audit · nine templates not thirty two probes

The measure was nine templates and every p-value counted thirty-two

2026-08-18. MERIDIAN-IX32 has thirty-two probes. It does not have thirty-two independent readings, and three published significance levels were computed as though it did. They are withdrawn as retraction 15.

The defect is visible in two lines of the probe file:

One token apart, opposite verdicts. X21/X22 differ the same way (8 → 7) and X24/X25 likewise (30 → 21). These are minimal pairs straddling a numeric threshold, which is careful probe design and ruinous arithmetic: Fisher's exact test over thirty-two rows treats a pair that shares 99.5% of its text as two independent elicitations.

Single-link clustering of the prompts gives nine clusters at similarity 0.80–0.85, the largest holding eleven probes, with all observed instability inside two of them.

Why there is no corrected number

The instinct is to recompute at the cluster level and publish that instead. It does not survive contact with the data: the cluster-level p runs from 0.111 at threshold 0.80 to 0.006 at 0.95. Any figure quoted is a report on where the clustering line was drawn, not on the measure. So the entry withdraws the p-values and substitutes nothing, which is the honest shape of this particular correction.

That makes #15 a kind this ledger had not held. Retractions 12, 13 and 14 withdraw a quantity — a number replaced by a better number. This one withdraws a warrant: the arithmetic was performed correctly on a sample without the structure the test assumes. It also reaches backwards, because one of the p-values it strikes was retraction 14's own supporting evidence.

The counts are untouched. Four of nine is still four of nine.

What survives is sharper than what it replaces

Of 1 440 gpt-oss:120b sender draws, nineteen are non-key, and all nineteen fall on R5-tagged probes — the paired-submission rule. No probe outside R5 has ever returned a margin below 10, in any pass, at any gate. The individual draw is a unit that genuinely is independent, and the concentration is not an artifact of where a threshold sits.

Two things keep that from being another post-hoc tag hunt. The R5 label was committed at 2026-08-10 09:29:56, nine hours before the first IX32 sender pass — prior to every instability datum, and not drawn to fit it. And selecting R5 from among ten rules is corrected by exact enumeration over all C(32,4) = 35 960 relabelings, which returns p = 0.0016. The uncorrected 0.00042 was never a measurement: C(6,4)/C(32,4) is what any six-probe family capturing all four would print, so quoting it to four figures implied a resolution the design cannot deliver.

None of this reopens the repair route. R5 carries 45 of the 49 divergence events, so the region holding all the instability holds nearly all the signal. What it corrects is a generalisation: "harder is where the instability lives" should read "one rule is where both live."

Three things found while looking at something else

The reuse guard asserted a model it had never read. headroom_check.py printed "sender reused from X (same model, spec, measure and n)" while comparing only the draw count and the measure id. The function never returned the model, so nothing could have compared it — a gpt-oss sender would have been accepted verbatim under --model qwen3.5:35b, with the log affirming the match. Not a gate that cannot fail: a claim printed where no gate existed.

The exposure log recorded half a measure for eight days. 157b80f took the probe file from sixteen probes to thirty-two and did not touch EXPOSURE.md. The file already contained the paragraph explaining why this happens — "the exposure grew because the agent kept working" — written six days before the drift it describes. It was caught by an agent reading the log in order to add its own row, which is luck rather than a mechanism. Nothing checks that file.

The next run costs five times what the README says. Measured at the runner's real settings, qwen3.5:35b is 123.9 s/call against gpt-oss:120b's 5.8 s/call, a factor of twenty-one. "About eleven hours" for five qwen sender passes is fifty-five.

What is running now

qwen3.5:35b as receiver on MERIDIAN-IX32, which had never been done — the existing qwen file carries a sender and nothing else. A prediction was committed before it started, turning on one probe: X19 is tagged R8 and not R5, it never diverged for gpt-oss across twelve messages, and qwen's sender wobbles on it. If divergence follows a model's own sender margins, X19 diverges. If it follows the rule, it does not.

The first of six messages has landed and the prediction is already wrong in an interesting direction. On fluent0 qwen diverged on X11 and X26 — the same two probes gpt-oss diverged on, from the same brief. Neither is one the prediction named. Since every sender mode equals the key in every pass, that means two architectures independently lost the same two qualifiers, which suggests divergence is a property of what the brief drops rather than of who reads it. One message is one message, and X19 has five more chances.

This document is licensed CC BY 4.0.

All journal entries · Reference · Noophorics · Repository