At the test window's size each H1 cell detects about three exact-match points, and absence is out of reach¶

scripts/decomposition_sensitivity.py simulates one H1 cell as the gate reads it: the placebo's own-against-sibling contrast on each half as the variance model, the test window's projected changes split between the halves in their dev-window proportion, five seeds at each organization's half-size seed effect, and the stratified crossed interval. Marginal power 0.928 per cell, so three independent cells pass together about four times in five. Generated from decomposition-sensitivity.json (100 trials a step, 1000 resamples).

The detectable effect is two to three times the smallest effect of interest, and no null study in the simulation read absent: the interval at the test window's size is wider than the (-0.01, +0.01) band, so a cell can pass or be inconclusive and cannot be absent. The spec named this case and committed to stating it before the test rather than discovering it after. Every pre-committed negative reading ("an organization adds nothing", "no transferable style") rests on absence, so as registered those readings could never be reached. What replaces them is the next entry. One caveat on the variance model: Qt's placebo halves trained on about 4,200 examples each, twice the half size every adapter will train at, so Qt's row describes a better-trained adapter's variance than the design will have.

H1's own regime holds nominal. The coverage runs with the second half's arms swapped (--flip-second), as H1 swaps them, so a seed shift enters the halves with opposite signs. Generated from stratified-coverage-h1-0.975.json and stratified-coverage-h1-0.95.json.

Within Monte Carlo error of nominal at both levels, in both regimes now measured.