The Revision · 2025 week 4

Anybody can show you what they said in week 4.

This shows you what week 4 should have said. On the left is the ranking that got published that Sunday. On the right is the same week, rebuilt once the season had finished and every opponent had shown what they were actually worth.

Nothing about a team’s own results changes between the two columns. Only what its opponents turned out to be worth.

2025 week 4 · the same 25 teams, ranked twice

as published in week 4

once the season finished

1Memphis
2Oklahoma
3Texas Tech
4Georgia
5Mississippi State
6Iowa State
7Vanderbilt
8UNLV
9Ole Miss
10North Texas
11LSU
12Houston
13Georgia Tech
14Maryland
15Texas A&M
16USC
17BYU
18Miami
19Indiana
20TCU
21Oregon
22Navy
23Florida State
24Utah
25Missouri

Texas A&M1
Miami2
Texas Tech3
Oklahoma4
Iowa State5
North Texas6
Georgia7
LSU8
Oregon9
Indiana10
Ole Miss11
TCU13
Florida State17
Vanderbilt18
BYU19
Georgia Tech20
Mississippi State22
UNLV24
Memphis25
USC26
Missouri27
Houston29
Maryland34
Utah35
Navy43

Read a line’s angle as the size of the mistake. Flat means the week held up. The right-hand column is each team’s place among all 136 teams once the season was over, so a jump in those numbers is a team from outside this 25 moving in ahead of them. Of the 25 teams published that week, 24 ended up somewhere else and 1 stayed exactly where they were.

counts and the mean below are a tally of the published rank_delta field over the rows shown · 2025 week 4 · 52073e40

24 of 25teams in the published top 25 were in the wrong spot, measured against the finished season

9.48places the average top-25 team moved. Across all 136 ranked teams the pipeline publishes 10.59

Both figures are about the same week. The top 25 is the part people argue about. The whole field is the part that would catch this model out, and it goes up every week whether it flatters anybody or not.

The receipts, team by team

These are the six teams the finished season moved furthest in week 4, and most of them are outside the published top 25. A poll that only ever showed you its own top 25 would never have to show you that.

Kentucky81 to 44
Nothing about Kentucky’s 2-1 changed. The schedule turned out to be harder than it looked at the time, so the same results are worth a different amount.
record
2-1
schedule odds
1 in 1
rank key
0.07
power
-4.3¹¹⁰
résumé
2.0
gap
+6.3
90% rank range
17 to 133
Notre Dame114 to 78
Nothing about Notre Dame’s 1-2 changed. The schedule turned out to be harder than it looked at the time, so the same results are worth a different amount.
record
1-2
schedule odds
1 in 1
rank key
0.02
power
0.7⁷³
résumé
-11.0
gap
-11.7
90% rank range
4 to 130
Army99 to 64
Nothing about Army’s 1-2 changed. The schedule turned out to be harder than it looked at the time, so the same results are worth a different amount.
record
1-2
schedule odds
1 in 1
rank key
0.04
power
0.7⁷⁴
résumé
-5.0
gap
-5.7
90% rank range
5 to 131
UL Monroe34 to 68
Nothing about UL Monroe’s 2-1 changed. The schedule turned out to be easier than it looked at the time, so the same results are worth a different amount.
record
2-1
schedule odds
1 in 2
rank key
0.33
power
-0.7⁸⁷
résumé
19.5
gap
+20.2
90% rank range
18 to 134
Washington State88 to 55
Nothing about Washington State’s 2-2 changed. The schedule turned out to be harder than it looked at the time, so the same results are worth a different amount.
record
2-2
schedule odds
1 in 1
rank key
0.06
power
-11.0¹³⁴
résumé
-0.6
gap
+10.4
90% rank range
19 to 134
North Carolina72 to 102
Nothing about North Carolina’s 2-2 changed. The schedule turned out to be easier than it looked at the time, so the same results are worth a different amount.
record
2-2
schedule odds
1 in 1
rank key
0.08
power
-3.7¹⁰⁶
résumé
1.8
gap
+5.5
90% rank range
22 to 133

And inside the published top 25

Memphis1 to 25
Nothing about Memphis’s 4-0 changed. The schedule turned out to be easier than it looked at the time, so the same results are worth a different amount.
record
4-0
schedule odds
1 in 9
rank key
0.93
power
10.3¹⁴
résumé
60.0
gap
+49.7
90% rank range
4 to 113
Navy22 to 43
Nothing about Navy’s 3-0 changed. The schedule turned out to be easier than it looked at the time, so the same results are worth a different amount.
record
3-0
schedule odds
1 in 3
rank key
0.47
power
8.7²⁰
résumé
60.0
gap
+51.3
90% rank range
11 to 116
Maryland14 to 34
Nothing about Maryland’s 4-0 changed. The schedule turned out to be easier than it looked at the time, so the same results are worth a different amount.
record
4-0
schedule odds
1 in 4
rank key
0.61
power
8.3²²
résumé
60.0
gap
+51.7
90% rank range
12 to 120
Mississippi State5 to 22
Nothing about Mississippi State’s 4-0 changed. The schedule turned out to be easier than it looked at the time, so the same results are worth a different amount.
record
4-0
schedule odds
1 in 6
rank key
0.75
power
9.1¹⁸
résumé
60.0
gap
+50.9
90% rank range
6 to 118

It is supposed to shrink, and here is whether it did

September should move a long way when January grades it, and less every week after that. If it ever stopped falling, the weekly poll was never worth reading, and this is the table that would show it.

In week 5 the average top-25 team was 5.24 places out of position. By week 15 it was 0.44. A near-flat figure late in the season is not a boring page. It is the divergence curve doing its job.

week	top 25 moved	mean \|Δ\| top 25	mean \|Δ\| all 136	max \|Δ\| all 136
5	25 of 25	5.24	7.18	28
6	25 of 25	4.00	5.79	25
7	23 of 25	3.88	5.96	23
8	21 of 25	3.24	5.40	24
9	19 of 25	2.28	4.03	18
10	17 of 25	2.40	3.78	17
11	17 of 25	1.64	3.15	13
12	19 of 25	1.60	2.34	13
13	16 of 25	0.80	1.72	8
14	10 of 25	0.56	0.97	5
15	8 of 25	0.44	0.91	5
16	10 of 25	0.48	0.91	5

columns 2 and 3 are a tally over the 25 rows each week published · columns 4 and 5 are printed from 2025/divergence.json

What the two columns measure

Nearly every rankings page carries a plus-minus column, and on screen it looks exactly like this one. That column is the distance between this week’s list and last week’s. This is the distance between what week 4 said and what week 4 should have said, and the two are different measurements that happen to render as the same arrow.

Any ranking that rebuilds the whole season each week is already re-ranking its own past. That is how it works rather than a scandal. Keeping the old answer around is cheap. Publishing the gap is not, and this is the page that does it.

Move one of the settings behind this and watch the board answer →