Loading 2026. Still showing 2025 · Selection Sunday.
Published defaultNRR weights
35%Quality wins
25%Adjusted record
15%Loss integrity
15%Adjusted performance
7.5%Win control
2.5%Conference title
21/25same Top 25 teams
19/25within three positions
2.6average rank gap
0.94rank correlation
Comparison room
Largest NRR–CFP disagreements
A gap is not proof that either side is correct. Select a card, inspect the résumé and document whether a future revision is justified.
Selection Sunday · Dec 72025 National Résumé Rating Top 25
18-team path2× byeBowl byeHostRoadCC
NRRTeamRecordIndexMoveCFPΔSAROAPQ1
CFP comparison · 4 teams
Official CFP Top 25, below the NRR cutline.
The proposed default
Six scored measures. Quantify the résumé. Explain the judgment.
NRR creates a public baseline; the committee still owns—and must explain—the final order. The weights are 35% quality-win depth, 25% schedule-adjusted record, 15% loss integrity, 15% opponent-adjusted performance, 7.5% win control, and 2.5% conference champion.
Selected measure
Quality-win depth
35% of NRR
What it asks
How strong were the best opponents actually defeated, and where were those wins earned?
How it works
1
Grade every win with 80% of the opponent’s quality at the selected checkpoint and 20% of the opponent’s frozen pregame quality, then account for location and that game’s opponent-adjusted play efficiency.
2
A win over a stronger opponent adds more evidence. Every distinct opponent uses the same rule, so a fourth quality win counts just like the first.
3
Combine evidence from all wins into one depth score. A small smooth tail keeps even weak wins positive while preserving differences between them.
4
Apply a tiny debit to this component only when raw opponent evidence is negligible, the margin is three points or fewer, and the Game Grade is below 10. The victory still counts fully in the record and Schedule-Adjusted Record.
What keeps it fair
The same continuous equation applies to every win, with no flat minimum or hard evidence cap. The narrow-win debit uses raw evidence, a margin of three points or fewer, and a Game Grade below 10; it can remove at most 1.5 QWD points. School name, conference, brand and recruiting reputation never enter the score.
Show technical calculation and complete inputs Optional detail
Exact calculation
Build both the current and frozen pregame Opponent Value snapshots from 45% official standing, 30% schedule-connected performance, 15% schedule-adjusted record, and 10% schedule caliber. Score the result with 80% of the current checkpoint value plus 20% of the frozen game-time value. For an opponent ranked before kickoff, its published standing score is a floor on that frozen snapshot. Unranked opponents receive no poll floor. Rank every FBS opponent and label wins Top 5, Top 10, Top 15, Top 20, or Top 25. Road context is worth up to six points and neutral context up to three, scaled by the blended opponent quality. Build each Game Grade from 85% opponent-and-venue value and 15% game-level opponent-adjusted play efficiency. For that game, subtract the opponent’s EPA/play from the team’s EPA/play, add the opponent’s national baseline, then reliability-shrink the result toward the team’s season baseline before standardizing it nationally at the checkpoint. This game-specific input is distinct from the separate 15% season-wide Opponent-Adjusted Performance component. Keep a smooth positive lower tail in opponent credit instead of clipping every sub-35 opponent to zero. Convert that credit with one continuous, unbounded diminishing-return evidence curve on a 22-point scale, then add a small grade-sensitive positive tail with 0.05 units at Game Grade 20 and no flat minimum. Add the evidence from every distinct defeated opponent and use one smooth saturation curve plus a 1.5-point lower-tail calibration. Only when raw opponent evidence is near zero, the margin is three points or fewer, and the Game Grade is below 10 does the model subtract a small continuous debit. No win receives a smaller coefficient merely because it is fourth or fifth.
Complete inputs
Wins only
80% evolving checkpoint opponent quality
20% frozen pregame opponent quality
Pregame AP standing floor for ranked opponents
No poll floor for unranked opponents
45% official standing inside each opponent snapshot
Curved Top-25 standing scale
Smooth below-threshold opponent credit
Capped formula baseline for unranked opponents
30% schedule-connected performance
15% schedule-adjusted record
10% schedule caliber
No NRR-to-NRR circular input
Weekly national Opponent Value order
Top 5 / 10 / 15 / 20 / 25 tiers and cumulative counts
The weekly official CFP checkpoint remains the largest single Opponent Value input, but it no longer overwhelms completed performance and record when a credible opponent sits outside the poll. The formula still does not use one team’s NRR Index as the main evidence for another team’s NRR. Venue credit is continuous and opponent-scaled; a weak road win cannot receive the same adjustment as an elite road win. Elite evidence remains nonlinear and unbounded, while the gentler scale limits one marquee result’s dominance. A smooth lower opponent-credit tail and grade-sensitive evidence tail keep weak wins positive without flattening their differences. The exception-only debit still checks raw opponent evidence, is smooth and tiny, and is isolated to Quality-Win Depth; it never converts a win into a loss or removes the win from Schedule-Adjusted Record. Four equivalent quality wins receive four equivalent evidence contributions. Below-line support is disclosed separately, rematches occupy one opponent entry, and the same conference-neutral rule runs for every team.
Method and evidence
Transparent when you want it. Quiet when you do not.
The technical record stays public without crowding the rankings themselves.
How version 0.57 works 11 steps
01
Results only
Completed games create the rating. No preseason rank, brand value, recruiting score, or future result enters NRR.
02
National play-level fit
Completed plays estimate how well every offense and defense performed against the same national network. A soft-schedule deduction fades as a team proves itself against stronger, better-connected opponents. School and conference names never enter the calculation.
03
Schedule-adjusted record
Sixty-five percent is the team’s actual record. Another 17.5% asks how that record compares with one fixed Top-25-caliber benchmark facing the same opponents and locations. The final 17.5% measures the full schedule’s challenge; every opponent counts, while the hardest games carry more context. Conference-title losses count as full results here, even though Loss Integrity separately limits their downside.
04
Quality-win depth
Every win receives an Opponent Value, a Game Grade, and distinct positive evidence. The added tail is 0.05 × exp((Game Grade − 20) / 80): it equals 0.05 units at grade 20 and varies with the grade, rather than imposing a flat minimum. A separate narrow-escape debit still tests raw opponent evidence before that tail is added. Opponent quality is repriced with an 80/20 blend: 80% reflects current-season evidence, while 20% preserves what the opponent looked like before the game.
05
Peer head-to-head
A direct win adds a small 0.35 raw-composite tiebreak only when the teams have the same record and sit within one point in the frozen provisional résumé Index. The narrow rule can separate those two peers without becoming a global bonus over unrelated teams with better records.
06
Loss integrity
Every loss uses the same 80% evolving / 20% game-time Opponent Value as Quality-Win Depth, then adds margin, game state, venue, and repeat-matchup evidence. A bad loss costs more than a competitive road loss to an elite opponent, but no loss earns positive credit. Additional losses compound the damage.
07
Win control
Final margin supplies most of this 7.5% measure, with fourth-quarter checkpoints showing whether the winner stayed in control. Kneels, nullified plays, and late snaps after published blowout thresholds are removed.
08
Conference champion
A completed conference champion receives 100 in a binary 2.5% measure; every other team receives 0. There is no runner-up partial score and no conference-reputation multiplier.
09
Record foundation
The six measures pass through one record-first credibility factor. Strong evidence can improve a résumé, but it cannot make a multi-loss team equivalent to a team that consistently won. Soft schedules matter more as losses accumulate.
10
One NRR Index
The six weighted measures and the narrow peer head-to-head tiebreak produce one public Index: 20 + 0.85 × raw composite. That number determines the published NRR order. The committee still owns the official ranking and must explain material departures instead of hiding judgment inside an unpublished formula.
11
Weekly rebuild
The complete calculation is rebuilt from the games available at each release. Year, week, components, score, rank, and movement all come from the same repeatable code path intended for the 2026 live season.
Why 0.05 is not a minimum Worked example
The v0.57 formula adds a small amount of evidence that varies with the Game Grade. The added amount is 0.05 at grade 20; lower grades receive less and higher grades receive more.
Tail-only illustration: hold raw opponent evidence at zero to isolate the added amount. These are hypothetical inputs, not team ratings.
Game Grade 0: 0.0389 evidence units.
Game Grade 10: 0.0441 evidence units.
Game Grade 20: 0.0500 evidence units.
Game Grade 30: 0.0567 evidence units.
Archived example: Indiana’s win over Indiana State has a Game Grade of 14.19 and 0.0471 evidence units in the 2025 Selection Sunday archive. That published value is below 0.05. It includes raw opponent evidence as well as the tail.
A narrow-escape debit is a separate calculation using raw evidence before the tail is added. The win still counts in the completed record. This example explains the existing formula; it does not validate the ranking method.
The October 2026 frozen property review found score-homogeneity failures and fixed-context weak-win QWD decreases. Limited synthetic positive-response checks passed; proofs, complete-input refits and future validation remain open. See the property review in the Evidence center for exact assumptions and counterexamples.
Official NRR starts with the first CFP ranking. The 2026 tab is a labeled early-season preview using AP standings as its reference; that preview does not satisfy the official publication gate. Performance Lab remains a separate experimental view.
Version 0.57 tests twelve CFP-window rankings and a 2026 early-season preview. The v0.55 opponent-revaluation archive remains available as a prior-method record. A truly unseen season remains future validation work.
Non-FBS opponents receive a low fixed strength floor until their complete schedules are added.
Conference champions receive one binary 2.5% measure. That deliberately adds title value beyond the win itself, while the published component prevents it from becoming a hidden or conference-specific bonus.
Opponent-adjusted performance is descriptive evidence from completed plays, not the separate predictive Team Strength Rating proposed for availability and forward-looking context.
Version 0.57 uses 35 / 25 / 15 / 15 / 7.5 / 2.5 for Quality-Win Depth, Schedule-Adjusted Record, Loss Integrity, Opponent-Adjusted Performance, Win Control, and Conference Champion. A small grade-sensitive tail gives every win positive, distinct evidence without a flat minimum; the narrow-escape debit still uses raw opponent evidence. Schedule-Adjusted Record gives 65% to achieved record and 17.5% each to benchmark overperformance and schedule challenge. The curves and weights remain provisional until additional held-out seasons and the live 2026 weekly run are complete.
Record Foundation and Loss Integrity both react after a loss, but they answer different questions: whether the full record remains credible and how damaging the particular loss was. Their combined leverage remains a published sensitivity audit rather than a hidden adjustment.
The record-first design still respects an excellent one-loss season against a weak schedule, but two or three losses against that same slate receive progressively less credibility. Quality-Win Depth and the published schedule diagnostic expose the underlying evidence; the model does not add a hidden school, conference, or Group-of-Five penalty to force a preferred order.
Published score-state filters reduce late-blowout noise but do not identify substitutions or prove when every possession became noncompetitive. Lineup-aware validation remains open.
The ranked-versus-unranked qualification line is intentionally discrete: an official Top-25 opponent receives official-standing evidence that an unranked opponent does not. The historical market audit tests whether the ordering on both sides remains credible, but no point spread changes a score, rank, quality-win designation, loss cost, or schedule value. The same rule applies nationally with no conference, school, brand, or named-matchup adjustment.