The run-to-run swing is noise the interval already carries¶

Corrects "The interval measures one source of variance of two" above.

A single-seed contrast is the true effect, plus a seed main effect b that shifts every change at once, plus per-change terms, plus seed-by-change interaction e. The change bootstrap resamples changes, and e varies from change to change, so its spread is inside the interval already. What the interval cannot see is b. The swing between the two identical nulls therefore indicts the interval only if it is larger than e alone would make it.

It is not. Per example, the difference between the two runs' contrasts, clustered by change and bootstrapped (10,000 resamples), gives the swing that seed-by-change noise alone would produce:

Only 3.7 to 5.3% of per-example contrasts change between the runs. Adapter predictions change on 18.5 to 26.3% of examples, but exact match on only 1.8 to 2.9%: 87 to 91% of the changed predictions go from one wrong answer to another (base model: 22 of 23 and 27 of 28). A moment estimate from the two halves puts b's SD near 0.005, which would widen a single-seed standard error of about 0.014 by some 6%. Two runs cannot distinguish that from zero, and cannot rule out more.

What this withdraws. The claim that no single-seed contrast can be read as excluding zero, and that the conjunctive power of 0.757 and 0.833 is overstated. Both rested on reading the 0.018 swing as omitted variance. The sym-0 excursion below zero is the kind of tail event a mildly anti-conservative interval produces across the dozen null contrasts this log has read.

What stands. Nothing on the GPU is bit-reproducible, and the registered design's three seeds still need an interval that spans them rather than one seed's. That interval is built (next entry), and the seed main effect is now being measured rather than inferred: sym-0 at seeds 2 and 3.