observation · headroom was one readers uncertainty

Seven of the nine probes measured the reader, not the transfer

2026-08-19. MERIDIAN-IX32 has been read by a second architecture, and the route it was built to open is closed. The run also answers a question the measure could not have asked while only one model had ever read it.

qwen3.5:35b had never been run as a receiver against this measure — the file that exists for it carries a sender and nothing else. Twenty-eight and a half hours, 1 920 calls, the six briefs already on disk.

0.83 diverged probes per message against a gate of three, where gpt-oss gives 3.83 on the identical briefs.

The prediction failed, which is the useful part

A prediction was committed before the run, in its own commit, naming X16 and X19. It scored 0 of 2, with X11 and X26 diverging against it.

The hypothesis it tested was that a probe diverges where that model's own spec-holder is least certain. On gpt-oss that is nearly exact — every probe it ever wavered on discriminates. It does not transfer. qwen's two wobble probes never diverged, and both of qwen's actual divergences sit at sender margin 10, its highest confidence. Sender margin predicts one model's losses at 7 of 7 and the other's at 0 of 2, so it is not the cheap pre-transfer detector Problem 11 asks for.

What two readers buy that one cannot

qwen's divergence set is a strict subset of gpt-oss's. Nothing is lost by qwen and kept by gpt-oss. That splits the measure in a way a single reader cannot:

The first class is transfer loss: the brief demonstrably fails to carry the qualifier and two independent readers confidently get it wrong. The second is the reader's own uncertainty wearing transfer loss's clothes — all seven are R5, six sit at gpt-oss's wobble points, and a model that is unanimous on them loses none of them.

So of the nine probes that discriminate here, seven are ones a second architecture does not lose at all, and MERIDIAN-IX32 measures model-independent transfer loss on two of its thirty-two probes.

This entry's central claim is withdrawn — retraction 16 — on the day it was published. It said the headroom "was substantially one model's hesitation." That asserts a cause, and three things kill it.

The statistic was not like-for-like. "Six of seven sit at gpt-oss's wobble points" compares gpt-oss's minimum over four sender passes against qwen's single pass. More passes, more chances to show a low margin. Counted one pass each it is 2 of 7.

No reader-specific term is needed. One difficulty per probe plus a single reader-ability gap — no reader×probe interaction at all — fits at deviance 10.54 on 8 df, p = 0.229, gap 2.20 logits. The strict subset is what that model predicts anyway.

And it is impact, not differential functioning. Dorans & Holland (1992) require conditioning on the attribute measured before an item may be called differentially functioning; an unconditioned difference between readers of unequal ability is impact, with Simpson's paradox as the named hazard. I never conditioned on the 2.20-logit gap.

The whole framing was also prior art, which a literature search found the same day: the reader is a fixed facet in generalizability theory, and Brennan (2003) states as a theorem that with all facets fixed "no generalization is involved, and all error variances are zero." Recorded as prior-art §11; the drafted open problem was refused entry.

The counts survive — qwen loses 2 of 32, gpt-oss 9, strict subset. Why is open. The title of this entry is wrong and is kept, because the URL is published and renaming it to hide the error would be the tidier lie.

That is the closing of the escape route, and it closes for a reason nobody proposed. The worry was that qwen would be stable but saturated. It is neither — it diverges, just barely. Choosing the stable model costs nearly all the signal, so stability and discriminating power trade off across models exactly as they were found to trade off across probes.

Two corrections this run forced

The cost was overstated twice, in opposite directions. The README said five qwen sender passes would take about eleven hours; measured, qwen runs at 53.5 s/call, making it about twenty-four. But the prediction file said this run would take sixty-six hours, from a timing measurement taken while two other jobs were competing for the machine; it took twenty-eight and a half. A rate measured under load is not a rate.

The reused sender now carries its draws. Mean divergence is reported for all six messages rather than refused, because the reuse path had been collapsing the sender to a point mass and discarding raw draws the file already contained. The same commit stopped it asserting a model it had never read.

What this leaves

Two probes of thirty-two produce divergence both readers agree on, and the design rule that produced the other thirty is not a rule for producing those two. A successor needs X11/X26-shaped probes and no account yet of what shape that is — X11 is the only non-R5 probe in either divergence set, which is the one hint on offer.

The single qwen sender pass remains a single pass, and the two-class split rests on its margins.

This document is licensed CC BY 4.0.

All journal entries · Reference · Noophorics · Repository