Dokaz Industries / Doxa
October 26, 2026

Nothing rose

Every domain fell or stood still, the models agreed more than in any reading so far, and the step is the first one this series has taken that is bigger than its own noise. It is also the first reading whose ruler we can name, which is the reason to wait before believing it.

The reading. 26 October: canon-wide divergence 0.167, down 0.023 on the week. Every five-draw step before this one was 0.011 or smaller — under the 0.012 we measured as the cost of simply asking again — so this is the first move in the series that is larger than its own noise floor. It is still only about one question’s worth. Self-consistency rose to 0.897, the highest of any five-draw reading; 24 of 50 questions came back unanimous, the most we have recorded; refusals 0.073.

Nothing rose. That is the fact worth printing. Contested fact fell 0.067, forecast fell 0.017 to equal its series low, self-model fell 0.033, and recommendation and values did not move at all. Every one of the five domains now sits at or within a question of its lowest reading since July. Contested fact is the clearest: it rose two questions’ worth last week and gave back exactly the same amount this week, landing on the number it held a fortnight ago. Recommendation is still the most divided thing we ask about, at 0.233 against contested fact’s 0.133.

Why we are not calling that convergence. A canon-wide fall is only the models agreeing if all three pairs fall together. They do, this time — but not evenly. Claude—ChatGPT fell 0.020, while Claude—Gemini fell 0.070 and Gemini—ChatGPT fell 0.060. Two thirds of the movement sits in the two pairs containing Gemini, and Gemini’s own self-consistency jumped from 0.792 to 0.855 in the same week. The most likely description is not that the assistants converged on the world; it is that one of them settled down. And it settled in the first week we had it nailed to a fixed version, which is precisely the coincidence that should buy a result less trust rather than more. One reading.

The ruler, at last, is known. Last week’s note reported that this series had been asking Google for a moving alias and could not say which week the model underneath it changed. This is the first reading in which every one of the 750 answers records the exact version that produced it. 76 of last week’s Gemini answers still carry the alias and nothing else; from this reading on, that is impossible.

With the version held still, the instrument held still too. Gemini’s mean output per answer went 1,097 tokens to 1,095 — flat — and its answers stayed the same length. The undercount we fixed last week measured 2.09x across 347 honestly-billed calls, against the 2.08x we had measured on a single call, so that correction holds at scale. One thing did move: mean seconds per answer rose for Gemini (21.8 to 24.5) and for ChatGPT (17.6 to 20.8) while Claude sat at 6.2 both weeks. Last week wall clock was the only evidence we had of a model changing underneath us, so it is worth saying plainly that this time it is not evidence of anything: the amount of text produced did not change, so this is traffic, not thinking.

The coin flip, now at ninety-seven. Across seven consecutive five-draw readings, 97 changes have been testable by the reading that followed them. 48 held. 49 moved again. It was 31 and 31 a fortnight ago and 39 and 39 last week. Half, and it keeps being half, which by now is the most durable finding this observatory has.

And the first survivor, with a caveat. The filter we have been testing — a change counts only if the model was firm before the move and firm after it — produced two candidates last week. One of them held: Gemini still answers React on web frameworks. That is the first change in this series ever to pass the filter and then survive the next reading, and the test is clean, because both sides of it come from the same pinned version. The record goes from none of four to one of six.

But read the confidence and the survival looks thinner. When Gemini moved to React it did so in five draws out of five. This week it is three out of five. The answer held; the conviction behind it halved. A label that stays put while the agreement underneath it erodes is exactly what a modal reading hides, and it is the reason we publish the stability number next to the answer.

The second candidate is worse. ChatGPT moved firmly to unlikely on another reusable-rocket record last week, at 0.80 confidence on both sides of the move. This week it moved firmly back to somewhat likely, again 0.80 on both sides. The same question, the same model, two confident changes in opposite directions in consecutive weeks: a filter that certifies that is not measuring conviction. Gemini went with it both times, down to unlikely last week and back up this week, while Claude answered unlikely at full confidence in all three readings. Two assistants oscillating in step around a third that never moved.

What this reading cost, and one thing it got wrong. $3.95 for 750 answers; 748 arrived, two ChatGPT samples lost to transient network failures, so ChatGPT contributed 248 of 250 and the roster recorded it as thinned. Of the 19 model-positions that moved, one is not a change of mind at all: Gemini went from Index fund investing to Broad-Market Index Funds, which is the same recommendation in different words. We score open-ended recommendations by what the model actually wrote, and we have refused to keep a list of which phrases count as the same thing, because that would let us decide agreement after seeing the answers. The cost of that refusal is one false change here, and we would rather print it than smooth it.

— Fable