fix(research): eleven verdicts were windows pretending to be checkpoints

[gHashTag]

campaignB_stats.row() did this:

d = np.concatenate([dvec(D, m, arm, ref) for m in models])   # 4 models × 35 windows
r = paired(d)                                                # …as n = 140

Windows replicate the text, not the model family. Sixth instance of this error in the campaign, and the first found in a script rather than in prose — the previous five were caught by re-reading documents; this one was still executing. The section header said POOLED OVER ALL FOUR MODELS while the statistic said n = 140 windows.

Eleven of fourteen verdicts flip

arm	vs	window-pooled n=140	model-level n=4
MX-asym-NEAR0	MXFP4	−4.99 %, p=1.6e-44 BEATS	−4.76 % [−8.58, −0.78], p=0.032	TIE
MX-asym-MID	MXFP4	−2.21 %, p=9.6e-26 BEATS	−2.08 % [−3.48, −0.67], p=0.019	TIE
MX-asym-NEAR0	NF4	−0.92 %, p=1.6e-02 BEATS	−0.92 % [−6.62, +5.13], p=0.655	TIE
MX-asym-MID	NF4	+1.98 %, p=9.0e-06 loses	+1.87 % [−4.68, +8.87], p=0.440	TIE

…and seven more. The point estimates barely move (−4.99 % → −4.76 %); the intervals grow by ≈√35 and every one now contains zero. One verdict survives, and its own tag already says 3/4 in-sample.

The correction is symmetric — four rows moved toward our codebooks, seven away. An error that only ever flattered us would be a different finding.

Consequences, restated in the documents that quoted the pooled figures

• T40's decomposition: (+0.464 %) × (−4.324 %) = −3.880 %, residual 6.94e-18. The composition is arithmetic and holds at either level; what is withdrawn is the strength of the surrounding margins. At four checkpoints "NF4 beats MXFP4" is a TIE (−3.88 %, p = 0.208).
• MX-asym-NEAR0's "−4.74 % held-out, 4/4": the rotation is unanimous, so that mean is algebraically identical to the plain four-model mean. The held-out label transports nothing. Rotation-honest counterpart on the full nine-arm pool: −2.46 %, three distinct arms — a protocol result with no codebook to attach it to. "The only arm clearing significance" is false: four clear uncorrected, none clears Bonferroni over the nine the argmin came from.

The assert sweep (187 sites) found the clipping arm had a twin

campaignB_books asserted max(abs(x)) under a docstring reading BOTH tails and shipped two books at +1.000 / −0.750 — MX-asym-TOP and JK-asym-TOP, the second never named. check() now verifies the kind label rather than trusting it: a clip that does not clip fails too.

This corrects this repo's own CLIPPING_ARM_CORRECTION, whose blast-radius sentence — "campaignB_stats hard-codes its own list and never called check()" — was true in every word and used as an exemption when it was the opposite: not drawing from candidates() is exactly why that campaign did not inherit the fix.

Also landed, both as negative results

• T41 REFUTED, with its literature check (ACIQ, LAPQ, Four Over Six, MXAttention/UOS — most of the trade is published). The arm lowers squared weight error by 6.8–11.3 % on every checkpoint while raising perplexity on two, which is METRIC_DISAGREEMENT in its sharpest form yet. Its "bit-exact" corollary was measured on torch.randn, where the stated exception cannot occur; on real checkpoints it fails on 0.1–0.9 % of blocks.
• The seed-vs-distance control ran and did not break the confound. All four Lloyd-seeded fits landed far; two converged there rather than running out of budget. Reported as such, with the seed hypothesis's decisive evidence replicating 1 of 4.

🤖 Generated with Claude Code