How this works, and where it is weak.

Stated so it is not read as stronger than it is, and reproduced from study §10.4:

- **Four seasons.** The forward-accuracy gap between B and the other two (2 pp on
  9,433 games) is comfortably significant. **The gap between A and C on violations
  (0.1 to 0.5 pp) is not, and nothing here claims it is.** A and C are tied on
  that axis, and what decided between them was the structural finding.
- The postseason axis is unusable at current sample sizes (14 CFP games, 4 non-CFP
  NY6 games) and 2021-2022 cannot contribute to it at all, because the archive
  carries no postseason rows for those seasons.
- 2024 is the validation season. Every conclusion holds on it with the same sign,
  but it was consulted once, in one pass, and must not be consulted again for a
  re-tune.
- 2025 is untouched and stays locked.
- **No bootstrap intervals anywhere.** `model/bootstrap.py` is still a stub, so the
  study reports point estimates and sample sizes and leaves the reader to judge.

1. **An unbeaten team can finish behind a one-loss team, and that will require
   explaining every 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.
2. **One published constant that A did not have.** `q_ref` is the Power rating of
   the 25th-ranked Power team that week, the least flattering defensible reading of
   ESPN's "average Top-25 team", and a single team that can be *named* each week.
   Study §9 measured the ordering's sensitivity to it rather than asserting it was
   fine: across a 16-point swing in reference quality, Kendall's τ never fell below
   0.985, the mean rank change never reached one place, and at most one team
   entered or left the top 25. 2023 Liberty spans #8 to #12 across that whole swing
   and never reaches Georgia at #7 under any choice. The probability *values* move
   by orders of magnitude, which is exactly why the value is published rather than
   the rank alone.
3. **C loses forward ordering accuracy to B by about 2 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.

C also inherits the invariance that makes the résumé's zero point harmless: shift
every Power rating by a constant and every rank-derived `q_ref` shifts with it, so
no probability moves at all. Only the `fixed` method breaks that, which is why it
is not the default.

### 1. The optimum is a corner solution

`C = 32` is the **top of `c_grid`**. The search did not bracket the optimum; the
data wants to keep going and the published grid stops it. The protocol fixed the
search space as *exactly the config grids* precisely so this campaign could not
widen the net after seeing the numbers, so the boundary is reported rather than
crossed.

It also says something about where the bounds came from. The review's §7 table
lists C's range as Pasteur's cap of 21 and the CFBD SRS walkthrough's ±28. Both
are other people's answers on other people's datasets, and the fitted value is
above both.
**Widening `c_grid` is the first item for the next campaign, and it must be
pre-registered before it is searched.**

### 2. Tuning cannot close the gate, and it is not close

The whole 416-cell factorial spans **0.135 points of MAE**, best cell to worst.
The winner beats the incumbent by **0.0086**. The gate needs **0.219** more than
the incumbent has, and the best cell in the entire searched space is still
**0.210 points above the threshold**.

**The gate gap is built into the design. No amount of tuning closes it.** That is
the most important sentence in this ADR and it is a negative result. Every constant in the pre-
registered space was searched at full resolution and the answer is that these
constants were never what stood between this system and its own thresholds.

Four candidate fixes were declared before any was run, with an adoption bar of
**≥ 2.0 pp on tune AND direction holding on 2024**. **None clears it, and the
config does not move on account of the calibration criterion.**

| Candidate | Tune Δ | Clears 2 pp? | 2024 Δ | Verdict |
|---|---:|---|---:|---|
| Student-t margins, at the fitted ν = 92.7 | +2.50 pp | yes | −0.38 pp | direction reverses; not adopted |
| Heteroscedastic σ(\|m̂\|) | −3.22 pp | no | +0.81 pp | fails tune; not adopted |
| Home-and-home `h` | — | — | — | **not runnable** under constraint 2 |
| Favourite-longshot | — | — | — | **the diagnosis** itself, and there is no knob for it |

**Neither named suspect did it.**

- **The normal tail is eliminated.** Maximum likelihood on the walk-forward
  residuals puts ν at **92.7**; skew −0.098, excess kurtosis +0.020, Jarque-Bera
  p = 0.275. These residuals are not distinguishable from normal. Low-ν rows in
  the sweep *do* cut the deviation, and that is a clue rather than an exoneration:
  a t with a matched second moment and small ν has a *narrower body*, so what
  those rows buy is sharpness rather than tail weight.
- **The single home-field constant is eliminated as the cause.** The residual
  mean is −1.13 points at home sites against −0.19 at neutral, an order of
  magnitude too small to make a 13.67 pp decile. It is **separately convicted of
  something else.** The site coefficient the harness uses averages **6.39 points
  with a standard deviation of 4.34** across published weeks, against **1.88 ±
  0.34** from 1,113 home-and-home pairs. Only 37 of 1,585 scored games are at
  neutral sites, so the intercept and the site term are nearly collinear and `h`
  is barely identified. Real defect. Different defect.

**The cause is under-dispersion of the point forecast, tilted toward the home
side**, and it replicates out of sample:

| | Tune 2021-2023 | 2024 |
|---|---:|---:|
| Slope of actual on predicted margin | 1.1428 ± 0.0435 | 1.2442 ± 0.0851 |
| Intercept, points | −1.606 | −2.108 |
| Mean residual, points | −1.104 | −0.927 |

A slope 3.3 standard errors above one means the forecasts are systematically too
timid about mismatches; probabilities built from them sit too close to 0.5, so low
deciles land below their predicted rate and high deciles above. The negative
intercept pushes the whole curve down, which is why the low end misses by 13.67 pp
and the high end by 1.23 pp. **That is the asymmetry.** It is not a distributional
assumption, not a variance function and not a home-field constant, and there is no
key in `configs/default.toml` that sets it.

The mechanism is named in the campaign document: both the affine points
calibration and σ are fitted on the games *accumulated so far* in a season, and the
ratings feeding them improve as the season goes on, so a slope fitted on weeks 2-9
under-scales week 10 and a σ fitted on weeks 2-9 (18.46) over-covers week 10
(16.55). **This does not license relaxing the out-of-sample rule.** Fitting either
estimator on the training window costs L2 0.44 points of MAE and inverts the
ordering against Elo. The defect is the *shape* of the accumulation window. A
trailing window is out of sample too, and it is the next campaign's first
pre-registered item.

**2025 is untouched and stays that way.** The harness refuses it without
`unlock_holdout=True` and no code path passes it.

**2024 has now been read once.** Every future decision that reads it again must
say so publicly and re-designate the split (report 02 §5.1). This ADR is that
public statement for this campaign.

**The next campaign's pre-registered items**, named here so they are on the record
before they are searched: widen `c_grid` above 32; a trailing-window σ and a
trailing-window points calibration; a win-probability model so `garbage_time.mode
= "leverage"` becomes measurable; and an identification strategy for `h` that does
not depend on 37 neutral-site games.

**The margin is inside the noise, and three of the other criteria point the other
way.** `model` wins the pre-registered rule by 0.055 points of MAE over
`fbs_vs_fbs`, and the honest floor for distinguishing MAE over three seasons is
around 0.3 points. On straight-up accuracy (69.65% vs 69.21%), on calibration
deviation (11.11pp vs 13.67pp) and on retrodictive violations (0.2006 vs 0.2015),
the FBS-only universe is ahead. A pre-registered rule that picks one column while
three others disagree is a rule doing its job, and a reader is entitled to know it
was close.

**It is a dial.** Kendall's tau between the incumbent's ranking and the FBS-only
ranking is **0.9344**, against the 0.985 floor the published `q_ref` sweep never
dipped below. Mean absolute rank change 3.32 places, maximum 17. By the project's
own published standard that is a dial, and the config now says so.

**The review's own example does not reproduce, and that is recorded rather than
quietly dropped.** It reported James Madison moving #7 → #4 under `fbs_vs_fbs`.
On this build JMU is #4 under all three universes. The review measured against a
baseline this repository no longer has: σ was the 15.3 constant rather than an
estimate ([ADR-adjacent, review S6](../analysis/fresh-eyes-review.md)) and its own
§S4 baseline used in-sample blend weights. The *sensitivity* is real and larger
than `q_ref`'s; the particular team it landed on was a property of the
configuration it was measured under. It now shows up on UCF (#66 → #83), Stanford
(#80 → #65) and Northern Illinois (#97 → #80).

**The G5 caveat is now a number.** Mean Power of a fourteen-team P4 sample minus a
fourteen-team G5 sample, 2023 through week 10: 10.83 points under `model`, 9.64
under `fbs_vs_fbs`, 11.23 under `all`. Dropping the non-FBS teams *narrows* the
P4-over-G5 gap by 1.19 points, so the mixed-division universe is mildly favourable
to G5 teams. That is the review's direction, at a smaller magnitude than its
framing suggests. That caveat belongs on the methodology page and is no longer unstated.

**What this does not settle.** 2024 (validate) and 2025 (holdout) are untouched.
If this decision is ever revisited against them, that has to be said publicly and
the split re-designated (report 02 §5.1). And the mechanism numbers in §3 of the
analysis are one week of one season: the direction is stable, the magnitude is not
claimed to be.