated value is finding a player before any offer arrives, back when one is about to arrive; restricted to that pre-offer group, 268,026 player-weeks across 48,054 players, the numbers hold up (AUC 0.6662, lift 2.405, CI 2.112 to 2.713, base rate 0.57 percent), which is the honest lane to read the weekly board for.
One week of games already misses a tenth of the state
The version of the weekly board that actually ships anchors its window to a Friday and runs four days, which scores 254,872 player-weeks at AUC 0.6658 and lift 2.369 (CI 2.227 to 2.539), close to the full seven-day, Thursday-anchored version tested everywhere else in this piece (284,023 player-weeks, AUC 0.667, lift 2.359). The gap between the two is not a ranking problem, it is a coverage one: the shorter window drops every player whose only game that week fell on a Thursday, which is 10.5 percent of 2025's game rows and 9.2 percent of 2024's. A ranking fix would not recover those players. Only widening the window would.
What this means for a reader
Three honest reads. First, the weekly board works, in the specific sense that it beats chance and beats raw volume by a real, if modest, amount, and it is the only board this site can offer before a season board has enough games behind it to build on. Second, it is a materially weaker instrument than the season board, and neither this piece nor the site should describe a weekly appearance the way a season-long result gets described. Third, that is why the weekly board on this site shows a small set of names and their stat lines for a stated date range rather than a numbered order. Choosing which names make that list still runs the full model behind the scenes, but the board itself prints no rank order among them, because a single week's discrimination is not strong enough to defend an ordering in public. A name on a weekly board should be read as a player had a week worth noting, dated to when it happened, not as a ranking claim about anyone else in the state.
How we counted
This piece publishes Recruit Intel research finding RI-RF405 (RPM weekly board mode validated, approved, publishable, validated 2026-09-01), which reconstructs 34 weekly board windows across the 2024 and 2025 seasons from player_game_stats and scores each one through the shipped weekly code path of the RI Performance Score model (score_frame with min_games=1, weekly=True), then measures whether the top decile of each window lands a new Division I offer within 90 days at a higher rate than a volume-only ranking of the same players. Offers are read from v_player_offers_live, the quarantine-safe view; the season-mode comparison figures were reverified live in the same research session rather than quoted from an older publication, reproducing the previously published headline within 0.002 AUC and 0.04 lift. Unless stated otherwise, every figure in this piece covers the offensive families the board scores, quarterback, running back, wide receiver and tight end, 284,023 player-weeks combined. The board's defensive families are scored separately and stay provisional at a materially weaker AUC of 0.5973 on 55,584 player-weeks; the offensive headline figures above, all built on the 284,023-row set, do not include them.
Every rate in this piece is a floor rather than a point estimate: the outcome measured is whether an offer was captured in the source data, not a complete census of every offer actually extended, so lift, a within-week ratio, holds up more reliably across this dataset than any single rate taken on its own. This validation covers two seasons only, 2024 and 2025, 34 windows, and has not yet been tested on a season with a different box-score-collection footprint than those two. A weekly board can also only rank a player who has a box score on file, and that coverage is partial: at best 31.1 percent of the class of 2026 and 17.8 percent of the class of 2027 carry one, so no claim in this piece, or on the board itself, should be read as best in the state, only best among the players we have a box score for.
