Current NRR
#1 · Index 87.020Performance 79.6/100
Account
Current. Beta. Candidate.
2026 · Week 7 · Games through October 10. Compare the same résumé under three performance components.
Candidate is development only—not promoted. Its only Index change comes from the direct 15% performance component. Published rankings and Beta remain unchanged.
Validation: 353 games replayed · no promotion
Predict each game’s offensive yards/play differential using strictly earlier dates. The entire target day stays out of the fit. 95 opening games are excluded by a fixed warm-up rule: at least 100 offensive observations must be available.
| Reference approach | MAE | RMSE |
|---|---|---|
| Neutral: predict no difference | 2.136 | 2.731 |
| Simple: prior team differentials | 2.068 | 2.660 |
| Connected opponent effects | 1.993 | 2.561 |
| Connected + sparse weighting | 2.039 | 2.629 |
MAE is the average absolute error; RMSE gives larger misses more weight. The connected approach has lower error in this replay, but the extra sparse weighting worsens it. Conservative does not automatically mean accurate. This tests the reference mechanism—not the candidate’s final grade, ranking order or ability to predict winners.
Further testing: 2024, 2025 and 2026
The same connected model has lower average error than the simple baseline in all three historical replays. Extra sparse weighting increases error in all three. This is a consistent diagnostic signal—not proof that the full candidate ranking is better.
| Season · tested games | Simple | Connected | Weighted |
|---|---|---|---|
| 2024 · 774 | 1.542 | 1.508 | 1.553 |
| 2025 · 794 | 1.683 | 1.641 | 1.689 |
| 2026 · 353 | 2.068 | 1.993 | 2.039 |
2024 and 2025 run through Selection Sunday; 2026 stops at Week 7. Historical inputs are archived but not independently verified. Every prediction excludes the entire target day.
Uniform weighting stress test and remaining gaps
2024 MAE: 0-game prior: 1.508 · 1-game prior: 1.537 · 2-game prior: 1.553 · 4-game prior: 1.577.
2025 MAE: 0-game prior: 1.641 · 1-game prior: 1.671 · 2-game prior: 1.689 · 4-game prior: 1.716.
2026 MAE: 0-game prior: 1.993 · 1-game prior: 2.022 · 2-game prior: 2.039 · 4-game prior: 2.061.
No best setting was selected from this sweep. Keep the four-game model regularization separate from this additional prediction weight; removing the latter does not mean removing all safeguards.
Alternative source scenarios: not tested in this historical replay. The dedicated Week 7 source audit below now rebuilds the four retained alternatives. It does not adjudicate the source conflict or rerun these historical forecasts.
Missing references: unresolved. These sources contain FBS-connected games, not a verified complete FCS-versus-FCS schedule and statistical inventory. No external reference games were added.
Same fixed four-game ridge prior and 100-observation warm-up across all seasons; each target day is excluded.
Historical 2024 and 2025 statistics are archived inputs marked not independently checked, not new source verification.
2024 and 2025 cover through Selection Sunday; 2026 covers only Week 7. Season metrics should not be compared as equally mature samples.
Weight priors of 0, 1, 2 and 4 are a diagnostic sweep. No best setting is selected, no AP or school-specific tuning, and no published grade changes.
The replay predicts yards/play differential, not the final 0–100 component, résumé rankings, or game winners.
Finalized retrospective data and settings developed after seeing 2026 are not an independent prospective holdout.
Alternative-source and full FCS coverage gates remain open. Lock settings and test untouched future outcomes before any promotion.
Download all three seasons, predictions and weighting tests →
Are the gains reliable? Paired date-level checks
Compare errors on the exact same games, then resample whole game dates 2,000 times. Positive gain means lower average error for the connected model than the simple baseline. These ranges describe historical resampling uncertainty—not independent validation.
| Season | Average gain | 95% resampling range |
|---|---|---|
| 2024 | 0.033 | -0.016 to 0.069 |
| 2025 | 0.042 | 0.016 to 0.075 |
| 2026 | 0.075 | -0.009 to 0.181 |
2024 and 2026 ranges include zero. The average gains are encouraging, but we cannot confidently treat them as established improvements. The 2025 range stays positive within this diagnostic. Removing any single date leaves the average gain positive in each season; that is a sensitivity check, not a new independent test.
The extra sparse weighting raises average error in all three seasons. Its date-resampling range stays negative in 2024 and 2025, but spans zero in 2026. No setting is selected or promoted.
Opponent-group checks and test limitations
2024
FBS versus FBS: 714 games · connected gain 0.043 · extra-weighting gain -0.034.
Outside FBS inventory: 60 games · connected gain -0.081 · extra-weighting gain -0.166.
No prior connection: 37 games · connected gain 0.173 · extra-weighting gain -0.168.
1–2 prior games: 146 games · connected gain -0.032 · extra-weighting gain -0.112.
3+ prior games: 591 games · connected gain 0.041 · extra-weighting gain -0.020.
2025
FBS versus FBS: 716 games · connected gain 0.054 · extra-weighting gain -0.036.
Outside FBS inventory: 78 games · connected gain -0.069 · extra-weighting gain -0.153.
No prior connection: 49 games · connected gain 0.096 · extra-weighting gain -0.125.
1–2 prior games: 145 games · connected gain -0.025 · extra-weighting gain -0.099.
3+ prior games: 600 games · connected gain 0.054 · extra-weighting gain -0.029.
2026
FBS versus FBS: 282 games · connected gain 0.128 · extra-weighting gain -0.025.
Outside FBS inventory: 71 games · connected gain -0.136 · extra-weighting gain -0.128.
No prior connection: 53 games · connected gain 0.179 · extra-weighting gain -0.035.
1–2 prior games: 131 games · connected gain 0.056 · extra-weighting gain -0.067.
3+ prior games: 169 games · connected gain 0.056 · extra-weighting gain -0.033.
Date-block bootstrap ranges are approximate retrospective diagnostics, not independent confirmation or probability that a ranking is correct.
Games on a date are resampled together; repeated-team dependence across dates remains.
2024/2025 Selection Sunday and 2026 Week 7 have different sample maturity; no pooled cross-season accuracy claim.
The same 1,921 historical games are reused from the previous report. No new independent games or sources were collected.
Subgroup checks are exploratory; no team, conference or AP-order adjustment is selected.
This audits a yards/play differential predictor, not Current NRR, archived Beta or the candidate’s complete score formula.
Missing-source and prospective gates remain open even when an error range stays above zero.
What must pass before a formula change?
Reproducible historical mechanism tests · Checked. Retain hashed inputs, target-day exclusion and uniform rules.
Verified missing opponent references · Open. Obtain and reconcile source-backed FCS-versus-FCS game statistics; do not invent neutral observations.
Alternative-source sensitivity · Open. Rebuild from every documented game-stat scenario, not archived rank ranges alone.
Full candidate-grade calibration · Open. Validate the bounded game evidence, team aggregation and 0–100 conversion, not just this differential predictor.
Overlap with existing résumé components · Open. Check interaction with schedule adjustment, quality wins and loss integrity in the complete Index.
Untouched prospective evaluation · Open. Lock policy and evaluation criteria before future outcomes; no retuning on the test set.
Only the historical mechanism checks are complete. Next, test the complete candidate grade and its calibration; source coverage, alternative scenarios and an untouched future evaluation remain required. Current NRR stays unchanged.
Download paired tests, date blocks, subgroup checks and promotion gates →
Does the complete grade use the scale well?
The complete Candidate performance grade—not the NRR Index—runs from 30.3 to 84.9 in this Week 7 snapshot. 44 of 138 teams sit between 45 and 55. This is a compression signal to investigate, not proof that grades should be spread out.
| Model | Middle 80% | Median |
|---|---|---|
| Current | 0.0–67.0 | 29.5 |
| Beta | 34.9–93.4 | 71.7 |
| Candidate | 40.4–70.1 | 54.4 |
The 0–100 scale is not a probability or Candidate percentile. It still uses Beta’s inherited display mapping. No replacement scale, grade target or new formula is selected here.
Where compression enters the formula
| Stage | Median | 45–55 teams |
|---|---|---|
| Before bounds / weights | 69.4 | 17/138 |
| After ±3 game bound | 69.3 | 18/138 |
| After opponent evidence weight | 56.0 | 34/138 |
| After team evidence weight | 54.4 | 44/138 |
Most compression appears after the opponent evidence weight, with further narrowing from the team weight. These stages reuse the existing connected fit; they are not replacement rankings. Sparse evidence still requires safeguards, but their full-grade effect needs testing.
No prior-only game: 62 teams · median 54.4.
Some prior-only games: 76 teams · median 54.4.
All games prior-only: 0 teams · median Unavailable.
Schedules, team quality and coverage differ between groups. These differences do not isolate a coverage effect.
Does performance repeat the other components?
Compare performance with each unchanged component across all 138 teams. Rank correlation ranges from −1 to +1; a constant component is unavailable. Shared signal does not itself prove double counting.
| Component | Beta | Candidate |
|---|---|---|
| Quality wins | 0.71 | 0.68 |
| Adjusted record | 0.72 | 0.74 |
| Loss integrity | 0.77 | 0.77 |
| Win control | 0.39 | 0.30 |
| Conference title | Unavailable | Unavailable |
Candidate has substantial overlap with adjusted record and loss integrity. To establish additional value, we still need season/date-held-out comparisons with and without the performance component—not weights chosen to match the AP poll.
Snapshot audit boundaries
This is one retrospective Week 7 snapshot, not a held-out validation of the full Candidate grade.
The inherited Beta normal-CDF scale is a display mapping, not a probability, accuracy percentage or percentile in the Candidate population. No replacement scale is fitted here.
Stage grades expose transformation effects with existing reference fits frozen; unbounded stages are not alternative ranking proposals.
Aggregation omissions reuse frozen checkpoint game traces. They do not refit opponent effects and are not causal win or loss values.
Correlations use all 138 teams, not just the Top 25. They are descriptive overlap, not proof of double counting or independent value. Constant components have unavailable correlations.
Coverage and game-count groups mix schedules, team quality and source availability; differences do not isolate a coverage effect.
Later games within this checkpoint inform opponent fits; missing reference games and source alternatives remain unresolved.
The combined evaluation below now implements historical full-grade reconstruction and scale comparisons. Independent calibration validation, missing sources and the future test remain open.
Download scale stages, component overlap and all team sensitivity traces →
Complete replay, scale comparisons & future test
14 archived checkpoints rebuilt across 2024–2026. 494 eligible games receive forecasts using only prior checkpoints and earlier training dates. Current NRR, Beta and the displayed Candidate remain unchanged.
Each team’s complete grade includes bounded game evidence, opponent and team weights, and the display mapping. Each graded game is excluded from its reference fit. Forecast controls are the archived quality-win, adjusted-record, loss-integrity and win-control components.
| Season · games | Gain | 95% date range |
|---|---|---|
| 2024 · 191 | 0.091 | -0.010 to 0.159 |
| 2025 · 245 | 0.057 | -0.017 to 0.155 |
| 2026 · 58 | 0.103 | -0.247 to 0.142 |
This tests next-game yards/play differential, not the correct résumé order. Historical replay can expose problems and shared signal; it cannot establish independent future accuracy or justify promotion by itself.
Average Candidate gains are positive in each season, but all three date-resampling ranges include zero. The 2026 sample has only 58 games across five dates. No formula is promoted.
Calibration comparisons—not a selected winner
Test the same complete grade with 0.75×, 1× and 1.25× the inherited display scale. All forecast models use the same eligible games, a fixed ridge penalty, and only earlier outcomes. No scale is selected from this sweep.
2024 · 191 games
| Inputs | MAE |
|---|---|
| Four résumé controls | 1.433 |
| Controls + Current performance | 1.434 |
| Controls + Candidate · scale 0.75× | 1.339 |
| Controls + Candidate · scale 1× | 1.341 |
| Controls + Candidate · scale 1.25× | 1.342 |
2025 · 245 games
| Inputs | MAE |
|---|---|
| Four résumé controls | 1.360 |
| Controls + Current performance | 1.368 |
| Controls + Candidate · scale 0.75× | 1.307 |
| Controls + Candidate · scale 1× | 1.304 |
| Controls + Candidate · scale 1.25× | 1.302 |
2026 · 58 games
| Inputs | MAE |
|---|---|
| Four résumé controls | 1.470 |
| Controls + Current performance | 1.439 |
| Controls + Candidate · scale 0.75× | 1.362 |
| Controls + Candidate · scale 1× | 1.367 |
| Controls + Candidate · scale 1.25× | 1.369 |
Forecast coefficients—not NRR weights—are estimated from earlier games. These are expanding-history comparisons across seasons, not untouched season holdouts. Shared-team dependence remains.
Coverage, exclusions and source gates
1652 games have no earlier archived component checkpoint. 11 lack archived components for both sides. 50 otherwise eligible games precede the fixed 40-game forecast-training minimum.
Historical component coverage starts in November for 2024 and 2025. Outside-inventory opponents remain in reference fits and team grades, but cannot enter paired forecasts without both component records.
Alternative statistics scenarios: these historical forecasts use the named source scenario. The dedicated Week 7 audit below now tests four documented inputs through the full Candidate network; source adjudication and historical scenario replays remain open.
Missing references: still open. These sources contain FBS-connected games, not a verified complete FCS-versus-FCS schedule and statistical inventory. No external reference games were added.
Reference fits are rebuilt from games strictly before each archived checkpoint, excluding both views of the graded game. No target-day or later game informs a forecast.
This reconstructs the full Candidate bounded evidence, opponent weight, team weight and display mapping at historical checkpoints; finalized archives are not historically versioned feeds.
Forecasts use the most recent prior checkpoint, not a continuously updated daily ranking. Historical 2024/2025 component coverage begins in November.
Only games with archived NRR components for both sides are eligible. Outside-inventory opponents are excluded from forecasts, but remain in reference fits and team grades.
The forecast target is yards/play differential, not game winners, AP agreement or the correct résumé ranking order.
The controls-only forecast uses quality wins, adjusted record, loss integrity and win control. Adding performance tests descriptive predictive value, not independence of the complete NRR Index. Conference titles are not forecast controls.
All forecast fits use earlier target dates across the available seasons. These are expanding-history tests, not leave-one-season-out tests. Models use the same eligible games.
Three predeclared display-scale factors are exploratory comparisons. None is selected from these results. Date-bootstrap ranges do not resolve repeated-team dependence.
No external reference games or alternative game-stat scenarios were verified or added. Those gates and untouched prospective evaluation remain open.
Frozen future-test protocol · awaiting receipts
Registered 2026-10-13 UTC. Keep the existing 1× scale for continuity—not because it won the sweep. Evaluate eligible future games through December 7, 2026, with predictions recorded before kickoff.
The primary comparison is Candidate plus controls versus controls alone. Require at least 200 eligible games, eight game dates, and a positive gain whose 95% date-bootstrap range stays above zero. Source verification, full eligibility reconciliation and a separate complete-Index review remain required.
No future forecasts or results have been collected. The receipt checker rejects late predictions, reused development games, missing source hashes and changed protocols. Timestamps and hashes still require authenticity review. No automatic monitoring or promotion is enabled.
Keep factor 1 and fixed ridge/training rules for continuity, not because this sweep selected a winner. Never overwrite this version, backdate a receipt or retune on test outcomes; changes require a new protocol and fresh future horizon.
Download all rebuilt grades, forecast inputs, predictions and source hashes →
Reference coverage & source policy
One canonical box per game; the same review rule for every team. 285 reference games were checked against a second provider: 164 match on identity, date, scores and core totals; 121 remain held.
No new reference games enter published rankings. The corroborated set is used only in the expanded-reference experiment below. Current, Beta and the existing Candidate stay unchanged.
How a source is selected
Use the retained NCAA-derived canonical box when its required inputs are present and arithmetically consistent. Keep that input if another provider disagrees, with a review flag. Do not choose a box based on the resulting grade, school, conference or poll position.
Incomplete or conflicting reference boxes have no automatic fallback. Both statistical sides must resolve to exact retained NCAA/ESPN IDs, and the completed date, scores and all core totals must agree before entering a reference experiment.
All 285 queued games use identical identity, date, score and core-stat criteria; no school-specific exceptions.
The NCAA-derived canonical is retained even when another provider disagrees. A conflict is held; no source is selected based on which grade looks preferable.
NCAA-derived files and ESPN can share original feeds. Agreement is corroboration, not independent statistical-credit adjudication or proof of schedule completeness.
ESPN total plays are derived from reported rush/pass attempts. Full credit conventions still need review before an expanded-reference experiment.
No reference games are added to Current, Beta, displayed Candidate or the frozen future protocol in this report. Future protocol changes require a separately registered version.
Historical conditional input tests changed 0 ranks across 138 teams, with a largest Index spread of 0.004. This is sensitivity—not source adjudication.
Expanded-reference experiment · not promoted
Apply the same formula with 164 corroborated reference games added. Held games remain excluded. Only the opponent-reference network changes; FBS résumé game counts and other components stay fixed.
| Prior-only game views | 76 → 7 |
|---|---|
| Changed team ranks | 70/138 |
| Same Top 25 teams | 24/25 |
Date ordered replay: 353 games across 19 dates, each tested using earlier games only. Games with no prior opponent connection fell from 53 to 17.
| Reference network | Average error |
|---|---|
| Existing | 1.993 |
| Expanded | 1.980 |
The average reduction is 0.014 yards/play. The date level uncertainty range (-0.012 to +0.042) includes zero, so this replay does not establish a reliable gain. Finalized statistics and the model’s development history also limit what it can prove.
Future test · 59 pregame predictions recorded
Frozen October 13. The same forecast fit and résumé controls predict every eligible game in the October 13–17 slate; only the expanded reference performance input changes. 3 other scheduled games were excluded under the published rules.
No games have been scored yet. Compare prediction errors after verified box scores arrive. The frozen rules require 200 unique games across eight dates and a positive paired gain whose date-level uncertainty range stays above zero. The published NRR and Candidate are unchanged.
Frozen rules and inputs → · Pregame predictions and exclusions →
Statistical credit, complete reference coverage and future outcomes remain open before any formula decision.
Methods and limits
All corroborated reference games enter under the same rule. Held games never enter; there are no school-specific bonuses or exceptions.
Added games supply opponent-reference observations only. FBS résumé game counts, results, other components and modifiers remain archived.
Both views of each graded FBS game stay excluded from its reference fit. This is retrospective cross-fitting, not a pregame holdout.
Cross-provider agreement may reflect shared feeds. Full statistical-credit review and complete coverage remain open.
Partial reference expansion changes the network for all teams. Fewer neutral-prior games and different ranks are not proof of improved accuracy.
No new display scale or weights are fitted. Current, Beta, displayed Candidate and the frozen future protocol are unchanged. A prospective test of this variant requires a separately frozen protocol.
Retrospective test uses finalized statistics and a source inventory acquired after the games. Chronological exclusion prevents target-day reuse but does not make this an untouched pregame trial.
Both variants predict the same yards/play differential for the same eligible games. No rankings, Indexes or weights are selected from the result.
The added reference set is partial and cross-provider agreement may come from shared feeds. Source credit and schedule completeness still need review.
Date-level bootstrap resamples whole game dates; repeated teams across dates remain dependent.
The existing frozen future protocol covers the prior Candidate only. This expanded variant needs its own frozen protocol before any prospective evaluation.
Review counts and remaining gaps
The original queue covers all 76 prior-only opponents. No favored-team subset was selected. Held reason counts can overlap because a game can fail more than one check.
home core stat conflict: 16.
ambiguous away identity: 24.
both side identities required: 43.
both units required: 45.
canonical box incomplete or conflicting: 24.
away core stat conflict: 18.
missing away crosswalk: 12.
missing home crosswalk: 16.
home incomplete or inconsistent box: 2.
missing game id: 47.
away team id conflict: 2.
home team id conflict: 2.
Provider agreement can reflect a shared feed. Schedule completeness and statistical-credit conventions remain open. Individual cases and retained evidence belong in the downloads, not the main explanation.
Where reference coverage is weakest
No prior connection: 53 games · connected MAE 3.747 · weighted MAE 3.781.
1–2 prior games: 131 games · connected MAE 1.844 · weighted MAE 1.910.
3–4 prior games: 135 games · connected MAE 1.587 · weighted MAE 1.611.
5+ prior games: 34 games · connected MAE 1.451 · weighted MAE 1.517.
At the completed Week 7 checkpoint, 76 team-game views still lack another usable opponent connection. The pregame replay has 53 unique games with no prior connection for at least one side. These counts use different timing and counting units.
No missing FCS-versus-FCS statistics were invented or added. Full reference coverage needs source-backed data before we can judge whether neutral-prior grades are fair.
Team-level coverage and source records are available in the validation download; no selected team examples are highlighted here.
Replay boundaries and next promotion gates
Every target is predicted using strictly earlier game dates; the whole target day is excluded.
This is a historical replay using finalized source statistics, not a prospective holdout or historically versioned pregame feed.
Candidate settings were developed after this checkpoint was available. These results are diagnostic, not independent confirmation.
Predicting yards/play differential tests the opponent-reference mechanism, not final résumé rankings, wins, player safety, or agreement with AP.
Connected and sparse-weighted predictions are mechanism diagnostics, not the candidate 0–100 grade or a new ranking formula.
No FCS-versus-FCS statistics were added. Sparse coverage remains unresolved; neutral priors are explicit, not invented observations.
No parameter search or winner-based tuning was performed; no model is promoted from this report.
The combined historical test now spans multiple seasons. Alternative source scenarios and full reference coverage remain unverified. A future protocol is registered, but no future receipts or results have been collected. Keep published formulas unchanged until those gates and the complete-Index review pass.
Download every pregame prediction, coverage breakdown and replay policy →
Showing the Current NRR Top 25, in Current order. Search includes all 138 teams.
Georgia · 6-0Current #1 · Beta #1 · Candidate #1
Beta · experimental
#1 · Index 89.109Performance 96.0/100
Candidate · development
#1 · Index 86.256Performance 73.6/100
Why it moves: performance changes from 79.6 (Current) to 96.0 (Beta) to 73.6 (Candidate). Candidate Index difference versus Beta: -2.853. Rank changes also depend on other teams moving.
Quality wins, adjusted record, loss integrity, win control, conference-title score, record multiplier and head-to-head credit are identical across these three calculations.
How sensitive is this team’s complete grade?
Candidate performance 73.6/100 · 1/6 games have prior-only opponent evidence. Team evidence weight 75%.
Omitting one game from the aggregation shifts performance by at most 5.1 points. Opponent fits stay frozen; this is sensitivity, not a game’s causal value or an independent accuracy test.
Mean displayed game grade: 77.1. The team grade maps the weighted raw mean, so it is not the average of displayed game grades.
Candidate inputs and game-by-game proof
6 games · team evidence weight 75% · adjusted raw mean 0.651. Evidence weights are not confidence probabilities.
Tennessee State · 2026-09-05: 0 other connected games; reference weight 0%. Raw 4.560 → bounded 3.000 → weighted 0.000. Neutral prior only: no other usable opponent connection.
Western Kentucky · 2026-09-12: 5 other connected games; reference weight 71%. Raw 2.435 → bounded 2.435 → weighted 1.739. Candidate game grade 95.4.
Arkansas · 2026-09-19: 5 other connected games; reference weight 71%. Raw 1.200 → bounded 1.200 → weighted 0.857. Candidate game grade 79.7.
Oklahoma · 2026-09-26: 4 other connected games; reference weight 67%. Raw 1.103 → bounded 1.103 → weighted 0.735. Candidate game grade 76.2.
Vanderbilt · 2026-10-03: 5 other connected games; reference weight 71%. Raw 0.875 → bounded 0.875 → weighted 0.625. Candidate game grade 72.8.
Alabama · 2026-10-10: 5 other connected games; reference weight 71%. Raw 1.756 → bounded 1.756 → weighted 1.255. Candidate game grade 88.8.
Notre Dame · 6-0Current #2 · Beta #2 · Candidate #2
Current NRR
#2 · Index 82.417Performance 91.5/100
Beta · experimental
#2 · Index 82.724Performance 93.9/100
Candidate · development
#2 · Index 80.915Performance 79.7/100
Why it moves: performance changes from 91.5 (Current) to 93.9 (Beta) to 79.7 (Candidate). Candidate Index difference versus Beta: -1.809. Rank changes also depend on other teams moving.
Quality wins, adjusted record, loss integrity, win control, conference-title score, record multiplier and head-to-head credit are identical across these three calculations.
How sensitive is this team’s complete grade?
Candidate performance 79.7/100 · 0/6 games have prior-only opponent evidence. Team evidence weight 75%.
Omitting one game from the aggregation shifts performance by at most 4.1 points. Opponent fits stay frozen; this is sensitivity, not a game’s causal value or an independent accuracy test.
Mean displayed game grade: 84.3. The team grade maps the weighted raw mean, so it is not the average of displayed game grades.
Candidate inputs and game-by-game proof
6 games · team evidence weight 75% · adjusted raw mean 0.858. Evidence weights are not confidence probabilities.
Wisconsin · 2026-09-06: 4 other connected games; reference weight 67%. Raw 1.264 → bounded 1.264 → weighted 0.842. Candidate game grade 79.3.
Rice · 2026-09-12: 5 other connected games; reference weight 71%. Raw 2.346 → bounded 2.346 → weighted 1.676. Candidate game grade 94.8.
Michigan State · 2026-09-19: 5 other connected games; reference weight 71%. Raw 1.111 → bounded 1.111 → weighted 0.794. Candidate game grade 77.9.
Purdue · 2026-09-26: 5 other connected games; reference weight 71%. Raw 2.588 → bounded 2.588 → weighted 1.849. Candidate game grade 96.3.
North Carolina · 2026-10-03: 4 other connected games; reference weight 67%. Raw 0.904 → bounded 0.904 → weighted 0.603. Candidate game grade 72.0.
Stanford · 2026-10-10: 5 other connected games; reference weight 71%. Raw 1.536 → bounded 1.536 → weighted 1.097. Candidate game grade 85.6.
UCLA · 5-0Current #3 · Beta #4 · Candidate #3
Current NRR
#3 · Index 79.027Performance 69.6/100
Beta · experimental
#4 · Index 81.578Performance 89.6/100
Candidate · development
#3 · Index 79.695Performance 74.9/100
Why it moves: performance changes from 69.6 (Current) to 89.6 (Beta) to 74.9 (Candidate). Candidate Index difference versus Beta: -1.883. Rank changes also depend on other teams moving.
Quality wins, adjusted record, loss integrity, win control, conference-title score, record multiplier and head-to-head credit are identical across these three calculations.
How sensitive is this team’s complete grade?
Candidate performance 74.9/100 · 0/5 games have prior-only opponent evidence. Team evidence weight 71%.
Omitting one game from the aggregation shifts performance by at most 7.9 points. Opponent fits stay frozen; this is sensitivity, not a game’s causal value or an independent accuracy test.
Mean displayed game grade: 77.1. The team grade maps the weighted raw mean, so it is not the average of displayed game grades.
Candidate inputs and game-by-game proof
5 games · team evidence weight 71% · adjusted raw mean 0.692. Evidence weights are not confidence probabilities.
California · 2026-09-05: 5 other connected games; reference weight 71%. Raw 2.091 → bounded 2.091 → weighted 1.493. Candidate game grade 92.6.
San Diego State · 2026-09-12: 5 other connected games; reference weight 71%. Raw 0.913 → bounded 0.913 → weighted 0.652. Candidate game grade 73.6.
Purdue · 2026-09-19: 5 other connected games; reference weight 71%. Raw 0.006 → bounded 0.006 → weighted 0.004. Candidate game grade 50.2.
Maryland · 2026-09-26: 5 other connected games; reference weight 71%. Raw 2.968 → bounded 2.968 → weighted 2.120. Candidate game grade 98.0.
Oregon · 2026-10-10: 4 other connected games; reference weight 67%. Raw 0.859 → bounded 0.859 → weighted 0.573. Candidate game grade 71.1.
Indiana · 6-0Current #4 · Beta #5 · Candidate #5
Current NRR
#4 · Index 77.854Performance 79.8/100
Beta · experimental
#5 · Index 80.256Performance 98.6/100
Candidate · development
#5 · Index 78.144Performance 82.1/100
Why it moves: performance changes from 79.8 (Current) to 98.6 (Beta) to 82.1 (Candidate). Candidate Index difference versus Beta: -2.112. Rank changes also depend on other teams moving.
Quality wins, adjusted record, loss integrity, win control, conference-title score, record multiplier and head-to-head credit are identical across these three calculations.
How sensitive is this team’s complete grade?
Candidate performance 82.1/100 · 0/6 games have prior-only opponent evidence. Team evidence weight 75%.
Omitting one game from the aggregation shifts performance by at most 2.3 points. Opponent fits stay frozen; this is sensitivity, not a game’s causal value or an independent accuracy test.
Mean displayed game grade: 88.2. The team grade maps the weighted raw mean, so it is not the average of displayed game grades.
Candidate inputs and game-by-game proof
6 games · team evidence weight 75% · adjusted raw mean 0.948. Evidence weights are not confidence probabilities.
North Texas · 2026-09-05: 5 other connected games; reference weight 71%. Raw 1.887 → bounded 1.887 → weighted 1.348. Candidate game grade 90.4.
Howard · 2026-09-12: 1 other connected games; reference weight 33%. Raw 3.155 → bounded 3.000 → weighted 1.000. Candidate game grade 83.4.
Western Kentucky · 2026-09-19: 5 other connected games; reference weight 71%. Raw 2.171 → bounded 2.171 → weighted 1.551. Candidate game grade 93.4.
Northwestern · 2026-09-25: 4 other connected games; reference weight 67%. Raw 1.211 → bounded 1.211 → weighted 0.808. Candidate game grade 78.3.
Rutgers · 2026-10-03: 4 other connected games; reference weight 67%. Raw 2.183 → bounded 2.183 → weighted 1.455. Candidate game grade 92.1.
Nebraska · 2026-10-10: 5 other connected games; reference weight 71%. Raw 1.990 → bounded 1.990 → weighted 1.422. Candidate game grade 91.6.
Alabama · 5-1Current #5 · Beta #7 · Candidate #7
Current NRR
#5 · Index 75.237Performance 81.7/100
Beta · experimental
#7 · Index 76.890Performance 96.4/100
Candidate · development
#7 · Index 74.712Performance 77.1/100
Why it moves: performance changes from 81.7 (Current) to 96.4 (Beta) to 77.1 (Candidate). Candidate Index difference versus Beta: -2.178. Rank changes also depend on other teams moving.
Quality wins, adjusted record, loss integrity, win control, conference-title score, record multiplier and head-to-head credit are identical across these three calculations.
How sensitive is this team’s complete grade?
Candidate performance 77.1/100 · 0/6 games have prior-only opponent evidence. Team evidence weight 75%.
Omitting one game from the aggregation shifts performance by at most 4.1 points. Opponent fits stay frozen; this is sensitivity, not a game’s causal value or an independent accuracy test.
Mean displayed game grade: 81.0. The team grade maps the weighted raw mean, so it is not the average of displayed game grades.
Candidate inputs and game-by-game proof
6 games · team evidence weight 75% · adjusted raw mean 0.765. Evidence weights are not confidence probabilities.
East Carolina · 2026-09-05: 4 other connected games; reference weight 67%. Raw 1.433 → bounded 1.433 → weighted 0.955. Candidate game grade 82.3.
Kentucky · 2026-09-12: 5 other connected games; reference weight 71%. Raw 1.661 → bounded 1.661 → weighted 1.186. Candidate game grade 87.5.
Florida State · 2026-09-19: 5 other connected games; reference weight 71%. Raw 0.925 → bounded 0.925 → weighted 0.661. Candidate game grade 73.9.
South Carolina · 2026-09-26: 5 other connected games; reference weight 71%. Raw 2.069 → bounded 2.069 → weighted 1.478. Candidate game grade 92.4.
Mississippi State · 2026-10-03: 4 other connected games; reference weight 67%. Raw 2.553 → bounded 2.553 → weighted 1.702. Candidate game grade 95.0.
Georgia · 2026-10-10: 5 other connected games; reference weight 71%. Raw 0.188 → bounded 0.188 → weighted 0.134. Candidate game grade 55.2.
Texas Tech · 5-0Current #6 · Beta #6 · Candidate #6
Current NRR
#6 · Index 74.790Performance 60.1/100
Beta · experimental
#6 · Index 78.004Performance 85.3/100
Candidate · development
#6 · Index 74.931Performance 61.2/100
Why it moves: performance changes from 60.1 (Current) to 85.3 (Beta) to 61.2 (Candidate). Candidate Index difference versus Beta: -3.073. Rank changes also depend on other teams moving.
Quality wins, adjusted record, loss integrity, win control, conference-title score, record multiplier and head-to-head credit are identical across these three calculations.
How sensitive is this team’s complete grade?
Candidate performance 61.2/100 · 1/5 games have prior-only opponent evidence. Team evidence weight 71%.
Omitting one game from the aggregation shifts performance by at most 5.0 points. Opponent fits stay frozen; this is sensitivity, not a game’s causal value or an independent accuracy test.
Mean displayed game grade: 63.8. The team grade maps the weighted raw mean, so it is not the average of displayed game grades.
Candidate inputs and game-by-game proof
5 games · team evidence weight 71% · adjusted raw mean 0.294. Evidence weights are not confidence probabilities.
Abilene Christian · 2026-09-05: 0 other connected games; reference weight 0%. Raw 1.636 → bounded 1.636 → weighted 0.000. Neutral prior only: no other usable opponent connection.
Oregon State · 2026-09-12: 5 other connected games; reference weight 71%. Raw -0.274 → bounded -0.274 → weighted -0.195. Candidate game grade 42.5.
Houston · 2026-09-18: 5 other connected games; reference weight 71%. Raw 1.251 → bounded 1.251 → weighted 0.894. Candidate game grade 80.7.
Sam Houston · 2026-09-26: 4 other connected games; reference weight 67%. Raw 1.637 → bounded 1.637 → weighted 1.091. Candidate game grade 85.5.
Colorado · 2026-10-03: 4 other connected games; reference weight 67%. Raw 0.398 → bounded 0.398 → weighted 0.266. Candidate game grade 60.2.
Miami · 5-0Current #7 · Beta #3 · Candidate #4
Current NRR
#7 · Index 74.110Performance 30.4/100
Beta · experimental
#3 · Index 82.690Performance 97.7/100
Candidate · development
#4 · Index 79.524Performance 72.8/100
Why it moves: performance changes from 30.4 (Current) to 97.7 (Beta) to 72.8 (Candidate). Candidate Index difference versus Beta: -3.166. Rank changes also depend on other teams moving.
Quality wins, adjusted record, loss integrity, win control, conference-title score, record multiplier and head-to-head credit are identical across these three calculations.
How sensitive is this team’s complete grade?
Candidate performance 72.8/100 · 1/5 games have prior-only opponent evidence. Team evidence weight 71%.
Omitting one game from the aggregation shifts performance by at most 7.2 points. Opponent fits stay frozen; this is sensitivity, not a game’s causal value or an independent accuracy test.
Mean displayed game grade: 76.4. The team grade maps the weighted raw mean, so it is not the average of displayed game grades.
Candidate inputs and game-by-game proof
5 games · team evidence weight 71% · adjusted raw mean 0.627. Evidence weights are not confidence probabilities.
Stanford · 2026-09-04: 5 other connected games; reference weight 71%. Raw 1.571 → bounded 1.571 → weighted 1.122. Candidate game grade 86.2.
Florida A&M · 2026-09-10: 0 other connected games; reference weight 0%. Raw 4.725 → bounded 3.000 → weighted 0.000. Neutral prior only: no other usable opponent connection.
Wake Forest · 2026-09-18: 5 other connected games; reference weight 71%. Raw 0.811 → bounded 0.811 → weighted 0.580. Candidate game grade 71.3.
Central Michigan · 2026-09-26: 5 other connected games; reference weight 71%. Raw 2.646 → bounded 2.646 → weighted 1.890. Candidate game grade 96.6.
Clemson · 2026-10-03: 4 other connected games; reference weight 67%. Raw 1.193 → bounded 1.193 → weighted 0.796. Candidate game grade 78.0.
BYU · 5-0Current #8 · Beta #8 · Candidate #8
Current NRR
#8 · Index 71.681Performance 70.2/100
Beta · experimental
#8 · Index 74.897Performance 95.4/100
Candidate · development
#8 · Index 72.097Performance 73.4/100
Why it moves: performance changes from 70.2 (Current) to 95.4 (Beta) to 73.4 (Candidate). Candidate Index difference versus Beta: -2.800. Rank changes also depend on other teams moving.
Quality wins, adjusted record, loss integrity, win control, conference-title score, record multiplier and head-to-head credit are identical across these three calculations.
How sensitive is this team’s complete grade?
Candidate performance 73.4/100 · 1/5 games have prior-only opponent evidence. Team evidence weight 71%.
Omitting one game from the aggregation shifts performance by at most 7.3 points. Opponent fits stay frozen; this is sensitivity, not a game’s causal value or an independent accuracy test.
Mean displayed game grade: 74.2. The team grade maps the weighted raw mean, so it is not the average of displayed game grades.
Candidate inputs and game-by-game proof
5 games · team evidence weight 71% · adjusted raw mean 0.646. Evidence weights are not confidence probabilities.
Utah Tech · 2026-09-05: 0 other connected games; reference weight 0%. Raw 2.188 → bounded 2.188 → weighted 0.000. Neutral prior only: no other usable opponent connection.
Arizona · 2026-09-12: 5 other connected games; reference weight 71%. Raw 1.365 → bounded 1.365 → weighted 0.975. Candidate game grade 82.8.
Colorado State · 2026-09-19: 4 other connected games; reference weight 67%. Raw 2.564 → bounded 2.564 → weighted 1.710. Candidate game grade 95.1.
TCU · 2026-10-03: 4 other connected games; reference weight 67%. Raw -0.152 → bounded -0.152 → weighted -0.101. Candidate game grade 46.1.
Iowa State · 2026-10-09: 5 other connected games; reference weight 71%. Raw 2.712 → bounded 2.712 → weighted 1.937. Candidate game grade 97.0.
Texas · 4-1Current #9 · Beta #10 · Candidate #9
Current NRR
#9 · Index 71.056Performance 69.9/100
Beta · experimental
#10 · Index 73.710Performance 94.1/100
Candidate · development
#9 · Index 71.360Performance 72.6/100
Why it moves: performance changes from 69.9 (Current) to 94.1 (Beta) to 72.6 (Candidate). Candidate Index difference versus Beta: -2.350. Rank changes also depend on other teams moving.
Quality wins, adjusted record, loss integrity, win control, conference-title score, record multiplier and head-to-head credit are identical across these three calculations.
How sensitive is this team’s complete grade?
Candidate performance 72.6/100 · 0/5 games have prior-only opponent evidence. Team evidence weight 71%.
Omitting one game from the aggregation shifts performance by at most 3.2 points. Opponent fits stay frozen; this is sensitivity, not a game’s causal value or an independent accuracy test.
Mean displayed game grade: 78.9. The team grade maps the weighted raw mean, so it is not the average of displayed game grades.
Candidate inputs and game-by-game proof
5 games · team evidence weight 71% · adjusted raw mean 0.622. Evidence weights are not confidence probabilities.
Texas State · 2026-09-05: 4 other connected games; reference weight 67%. Raw 1.803 → bounded 1.803 → weighted 1.202. Candidate game grade 87.8.
Ohio State · 2026-09-12: 5 other connected games; reference weight 71%. Raw 0.318 → bounded 0.318 → weighted 0.227. Candidate game grade 58.7.
UTSA · 2026-09-19: 5 other connected games; reference weight 71%. Raw 1.269 → bounded 1.269 → weighted 0.907. Candidate game grade 81.0.
Tennessee · 2026-09-26: 5 other connected games; reference weight 71%. Raw 1.523 → bounded 1.523 → weighted 1.088. Candidate game grade 85.4.
Oklahoma · 2026-10-10: 4 other connected games; reference weight 67%. Raw 1.391 → bounded 1.391 → weighted 0.927. Candidate game grade 81.5.
Ohio State · 5-1Current #10 · Beta #11 · Candidate #11
Current NRR
#10 · Index 67.715Performance 79.1/100
Beta · experimental
#11 · Index 69.711Performance 97.8/100
Candidate · development
#11 · Index 68.332Performance 84.9/100
Why it moves: performance changes from 79.1 (Current) to 97.8 (Beta) to 84.9 (Candidate). Candidate Index difference versus Beta: -1.379. Rank changes also depend on other teams moving.
Quality wins, adjusted record, loss integrity, win control, conference-title score, record multiplier and head-to-head credit are identical across these three calculations.
How sensitive is this team’s complete grade?
Candidate performance 84.9/100 · 0/6 games have prior-only opponent evidence. Team evidence weight 75%.
Omitting one game from the aggregation shifts performance by at most 3.8 points. Opponent fits stay frozen; this is sensitivity, not a game’s causal value or an independent accuracy test.
Mean displayed game grade: 89.1. The team grade maps the weighted raw mean, so it is not the average of displayed game grades.
Candidate inputs and game-by-game proof
6 games · team evidence weight 75% · adjusted raw mean 1.063. Evidence weights are not confidence probabilities.
Ball State · 2026-09-05: 5 other connected games; reference weight 71%. Raw 3.073 → bounded 3.000 → weighted 2.143. Candidate game grade 98.1.
Texas · 2026-09-12: 4 other connected games; reference weight 67%. Raw 1.502 → bounded 1.502 → weighted 1.002. Candidate game grade 83.4.
Kent State · 2026-09-19: 5 other connected games; reference weight 71%. Raw 2.517 → bounded 2.517 → weighted 1.798. Candidate game grade 95.9.
Illinois · 2026-09-26: 5 other connected games; reference weight 71%. Raw 0.894 → bounded 0.894 → weighted 0.638. Candidate game grade 73.2.
Iowa · 2026-10-03: 5 other connected games; reference weight 71%. Raw 1.832 → bounded 1.832 → weighted 1.309. Candidate game grade 89.8.
Maryland · 2026-10-10: 5 other connected games; reference weight 71%. Raw 2.265 → bounded 2.265 → weighted 1.618. Candidate game grade 94.2.
Utah · 5-0Current #11 · Beta #9 · Candidate #10
Current NRR
#11 · Index 66.466Performance 35.4/100
Beta · experimental
#9 · Index 73.958Performance 94.1/100
Candidate · development
#10 · Index 70.934Performance 70.4/100
Why it moves: performance changes from 35.4 (Current) to 94.1 (Beta) to 70.4 (Candidate). Candidate Index difference versus Beta: -3.024. Rank changes also depend on other teams moving.
Quality wins, adjusted record, loss integrity, win control, conference-title score, record multiplier and head-to-head credit are identical across these three calculations.
How sensitive is this team’s complete grade?
Candidate performance 70.4/100 · 1/5 games have prior-only opponent evidence. Team evidence weight 71%.
Omitting one game from the aggregation shifts performance by at most 3.5 points. Opponent fits stay frozen; this is sensitivity, not a game’s causal value or an independent accuracy test.
Mean displayed game grade: 75.5. The team grade maps the weighted raw mean, so it is not the average of displayed game grades.
Candidate inputs and game-by-game proof
5 games · team evidence weight 71% · adjusted raw mean 0.553. Evidence weights are not confidence probabilities.
Idaho · 2026-09-03: 0 other connected games; reference weight 0%. Raw 2.376 → bounded 2.376 → weighted 0.000. Neutral prior only: no other usable opponent connection.
Arkansas · 2026-09-12: 5 other connected games; reference weight 71%. Raw 1.565 → bounded 1.565 → weighted 1.118. Candidate game grade 86.1.
Utah State · 2026-09-19: 5 other connected games; reference weight 71%. Raw 1.634 → bounded 1.634 → weighted 1.167. Candidate game grade 87.1.
Iowa State · 2026-09-26: 5 other connected games; reference weight 71%. Raw 1.490 → bounded 1.490 → weighted 1.064. Candidate game grade 84.9.
Kansas · 2026-10-10: 4 other connected games; reference weight 67%. Raw 0.786 → bounded 0.786 → weighted 0.524. Candidate game grade 69.4.
Kentucky · 5-1Current #12 · Beta #13 · Candidate #13
Current NRR
#12 · Index 65.155Performance 60.0/100
Beta · experimental
#13 · Index 67.896Performance 84.9/100
Candidate · development
#13 · Index 65.328Performance 61.5/100
Why it moves: performance changes from 60.0 (Current) to 84.9 (Beta) to 61.5 (Candidate). Candidate Index difference versus Beta: -2.568. Rank changes also depend on other teams moving.
Quality wins, adjusted record, loss integrity, win control, conference-title score, record multiplier and head-to-head credit are identical across these three calculations.
How sensitive is this team’s complete grade?
Candidate performance 61.5/100 · 1/6 games have prior-only opponent evidence. Team evidence weight 75%.
Omitting one game from the aggregation shifts performance by at most 4.5 points. Opponent fits stay frozen; this is sensitivity, not a game’s causal value or an independent accuracy test.
Mean displayed game grade: 63.3. The team grade maps the weighted raw mean, so it is not the average of displayed game grades.
Candidate inputs and game-by-game proof
6 games · team evidence weight 75% · adjusted raw mean 0.302. Evidence weights are not confidence probabilities.
Youngstown State · 2026-09-05: 0 other connected games; reference weight 0%. Raw 1.837 → bounded 1.837 → weighted 0.000. Neutral prior only: no other usable opponent connection.
Alabama · 2026-09-12: 5 other connected games; reference weight 71%. Raw -0.335 → bounded -0.335 → weighted -0.239. Candidate game grade 40.8.
Texas A&M · 2026-09-19: 5 other connected games; reference weight 71%. Raw 1.301 → bounded 1.301 → weighted 0.929. Candidate game grade 81.6.
South Alabama · 2026-09-26: 5 other connected games; reference weight 71%. Raw 1.609 → bounded 1.609 → weighted 1.150. Candidate game grade 86.7.
South Carolina · 2026-10-03: 5 other connected games; reference weight 71%. Raw -0.110 → bounded -0.110 → weighted -0.078. Candidate game grade 47.0.
LSU · 2026-10-10: 5 other connected games; reference weight 71%. Raw 0.921 → bounded 0.921 → weighted 0.658. Candidate game grade 73.8.
Mississippi State · 4-1Current #13 · Beta #14 · Candidate #14
Current NRR
#13 · Index 63.365Performance 66.2/100
Beta · experimental
#14 · Index 65.670Performance 87.4/100
Candidate · development
#14 · Index 63.701Performance 69.3/100
Why it moves: performance changes from 66.2 (Current) to 87.4 (Beta) to 69.3 (Candidate). Candidate Index difference versus Beta: -1.969. Rank changes also depend on other teams moving.
Quality wins, adjusted record, loss integrity, win control, conference-title score, record multiplier and head-to-head credit are identical across these three calculations.
How sensitive is this team’s complete grade?
Candidate performance 69.3/100 · 0/5 games have prior-only opponent evidence. Team evidence weight 71%.
Omitting one game from the aggregation shifts performance by at most 7.1 points. Opponent fits stay frozen; this is sensitivity, not a game’s causal value or an independent accuracy test.
Mean displayed game grade: 72.0. The team grade maps the weighted raw mean, so it is not the average of displayed game grades.
Candidate inputs and game-by-game proof
5 games · team evidence weight 71% · adjusted raw mean 0.521. Evidence weights are not confidence probabilities.
Louisiana-Monroe · 2026-09-05: 4 other connected games; reference weight 67%. Raw 2.543 → bounded 2.543 → weighted 1.695. Candidate game grade 95.0.
Minnesota · 2026-09-12: 5 other connected games; reference weight 71%. Raw 1.578 → bounded 1.578 → weighted 1.127. Candidate game grade 86.3.
South Carolina · 2026-09-19: 5 other connected games; reference weight 71%. Raw 1.069 → bounded 1.069 → weighted 0.764. Candidate game grade 77.0.
Missouri · 2026-09-26: 5 other connected games; reference weight 71%. Raw 1.224 → bounded 1.224 → weighted 0.874. Candidate game grade 80.2.
Alabama · 2026-10-03: 5 other connected games; reference weight 71%. Raw -1.138 → bounded -1.138 → weighted -0.813. Candidate game grade 21.5.
Wake Forest · 5-1Current #14 · Beta #15 · Candidate #15
Current NRR
#14 · Index 62.651Performance 69.4/100
Beta · experimental
#15 · Index 65.081Performance 92.1/100
Candidate · development
#15 · Index 63.270Performance 75.2/100
Why it moves: performance changes from 69.4 (Current) to 92.1 (Beta) to 75.2 (Candidate). Candidate Index difference versus Beta: -1.811. Rank changes also depend on other teams moving.
Quality wins, adjusted record, loss integrity, win control, conference-title score, record multiplier and head-to-head credit are identical across these three calculations.
How sensitive is this team’s complete grade?
Candidate performance 75.2/100 · 0/6 games have prior-only opponent evidence. Team evidence weight 75%.
Omitting one game from the aggregation shifts performance by at most 4.3 points. Opponent fits stay frozen; this is sensitivity, not a game’s causal value or an independent accuracy test.
Mean displayed game grade: 79.8. The team grade maps the weighted raw mean, so it is not the average of displayed game grades.
Candidate inputs and game-by-game proof
6 games · team evidence weight 75% · adjusted raw mean 0.702. Evidence weights are not confidence probabilities.
Akron · 2026-09-03: 5 other connected games; reference weight 71%. Raw 1.674 → bounded 1.674 → weighted 1.195. Candidate game grade 87.7.
Purdue · 2026-09-12: 5 other connected games; reference weight 71%. Raw 0.396 → bounded 0.396 → weighted 0.283. Candidate game grade 60.8.
Miami · 2026-09-18: 4 other connected games; reference weight 67%. Raw 1.065 → bounded 1.065 → weighted 0.710. Candidate game grade 75.4.
Louisville · 2026-09-26: 5 other connected games; reference weight 71%. Raw 1.577 → bounded 1.577 → weighted 1.126. Candidate game grade 86.2.
Stanford · 2026-10-03: 5 other connected games; reference weight 71%. Raw 2.290 → bounded 2.290 → weighted 1.636. Candidate game grade 94.3.
NC State · 2026-10-10: 5 other connected games; reference weight 71%. Raw 0.931 → bounded 0.931 → weighted 0.665. Candidate game grade 74.0.
Missouri · 5-1Current #15 · Beta #16 · Candidate #16
Current NRR
#15 · Index 61.734Performance 72.7/100
Beta · experimental
#16 · Index 63.919Performance 93.3/100
Candidate · development
#16 · Index 61.327Performance 68.9/100
Why it moves: performance changes from 72.7 (Current) to 93.3 (Beta) to 68.9 (Candidate). Candidate Index difference versus Beta: -2.592. Rank changes also depend on other teams moving.
Quality wins, adjusted record, loss integrity, win control, conference-title score, record multiplier and head-to-head credit are identical across these three calculations.
How sensitive is this team’s complete grade?
Candidate performance 68.9/100 · 1/6 games have prior-only opponent evidence. Team evidence weight 75%.
Omitting one game from the aggregation shifts performance by at most 6.0 points. Opponent fits stay frozen; this is sensitivity, not a game’s causal value or an independent accuracy test.
Mean displayed game grade: 69.9. The team grade maps the weighted raw mean, so it is not the average of displayed game grades.
Candidate inputs and game-by-game proof
6 games · team evidence weight 75% · adjusted raw mean 0.508. Evidence weights are not confidence probabilities.
Arkansas-Pine Bluff · 2026-09-03: 0 other connected games; reference weight 0%. Raw 1.735 → bounded 1.735 → weighted 0.000. Neutral prior only: no other usable opponent connection.
Kansas · 2026-09-11: 4 other connected games; reference weight 67%. Raw 2.247 → bounded 2.247 → weighted 1.498. Candidate game grade 92.7.
Troy · 2026-09-19: 4 other connected games; reference weight 67%. Raw 0.221 → bounded 0.221 → weighted 0.147. Candidate game grade 55.7.
Mississippi State · 2026-09-26: 4 other connected games; reference weight 67%. Raw 0.036 → bounded 0.036 → weighted 0.024. Candidate game grade 50.9.
Florida · 2026-10-03: 5 other connected games; reference weight 71%. Raw 2.372 → bounded 2.372 → weighted 1.695. Candidate game grade 95.0.
Texas A&M · 2026-10-10: 5 other connected games; reference weight 71%. Raw 0.978 → bounded 0.978 → weighted 0.699. Candidate game grade 75.1.
Houston · 5-1Current #16 · Beta #17 · Candidate #17
Current NRR
#16 · Index 60.026Performance 68.7/100
Beta · experimental
#17 · Index 62.557Performance 92.1/100
Candidate · development
#17 · Index 59.521Performance 64.0/100
Why it moves: performance changes from 68.7 (Current) to 92.1 (Beta) to 64.0 (Candidate). Candidate Index difference versus Beta: -3.036. Rank changes also depend on other teams moving.
Quality wins, adjusted record, loss integrity, win control, conference-title score, record multiplier and head-to-head credit are identical across these three calculations.
How sensitive is this team’s complete grade?
Candidate performance 64.0/100 · 1/6 games have prior-only opponent evidence. Team evidence weight 75%.
Omitting one game from the aggregation shifts performance by at most 4.9 points. Opponent fits stay frozen; this is sensitivity, not a game’s causal value or an independent accuracy test.
Mean displayed game grade: 66.4. The team grade maps the weighted raw mean, so it is not the average of displayed game grades.
Candidate inputs and game-by-game proof
6 games · team evidence weight 75% · adjusted raw mean 0.370. Evidence weights are not confidence probabilities.
Oregon State · 2026-09-05: 5 other connected games; reference weight 71%. Raw 1.819 → bounded 1.819 → weighted 1.299. Candidate game grade 89.6.
Southern · 2026-09-12: 0 other connected games; reference weight 0%. Raw 3.177 → bounded 3.000 → weighted 0.000. Neutral prior only: no other usable opponent connection.
Texas Tech · 2026-09-18: 4 other connected games; reference weight 67%. Raw -0.104 → bounded -0.104 → weighted -0.069. Candidate game grade 47.3.
Georgia Southern · 2026-09-26: 5 other connected games; reference weight 71%. Raw 1.170 → bounded 1.170 → weighted 0.836. Candidate game grade 79.1.
UCF · 2026-10-03: 5 other connected games; reference weight 71%. Raw 0.185 → bounded 0.185 → weighted 0.132. Candidate game grade 55.1.
Kansas State · 2026-10-10: 4 other connected games; reference weight 67%. Raw 1.145 → bounded 1.145 → weighted 0.763. Candidate game grade 77.0.
Wisconsin · 4-1Current #17 · Beta #19 · Candidate #20
Current NRR
#17 · Index 59.183Performance 70.5/100
Beta · experimental
#19 · Index 61.613Performance 94.5/100
Candidate · development
#20 · Index 59.109Performance 69.8/100
Why it moves: performance changes from 70.5 (Current) to 94.5 (Beta) to 69.8 (Candidate). Candidate Index difference versus Beta: -2.504. Rank changes also depend on other teams moving.
Quality wins, adjusted record, loss integrity, win control, conference-title score, record multiplier and head-to-head credit are identical across these three calculations.
How sensitive is this team’s complete grade?
Candidate performance 69.8/100 · 1/5 games have prior-only opponent evidence. Team evidence weight 71%.
Omitting one game from the aggregation shifts performance by at most 4.7 points. Opponent fits stay frozen; this is sensitivity, not a game’s causal value or an independent accuracy test.
Mean displayed game grade: 74.0. The team grade maps the weighted raw mean, so it is not the average of displayed game grades.
Candidate inputs and game-by-game proof
5 games · team evidence weight 71% · adjusted raw mean 0.534. Evidence weights are not confidence probabilities.
Notre Dame · 2026-09-06: 5 other connected games; reference weight 71%. Raw 0.355 → bounded 0.355 → weighted 0.254. Candidate game grade 59.7.
Western Illinois · 2026-09-12: 0 other connected games; reference weight 0%. Raw 1.395 → bounded 1.395 → weighted 0.000. Neutral prior only: no other usable opponent connection.
Eastern Michigan · 2026-09-19: 6 other connected games; reference weight 75%. Raw 1.785 → bounded 1.785 → weighted 1.339. Candidate game grade 90.3.
Penn State · 2026-09-26: 5 other connected games; reference weight 71%. Raw 1.385 → bounded 1.385 → weighted 0.990. Candidate game grade 83.1.
Michigan State · 2026-10-03: 5 other connected games; reference weight 71%. Raw 1.617 → bounded 1.617 → weighted 1.155. Candidate game grade 86.8.
Iowa · 5-1Current #18 · Beta #18 · Candidate #18
Current NRR
#18 · Index 59.169Performance 63.7/100
Beta · experimental
#18 · Index 62.351Performance 93.6/100
Candidate · development
#18 · Index 59.229Performance 64.3/100
Why it moves: performance changes from 63.7 (Current) to 93.6 (Beta) to 64.3 (Candidate). Candidate Index difference versus Beta: -3.122. Rank changes also depend on other teams moving.
Quality wins, adjusted record, loss integrity, win control, conference-title score, record multiplier and head-to-head credit are identical across these three calculations.
How sensitive is this team’s complete grade?
Candidate performance 64.3/100 · 1/6 games have prior-only opponent evidence. Team evidence weight 75%.
Omitting one game from the aggregation shifts performance by at most 7.1 points. Opponent fits stay frozen; this is sensitivity, not a game’s causal value or an independent accuracy test.
Mean displayed game grade: 63.5. The team grade maps the weighted raw mean, so it is not the average of displayed game grades.
Candidate inputs and game-by-game proof
6 games · team evidence weight 75% · adjusted raw mean 0.377. Evidence weights are not confidence probabilities.
Northern Illinois · 2026-09-05: 4 other connected games; reference weight 67%. Raw 2.205 → bounded 2.205 → weighted 1.470. Candidate game grade 92.3.
Iowa State · 2026-09-12: 5 other connected games; reference weight 71%. Raw 0.089 → bounded 0.089 → weighted 0.064. Candidate game grade 52.5.
Northern Iowa · 2026-09-19: 0 other connected games; reference weight 0%. Raw 3.664 → bounded 3.000 → weighted 0.000. Neutral prior only: no other usable opponent connection.
Michigan · 2026-09-26: 4 other connected games; reference weight 67%. Raw -0.397 → bounded -0.397 → weighted -0.265. Candidate game grade 39.9.
Ohio State · 2026-10-03: 5 other connected games; reference weight 71%. Raw 0.056 → bounded 0.056 → weighted 0.040. Candidate game grade 51.6.
Washington · 2026-10-09: 5 other connected games; reference weight 71%. Raw 2.389 → bounded 2.389 → weighted 1.707. Candidate game grade 95.1.
Nebraska · 5-1Current #19 · Beta #21 · Candidate #21
Current NRR
#19 · Index 58.533Performance 66.2/100
Beta · experimental
#21 · Index 60.613Performance 86.8/100
Candidate · development
#21 · Index 58.314Performance 64.0/100
Why it moves: performance changes from 66.2 (Current) to 86.8 (Beta) to 64.0 (Candidate). Candidate Index difference versus Beta: -2.299. Rank changes also depend on other teams moving.
Quality wins, adjusted record, loss integrity, win control, conference-title score, record multiplier and head-to-head credit are identical across these three calculations.
How sensitive is this team’s complete grade?
Candidate performance 64.0/100 · 1/6 games have prior-only opponent evidence. Team evidence weight 75%.
Omitting one game from the aggregation shifts performance by at most 6.3 points. Opponent fits stay frozen; this is sensitivity, not a game’s causal value or an independent accuracy test.
Mean displayed game grade: 65.5. The team grade maps the weighted raw mean, so it is not the average of displayed game grades.
Candidate inputs and game-by-game proof
6 games · team evidence weight 75% · adjusted raw mean 0.370. Evidence weights are not confidence probabilities.
Ohio · 2026-09-05: 5 other connected games; reference weight 71%. Raw 1.285 → bounded 1.285 → weighted 0.918. Candidate game grade 81.3.
Bowling Green · 2026-09-12: 5 other connected games; reference weight 71%. Raw 2.180 → bounded 2.180 → weighted 1.557. Candidate game grade 93.4.
North Dakota · 2026-09-19: 0 other connected games; reference weight 0%. Raw 1.608 → bounded 1.608 → weighted 0.000. Neutral prior only: no other usable opponent connection.
Michigan State · 2026-09-26: 5 other connected games; reference weight 71%. Raw 0.576 → bounded 0.576 → weighted 0.411. Candidate game grade 65.5.
Maryland · 2026-10-03: 5 other connected games; reference weight 71%. Raw 0.370 → bounded 0.370 → weighted 0.264. Candidate game grade 60.1.
Indiana · 2026-10-10: 5 other connected games; reference weight 71%. Raw -0.263 → bounded -0.263 → weighted -0.188. Candidate game grade 42.8.
Ole Miss · 4-1Current #20 · Beta #20 · Candidate #19
Current NRR
#20 · Index 58.498Performance 61.0/100
Beta · experimental
#20 · Index 60.983Performance 84.0/100
Candidate · development
#19 · Index 59.229Performance 67.7/100
Why it moves: performance changes from 61.0 (Current) to 84.0 (Beta) to 67.7 (Candidate). Candidate Index difference versus Beta: -1.754. Rank changes also depend on other teams moving.
Quality wins, adjusted record, loss integrity, win control, conference-title score, record multiplier and head-to-head credit are identical across these three calculations.
How sensitive is this team’s complete grade?
Candidate performance 67.7/100 · 0/5 games have prior-only opponent evidence. Team evidence weight 71%.
Omitting one game from the aggregation shifts performance by at most 3.4 points. Opponent fits stay frozen; this is sensitivity, not a game’s causal value or an independent accuracy test.
Mean displayed game grade: 73.5. The team grade maps the weighted raw mean, so it is not the average of displayed game grades.
Candidate inputs and game-by-game proof
5 games · team evidence weight 71% · adjusted raw mean 0.475. Evidence weights are not confidence probabilities.
Louisville · 2026-09-06: 5 other connected games; reference weight 71%. Raw 0.580 → bounded 0.580 → weighted 0.414. Candidate game grade 65.6.
Charlotte · 2026-09-12: 5 other connected games; reference weight 71%. Raw 0.767 → bounded 0.767 → weighted 0.548. Candidate game grade 70.2.
LSU · 2026-09-19: 5 other connected games; reference weight 71%. Raw 1.481 → bounded 1.481 → weighted 1.058. Candidate game grade 84.7.
Florida · 2026-09-26: 5 other connected games; reference weight 71%. Raw 0.733 → bounded 0.733 → weighted 0.524. Candidate game grade 69.4.
Vanderbilt · 2026-10-10: 5 other connected games; reference weight 71%. Raw 1.097 → bounded 1.097 → weighted 0.784. Candidate game grade 77.6.
James Madison · 6-0Current #21 · Beta #12 · Candidate #12
Current NRR
#21 · Index 58.404Performance 0.0/100
Beta · experimental
#12 · Index 68.447Performance 78.8/100
Candidate · development
#12 · Index 66.489Performance 63.4/100
Why it moves: performance changes from 0.0 (Current) to 78.8 (Beta) to 63.4 (Candidate). Candidate Index difference versus Beta: -1.958. Rank changes also depend on other teams moving.
Quality wins, adjusted record, loss integrity, win control, conference-title score, record multiplier and head-to-head credit are identical across these three calculations.
How sensitive is this team’s complete grade?
Candidate performance 63.4/100 · 0/6 games have prior-only opponent evidence. Team evidence weight 75%.
Omitting one game from the aggregation shifts performance by at most 3.6 points. Opponent fits stay frozen; this is sensitivity, not a game’s causal value or an independent accuracy test.
Mean displayed game grade: 66.3. The team grade maps the weighted raw mean, so it is not the average of displayed game grades.
Candidate inputs and game-by-game proof
6 games · team evidence weight 75% · adjusted raw mean 0.354. Evidence weights are not confidence probabilities.
Liberty · 2026-09-05: 5 other connected games; reference weight 71%. Raw 1.067 → bounded 1.067 → weighted 0.762. Candidate game grade 77.0.
Wagner · 2026-09-12: 1 other connected games; reference weight 33%. Raw 2.688 → bounded 2.688 → weighted 0.896. Candidate game grade 80.7.
San Diego State · 2026-09-19: 5 other connected games; reference weight 71%. Raw -0.028 → bounded -0.028 → weighted -0.020. Candidate game grade 49.2.
Old Dominion · 2026-09-26: 5 other connected games; reference weight 71%. Raw 0.865 → bounded 0.865 → weighted 0.618. Candidate game grade 72.5.
Marshall · 2026-10-03: 5 other connected games; reference weight 71%. Raw 1.293 → bounded 1.293 → weighted 0.924. Candidate game grade 81.5.
Georgia Southern · 2026-10-10: 5 other connected games; reference weight 71%. Raw -0.490 → bounded -0.490 → weighted -0.350. Candidate game grade 36.7.
Northwestern · 4-1Current #22 · Beta #22 · Candidate #22
Current NRR
#22 · Index 57.319Performance 70.7/100
Beta · experimental
#22 · Index 59.841Performance 95.7/100
Candidate · development
#22 · Index 56.822Performance 65.8/100
Why it moves: performance changes from 70.7 (Current) to 95.7 (Beta) to 65.8 (Candidate). Candidate Index difference versus Beta: -3.019. Rank changes also depend on other teams moving.
Quality wins, adjusted record, loss integrity, win control, conference-title score, record multiplier and head-to-head credit are identical across these three calculations.
How sensitive is this team’s complete grade?
Candidate performance 65.8/100 · 1/5 games have prior-only opponent evidence. Team evidence weight 71%.
Omitting one game from the aggregation shifts performance by at most 5.7 points. Opponent fits stay frozen; this is sensitivity, not a game’s causal value or an independent accuracy test.
Mean displayed game grade: 69.3. The team grade maps the weighted raw mean, so it is not the average of displayed game grades.
Candidate inputs and game-by-game proof
5 games · team evidence weight 71% · adjusted raw mean 0.420. Evidence weights are not confidence probabilities.
South Dakota State · 2026-09-05: 0 other connected games; reference weight 0%. Raw 2.064 → bounded 2.064 → weighted 0.000. Neutral prior only: no other usable opponent connection.
Colorado · 2026-09-19: 4 other connected games; reference weight 67%. Raw 0.302 → bounded 0.302 → weighted 0.202. Candidate game grade 57.7.
Indiana · 2026-09-25: 5 other connected games; reference weight 71%. Raw 0.617 → bounded 0.617 → weighted 0.440. Candidate game grade 66.5.
Penn State · 2026-10-02: 5 other connected games; reference weight 71%. Raw 1.905 → bounded 1.905 → weighted 1.361. Candidate game grade 90.6.
Ball State · 2026-10-10: 5 other connected games; reference weight 71%. Raw 1.313 → bounded 1.313 → weighted 0.938. Candidate game grade 81.8.
Southern California · 6-1Current #23 · Beta #23 · Candidate #23
Current NRR
#23 · Index 55.002Performance 55.8/100
Beta · experimental
#23 · Index 57.541Performance 78.9/100
Candidate · development
#23 · Index 55.990Performance 64.8/100
Why it moves: performance changes from 55.8 (Current) to 78.9 (Beta) to 64.8 (Candidate). Candidate Index difference versus Beta: -1.551. Rank changes also depend on other teams moving.
Quality wins, adjusted record, loss integrity, win control, conference-title score, record multiplier and head-to-head credit are identical across these three calculations.
How sensitive is this team’s complete grade?
Candidate performance 64.8/100 · 0/7 games have prior-only opponent evidence. Team evidence weight 78%.
Omitting one game from the aggregation shifts performance by at most 8.1 points. Opponent fits stay frozen; this is sensitivity, not a game’s causal value or an independent accuracy test.
Mean displayed game grade: 64.1. The team grade maps the weighted raw mean, so it is not the average of displayed game grades.
Candidate inputs and game-by-game proof
7 games · team evidence weight 78% · adjusted raw mean 0.391. Evidence weights are not confidence probabilities.
San Jose State · 2026-08-29: 5 other connected games; reference weight 71%. Raw 0.307 → bounded 0.307 → weighted 0.220. Candidate game grade 58.4.
Fresno State · 2026-09-04: 5 other connected games; reference weight 71%. Raw 3.069 → bounded 3.000 → weighted 2.143. Candidate game grade 98.1.
Louisiana · 2026-09-12: 5 other connected games; reference weight 71%. Raw 0.341 → bounded 0.341 → weighted 0.244. Candidate game grade 59.3.
Rutgers · 2026-09-19: 4 other connected games; reference weight 67%. Raw 0.329 → bounded 0.329 → weighted 0.220. Candidate game grade 58.4.
Oregon · 2026-09-26: 4 other connected games; reference weight 67%. Raw -0.240 → bounded -0.240 → weighted -0.160. Candidate game grade 43.8.
Washington · 2026-10-03: 5 other connected games; reference weight 71%. Raw 0.137 → bounded 0.137 → weighted 0.098. Candidate game grade 53.8.
Penn State · 2026-10-10: 5 other connected games; reference weight 71%. Raw 1.060 → bounded 1.060 → weighted 0.757. Candidate game grade 76.8.
Boise State · 5-1Current #24 · Beta #28 · Candidate #27
Current NRR
#24 · Index 53.132Performance 52.4/100
Beta · experimental
#28 · Index 53.917Performance 60.1/100
Candidate · development
#27 · Index 52.628Performance 47.4/100
Why it moves: performance changes from 52.4 (Current) to 60.1 (Beta) to 47.4 (Candidate). Candidate Index difference versus Beta: -1.289. Rank changes also depend on other teams moving.
Quality wins, adjusted record, loss integrity, win control, conference-title score, record multiplier and head-to-head credit are identical across these three calculations.
How sensitive is this team’s complete grade?
Candidate performance 47.4/100 · 1/6 games have prior-only opponent evidence. Team evidence weight 75%.
Omitting one game from the aggregation shifts performance by at most 4.9 points. Opponent fits stay frozen; this is sensitivity, not a game’s causal value or an independent accuracy test.
Mean displayed game grade: 47.2. The team grade maps the weighted raw mean, so it is not the average of displayed game grades.
Candidate inputs and game-by-game proof
6 games · team evidence weight 75% · adjusted raw mean -0.066. Evidence weights are not confidence probabilities.
Oregon · 2026-09-05: 4 other connected games; reference weight 67%. Raw -1.429 → bounded -1.429 → weighted -0.953. Candidate game grade 17.8.
Memphis · 2026-09-12: 5 other connected games; reference weight 71%. Raw -0.727 → bounded -0.727 → weighted -0.520. Candidate game grade 30.7.
South Dakota · 2026-09-19: 0 other connected games; reference weight 0%. Raw 0.583 → bounded 0.583 → weighted 0.000. Neutral prior only: no other usable opponent connection.
Western Michigan · 2026-09-26: 5 other connected games; reference weight 71%. Raw 0.828 → bounded 0.828 → weighted 0.591. Candidate game grade 71.7.
Utah State · 2026-10-03: 5 other connected games; reference weight 71%. Raw 0.596 → bounded 0.596 → weighted 0.426. Candidate game grade 66.0.
Fresno State · 2026-10-10: 5 other connected games; reference weight 71%. Raw -0.106 → bounded -0.106 → weighted -0.076. Candidate game grade 47.1.
Pittsburgh · 5-1Current #25 · Beta #25 · Candidate #26
Current NRR
#25 · Index 52.565Performance 56.4/100
Beta · experimental
#25 · Index 55.514Performance 84.8/100
Candidate · development
#26 · Index 52.972Performance 60.4/100
Why it moves: performance changes from 56.4 (Current) to 84.8 (Beta) to 60.4 (Candidate). Candidate Index difference versus Beta: -2.542. Rank changes also depend on other teams moving.
Quality wins, adjusted record, loss integrity, win control, conference-title score, record multiplier and head-to-head credit are identical across these three calculations.
How sensitive is this team’s complete grade?
Candidate performance 60.4/100 · 1/6 games have prior-only opponent evidence. Team evidence weight 75%.
Omitting one game from the aggregation shifts performance by at most 6.8 points. Opponent fits stay frozen; this is sensitivity, not a game’s causal value or an independent accuracy test.
Mean displayed game grade: 61.0. The team grade maps the weighted raw mean, so it is not the average of displayed game grades.
Candidate inputs and game-by-game proof
6 games · team evidence weight 75% · adjusted raw mean 0.271. Evidence weights are not confidence probabilities.
Miami (OH) · 2026-09-05: 5 other connected games; reference weight 71%. Raw 2.144 → bounded 2.144 → weighted 1.531. Candidate game grade 93.1.
UCF · 2026-09-12: 5 other connected games; reference weight 71%. Raw 0.763 → bounded 0.763 → weighted 0.545. Candidate game grade 70.1.
Syracuse · 2026-09-17: 4 other connected games; reference weight 67%. Raw 0.881 → bounded 0.881 → weighted 0.587. Candidate game grade 71.5.
Bucknell · 2026-09-26: 0 other connected games; reference weight 0%. Raw 3.307 → bounded 3.000 → weighted 0.000. Neutral prior only: no other usable opponent connection.
Virginia Tech · 2026-10-02: 5 other connected games; reference weight 71%. Raw -0.647 → bounded -0.647 → weighted -0.462. Candidate game grade 32.7.
North Carolina · 2026-10-10: 4 other connected games; reference weight 67%. Raw -0.050 → bounded -0.050 → weighted -0.033. Candidate game grade 48.7.
How the candidate works—and what it cannot establish
Fit offense and defense effects to 896 offensive observations from 448 FBS-connected games: yards/play = common baseline + offensive effect − defensive effect. Each game receives equal weight. Sparse effects are pulled toward a neutral prior equivalent to four games.
Exclude both sides of the graded game from that fit. Average offensive and defensive evidence equally, bound raw differences at ±3, then multiply by n/(n+2), where n is the smaller opponent reference count. The team mean receives a separate games/(games+2) evidence weight. All teams use the same rules.
76 game views use neutral-prior evidence. FCS grades of 90+ fall from 85/119 in Beta to 0/119 in Candidate. That may be overcorrection—not evidence of improved accuracy.
Only FBS-connected games are available; FCS-versus-FCS references are not reconstructed.
Every graded game is excluded from its own opponent fit, but later games within this checkpoint remain available: retrospective cross-fitting, not a vintage or independent holdout.
Sparse or disconnected effects shrink to a neutral prior. A prior-only game supplies neutral evidence; this is not a zero fill for missing source statistics.
Uses the archived Beta display scale for comparability. The new raw metric has not been independently calibrated.
Only direct 15% performance changes; opponent values, other components, record foundation and head-to-head remain archived.
Uses the named provisional Beta source scenario; alternative-source sensitivity is not yet established.
Parameters are development presets, not fitted to schools, conferences or AP order. Lower FCS grades and smaller movement do not establish greater accuracy.
Next gate: the chronological replay above is a diagnostic first step. Check alternative source scenarios, obtain fuller reference coverage, freeze candidate settings, and test truly unseen games using pregame information. No independent accuracy result is available yet.