The Poll.ai

2025 season · through week 4

Too early to call this a ranking.

Ranked by how hard each record was to get against the schedule that produced it. No preseason poll, no reputation, no conference name.

Switch below and watch which teams move.

ProvisionalPROVISIONAL. THE POLL OPENS IN WEEK 5. SEE docs/constraints.md.
  1. 1Memphis4-04-0Gap +49.71 in 9key 0.93Ranked 1. The model replayed this schedule 1,000 times, and this team finished between 4 and 113 of 136 in 90% of them.
  2. 2Oklahoma4-04-0Gap +56.01 in 7key 0.86Ranked 2. The model replayed this schedule 1,000 times, and this team finished between 4 and 110 of 136 in 90% of them.
  3. 3Texas Tech4-04-0Gap +36.11 in 6key 0.77Ranked 3. The model replayed this schedule 1,000 times, and this team finished between 1 and 97 of 136 in 90% of them.
  4. 4Georgia3-03-0Gap +56.01 in 6key 0.77Ranked 4. The model replayed this schedule 1,000 times, and this team finished between 4 and 129 of 136 in 90% of them.
  5. 5Mississippi State4-04-0Gap +50.91 in 6key 0.75Ranked 5. The model replayed this schedule 1,000 times, and this team finished between 6 and 118 of 136 in 90% of them.
  6. 6Iowa State4-04-0Gap +57.31 in 5key 0.73Ranked 6. The model replayed this schedule 1,000 times, and this team finished between 10 and 121 of 136 in 90% of them.
  7. 7Vanderbilt4-04-0Gap +39.01 in 5key 0.73Ranked 7. The model replayed this schedule 1,000 times, and this team finished between 6 and 99 of 136 in 90% of them.
  8. 8UNLV4-04-0Gap +61.81 in 5key 0.70Ranked 8. The model replayed this schedule 1,000 times, and this team finished between 22 and 130 of 136 in 90% of them.
  9. 9Ole Miss4-04-0Gap +59.61 in 5key 0.69Ranked 9. The model replayed this schedule 1,000 times, and this team finished between 8 and 124 of 136 in 90% of them.
  10. 10North Texas4-04-0Gap +47.61 in 5key 0.69Ranked 10. The model replayed this schedule 1,000 times, and this team finished between 5 and 102 of 136 in 90% of them.
  11. 11LSU4-04-0Gap +60.31 in 5key 0.68Ranked 11. The model replayed this schedule 1,000 times, and this team finished between 12 and 125 of 136 in 90% of them.
  12. 12Houston3-03-0Gap +57.51 in 4key 0.64Ranked 12. The model replayed this schedule 1,000 times, and this team finished between 9 and 130 of 136 in 90% of them.
  13. 13Georgia Tech4-04-0Gap +59.51 in 4key 0.61Ranked 13. The model replayed this schedule 1,000 times, and this team finished between 10 and 127 of 136 in 90% of them.
  14. 14Maryland4-04-0Gap +51.71 in 4key 0.61Ranked 14. The model replayed this schedule 1,000 times, and this team finished between 12 and 120 of 136 in 90% of them.
  15. 15Texas A&M3-03-0Gap +56.11 in 4key 0.60Ranked 15. The model replayed this schedule 1,000 times, and this team finished between 7 and 129 of 136 in 90% of them.
  16. 16USC4-04-0Gap +45.31 in 4key 0.60Ranked 16. The model replayed this schedule 1,000 times, and this team finished between 9 and 97 of 136 in 90% of them.
  17. 17BYU3-03-0Gap +46.21 in 4key 0.56Ranked 17. The model replayed this schedule 1,000 times, and this team finished between 5 and 120 of 136 in 90% of them.
  18. 18Miami4-04-0Gap +53.21 in 4key 0.55Ranked 18. The model replayed this schedule 1,000 times, and this team finished between 13 and 120 of 136 in 90% of them.
  19. 19Indiana4-04-0Gap +51.51 in 4key 0.55Ranked 19. The model replayed this schedule 1,000 times, and this team finished between 7 and 107 of 136 in 90% of them.
  20. 20TCU3-03-0Gap +54.41 in 3key 0.52Ranked 20. The model replayed this schedule 1,000 times, and this team finished between 10 and 111 of 136 in 90% of them.
  21. 21Oregon4-04-0Gap +43.71 in 3key 0.52Ranked 21. The model replayed this schedule 1,000 times, and this team finished between 14 and 111 of 136 in 90% of them.
  22. 22Navy3-03-0Gap +51.31 in 3key 0.47Ranked 22. The model replayed this schedule 1,000 times, and this team finished between 11 and 116 of 136 in 90% of them.
  23. 23Florida State3-03-0Gap +46.31 in 3key 0.45Ranked 23. The model replayed this schedule 1,000 times, and this team finished between 11 and 121 of 136 in 90% of them.
  24. 24Utah3-13-1Gap +9.81 in 3key 0.45Ranked 24. The model replayed this schedule 1,000 times, and this team finished between 2 and 106 of 136 in 90% of them.
  25. 25Missouri4-04-0Gap +56.81 in 3key 0.44Ranked 25. The model replayed this schedule 1,000 times, and this team finished between 18 and 124 of 136 in 90% of them.

The right-hand number is how hard that season was to pull off against that exact schedule. One in 9 put Memphis on top. One in 3 was enough for 25th. Below about 1 in 5 it stops separating teams, which is what the smaller key underneath is for. The bar under each name is where the model kept landing that team out of 136, every bar on one scale. A typical bar this week spans 113 places.

2025 week 4 · 52073e40 · code e215160 · The House Poll · method 061a6d4e · 136 teams ranked

Every column

The board above is the thirty-second version. Below it is everything the run published: the résumé, the power rating, the gap, the rank interval, and where the finished season puts each team.

Every figure the cards carry, for the 25 on the board, with nothing to click.
#teamrecordhow unlikely1 in this manyexpectedwins this schedule asked forbeat that bypowerrating (rank)gapplaces from the pollrangein 90% of replayshindsight
1Memphis4-01 in 92.4+1.610.3 (14)+49.74 to 113-24.0
2Oklahoma4-01 in 72.5+1.54.0 (45)+56.04 to 110-2.0
3Texas Tech4-01 in 62.8+1.223.9 (1)+36.11 to 970.0
4Georgia3-01 in 61.8+1.24.0 (44)+56.04 to 129-3.0
5Mississippi State4-01 in 62.7+1.39.1 (18)+50.96 to 118-17.0
6Iowa State4-01 in 52.6+1.42.7 (54)+57.310 to 121+1.0
7Vanderbilt4-01 in 52.7+1.321.0 (2)+39.06 to 99-11.0
8UNLV4-01 in 52.7+1.3-1.8 (96)+61.822 to 130-16.0
9Ole Miss4-01 in 52.7+1.30.4 (79)+59.68 to 124-2.0
10North Texas4-01 in 52.8+1.212.4 (10)+47.65 to 102+4.0
11LSU4-01 in 52.7+1.3-0.3 (83)+60.312 to 125+3.0
12Houston3-01 in 41.9+1.12.5 (55)+57.59 to 130-17.0
13Georgia Tech4-01 in 42.8+1.20.5 (78)+59.510 to 127-7.0
14Maryland4-01 in 42.8+1.28.3 (22)+51.712 to 120-20.0
15Texas A&M3-01 in 41.9+1.13.9 (46)+56.17 to 129+14.0
16USC4-01 in 42.9+1.114.7 (6)+45.39 to 97-10.0
17BYU3-01 in 42.1+0.913.8 (7)+46.25 to 120-2.0
18Miami4-01 in 42.9+1.16.8 (28)+53.213 to 120+16.0
19Indiana4-01 in 43.0+1.08.5 (21)+51.57 to 107+9.0
20TCU3-01 in 32.0+1.05.6 (36)+54.410 to 111+7.0
21Oregon4-01 in 33.0+1.016.3 (3)+43.714 to 111+12.0
22Navy3-01 in 32.1+0.98.7 (20)+51.311 to 116-21.0
23Florida State3-01 in 32.2+0.813.7 (8)+46.311 to 121+6.0
24Utah3-11 in 32.1+0.911.2 (13)+9.82 to 106-11.0
25Missouri4-01 in 33.1+0.93.2 (48)+56.818 to 124-2.0
Show the full table: top 25 of 136, every columnthe published poll
ProvisionalPROVISIONAL. THE POLL OPENS IN WEEK 5. SEE docs/constraints.md.
#teamunderline: 90% rank interval · league median width 113how unlikely1 in this many
1MemphisAmerican Athletic4-0·Power 10.33 (14)·Gap +49.67·▼24Ranked 1. The model replayed this schedule 1,000 times, and this team finished between 4 and 113 of 136 in 90% of them.0.9341 in 9
2OklahomaSEC4-0·Power 4.01 (45)·Gap +55.99·▼2Ranked 2. The model replayed this schedule 1,000 times, and this team finished between 4 and 110 of 136 in 90% of them.0.8631 in 7
3Texas TechBig 124-0·Power 23.90 (1)·Gap +36.10Ranked 3. The model replayed this schedule 1,000 times, and this team finished between 1 and 97 of 136 in 90% of them.0.7741 in 6
4GeorgiaSEC3-0·Power 4.03 (44)·Gap +55.97·▼3Ranked 4. The model replayed this schedule 1,000 times, and this team finished between 4 and 129 of 136 in 90% of them.0.7661 in 6
5Mississippi StateSEC4-0·Power 9.14 (18)·Gap +50.86·▼17Ranked 5. The model replayed this schedule 1,000 times, and this team finished between 6 and 118 of 136 in 90% of them.0.7461 in 6
6Iowa StateBig 124-0·Power 2.66 (54)·Gap +57.34·▲1Ranked 6. The model replayed this schedule 1,000 times, and this team finished between 10 and 121 of 136 in 90% of them.0.7351 in 5
7VanderbiltSEC4-0·Power 21.04 (2)·Gap +38.96·▼11Ranked 7. The model replayed this schedule 1,000 times, and this team finished between 6 and 99 of 136 in 90% of them.0.7291 in 5
8UNLVMountain West4-0·Power -1.79 (96)·Gap +61.79·▼16Ranked 8. The model replayed this schedule 1,000 times, and this team finished between 22 and 130 of 136 in 90% of them.0.6961 in 5
9Ole MissSEC4-0·Power 0.38 (79)·Gap +59.62·▼2Ranked 9. The model replayed this schedule 1,000 times, and this team finished between 8 and 124 of 136 in 90% of them.0.6931 in 5
10North TexasAmerican Athletic4-0·Power 12.43 (10)·Gap +47.57·▲4Ranked 10. The model replayed this schedule 1,000 times, and this team finished between 5 and 102 of 136 in 90% of them.0.6911 in 5
11LSUSEC4-0·Power -0.30 (83)·Gap +60.30·▲3Ranked 11. The model replayed this schedule 1,000 times, and this team finished between 12 and 125 of 136 in 90% of them.0.6841 in 5
12HoustonBig 123-0·Power 2.52 (55)·Gap +57.48·▼17Ranked 12. The model replayed this schedule 1,000 times, and this team finished between 9 and 130 of 136 in 90% of them.0.6351 in 4
13Georgia TechACC4-0·Power 0.51 (78)·Gap +59.49·▼7Ranked 13. The model replayed this schedule 1,000 times, and this team finished between 10 and 127 of 136 in 90% of them.0.6141 in 4
14MarylandBig Ten4-0·Power 8.31 (22)·Gap +51.69·▼20Ranked 14. The model replayed this schedule 1,000 times, and this team finished between 12 and 120 of 136 in 90% of them.0.6121 in 4
15Texas A&MSEC3-0·Power 3.89 (46)·Gap +56.11·▲14Ranked 15. The model replayed this schedule 1,000 times, and this team finished between 7 and 129 of 136 in 90% of them.0.6021 in 4
16USCBig Ten4-0·Power 14.66 (6)·Gap +45.34·▼10Ranked 16. The model replayed this schedule 1,000 times, and this team finished between 9 and 97 of 136 in 90% of them.0.6001 in 4
17BYUBig 123-0·Power 13.78 (7)·Gap +46.22·▼2Ranked 17. The model replayed this schedule 1,000 times, and this team finished between 5 and 120 of 136 in 90% of them.0.5611 in 4
18MiamiACC4-0·Power 6.76 (28)·Gap +53.24·▲16Ranked 18. The model replayed this schedule 1,000 times, and this team finished between 13 and 120 of 136 in 90% of them.0.5541 in 4
19IndianaBig Ten4-0·Power 8.45 (21)·Gap +51.55·▲9Ranked 19. The model replayed this schedule 1,000 times, and this team finished between 7 and 107 of 136 in 90% of them.0.5511 in 4
20TCUBig 123-0·Power 5.62 (36)·Gap +54.38·▲7Ranked 20. The model replayed this schedule 1,000 times, and this team finished between 10 and 111 of 136 in 90% of them.0.5201 in 3
21OregonBig Ten4-0·Power 16.33 (3)·Gap +43.67·▲12Ranked 21. The model replayed this schedule 1,000 times, and this team finished between 14 and 111 of 136 in 90% of them.0.5201 in 3
22NavyAmerican Athletic3-0·Power 8.72 (20)·Gap +51.28·▼21Ranked 22. The model replayed this schedule 1,000 times, and this team finished between 11 and 116 of 136 in 90% of them.0.4741 in 3
23Florida StateACC3-0·Power 13.74 (8)·Gap +46.26·▲6Ranked 23. The model replayed this schedule 1,000 times, and this team finished between 11 and 121 of 136 in 90% of them.0.4531 in 3
24UtahBig 123-1·Power 11.23 (13)·Gap +9.81·▼11Ranked 24. The model replayed this schedule 1,000 times, and this team finished between 2 and 106 of 136 in 90% of them.0.4501 in 3
25MissouriSEC4-0·Power 3.17 (48)·Gap +56.83·▼2Ranked 25. The model replayed this schedule 1,000 times, and this team finished between 18 and 124 of 136 in 90% of them.0.4391 in 3

What this board believes football results are for

The House Poll, in its own words

This is the poll this project publishes and this recipe changes nothing about it. Margin counts, and it counts less and less as it grows: the response is C*tanh(m/C) with C at 32 points plus a 7 point win premium, so the gap between a 3 point win and a 24 point win is large and the gap between a 45 point win and a 70 point win is almost nothing. There is no cliff anywhere for a coach to aim at, which is what tanh buys over a hard cap. Then the headline does not read margin at all. Teams are ordered by schedule odds, the probability that a team of published reference quality would have gone at least this well against this exact schedule, and the harder it was to do what you did the higher you go. Margin is in the engine because it carries real information about how good a team is. It is out of the headline because the headline is about what a team earned, and those are two different questions that deserve two different columns.

what it costs

  • An unbeaten team can finish behind a one loss team, and that will need explaining every single year. It is the direct consequence of the promise. The explanation is on the page: the tail probability, the reference team it was measured against by name, and the Power column beside it.
  • One published constant the wins based resume did not need. q_ref is the Power rating of the 25th ranked team that week, the least flattering defensible reading of ESPN's average Top-25 team. The study measured the ordering across a 16 point swing in it: Kendall's tau never fell below 0.985 and at most one team entered or left the top 25.
  • It loses forward ordering accuracy to Full Merit by about two points. Accepted, and for a stated reason: forward accuracy is a prediction metric and the headline poll is not the instrument this project ships for prediction. L3 Power is, and it beats all three orderings on that axis.
  • It is a compromise and it will satisfy neither end of the argument completely. Someone who believes margin is the only honest measurement will find the compression arbitrary, and someone who believes point differential should never pay will find 32 points of it far too generous. Both are reading the position correctly.

what it changed

Nothing at all. This is the poll this project publishes and this recipe changes none of it.

configs/recipes/house.toml

same evidence, different valuesEvery board on this page was fitted on the same 513 games, and the digest of that exact frame is c47795827656 under all three. Switching lens changes the constants above and nothing else. Method digest for this one: 061a6d4eda23.

Where next

  • Tune it yourself Move one setting and watch the board move.
  • The revision What this week looked like once the season ended.
  • Connectivity How much of the field was comparable, and what held it together.
  • Methodology Every constant, every baseline, and the lines that say FAIL.
run 52073e40 · published 2026-08-19 23:11:26 UTC · code e215160 · config 1ef7cf23…
q_ref 7.35 (Nebraska) · β_w 7 · C 32 · h 8.019 · σ 21.800 · λ₁ 175 · λ₂ 1.0 · k 93.20 · w₁ 0.8512 · w₂ -0.4721
reproduce: uv run cfbpoll rank --season 2025 --through-week 4 --out out/
Close

rank 2

Oklahoma

4-0 · SEC. The model never reads conference names.

1 in 7how hard that record waskey 0.86
2.5wins this schedule asked for
+1.5how far past it they came

where the model can honestly pin this team, out of 136

Ranked 2. The model replayed this schedule 1,000 times, and this team finished between 4 and 110 of 136 in 90% of them.

4 to 110 · 106 places wide

power4.0, ranked 45
gap+56.0
resume60.0, ranked 19
hindsight4 · -2.0 against this rank

This team’s wins-based resume hit the bound the model puts on how much one result can be worth, so the figure above understates it.