The interval measures one source of variance of two, and a null crossed zero¶

Corrected later on 2026-09-18. The inference below overreaches. The run-to-run swing is what seed-by-change noise produces (z = -1.47 and 0.00 against it), and a single-seed change bootstrap already carries that noise. The only variance it omits is a seed main effect, which two runs give no evidence of. The retraction of single-seed results and of the conjunctive power is withdrawn pending a measurement. See "The run-to-run swing is noise the interval already carries" below.

The most consequential finding of the night, and it is about the gate rather than about organizations.

A pure null excluded zero. sym-0 plants nothing, so at a fraction of zero it is the same experiment as marker-0. Verified identical: one SHA-256 per half across the two, the same held-out examples arm for arm, the same split, training and bootstrap seeds. The two runs disagree:

Nothing on the GPU is bit-reproducible. Final training losses differ in the fourth decimal (a: 0.30191 against 0.30273). Adapter arms change 19 to 27% of their greedy predictions between the runs, 83 to 118 of about 440. The base model, with no adapter, changes 23 to 28 of them. So inference is not deterministic either, although HFGenerator justifies greedy decoding on exactly the ground that exact match needs a deterministic output. The mechanism is documented: Yuan et al. (arXiv:2506.09501, "Understanding and Mitigating Numerical Sources of Nondeterminism in LLM Inference", read from the abstract page) report bfloat16 greedy decoding varying by up to 9% in accuracy on DeepSeek-R1-Distill-Qwen-7B across differences in GPU count, GPU type and evaluation batch size, and attribute it to floating-point arithmetic not being associative. Their remedy, LayerCast, keeps weights in 16-bit and computes in FP32. What was seen here is the stronger case: the same GH200 type and the same configuration, run twice.

The consequence is the gate, not these two runs. The pairs cluster bootstrap resamples changes and holds the trained adapters and the decoder fixed. It measures sampling variance only. The variance from retraining, seed and GPU nondeterminism alike, is about 0.018 on a single contrast, the same size as the organizational effects this study has been reading (+0.016 to +0.034). An interval that omits it is too narrow, and a null crossed zero because of it. Bouthillier et al. (MLSys 2021, arXiv:2103.03098) show that variance from initialization and data sampling "impact markedly the results" and that a comparison needs trials across those sources, not a bootstrap over test data alone. The registered design has three seeds for this reason, but the gate reads the median seed's interval, which is still a single-seed interval.

What the literature does, and why neither half is enough. Du et al. (arXiv:2511.19794, November 2025) bootstrap over per-seed deltas on a fixed test set, the opposite choice: seed variance in, sampling variance out, and three values to resample. At three seeds their protocol never declared a 0.5 to 2.0 point gain significant, even where unpaired t-tests gave p < 0.05. This study's effects are 1.6 to 3.4 points. A two-level bootstrap, resampling seeds and then changes within each, carries both sources and is the interval the gate should read.

What it changes in numbers already reported.

• No single-seed contrast in this log can be read as excluding zero. That includes Qt's +0.034 [+0.010, +0.060] on the full dev window and qt-creator's +0.079 [+0.041, +0.118]. The second is about four times the run-to-run swing and is likely to survive; the first is not.
• The conjunctive power of 0.757 and 0.833 is overstated. It was simulated from example-level resampling of one run, so it too omits seed variance, and the true power at the registered design is lower.
• The gate's non-monotonicity result is unaffected in kind: the swings there (+0.214 against -0.086) are ten times the run-to-run variation.

What it does not change. The calibration's negative control is still clean in the sense that matters: both nulls fail the gate, and the one excursion lies on the refuting side. The contrast does not manufacture a pass from nothing. It can, however, manufacture an interval that looks decisive, which is the reason to widen it.