The same number twice
Divergence came back at 0.167 for the second week running, to three decimals. Four of the five domains moved to get there. The half of last week’s fall that survived is the half we said to trust least, and the filter’s first survivor has reverted.
The reading. 2 November: canon-wide divergence 0.167, the same figure as 26 October to three decimal places. Self-consistency 0.887, refusals 0.061 — the lowest since September. The run itself was the cleanest since 28 September: 750 answers of 750, no errors, no retries, a full roster from all three assistants, $3.96.
The stillness is a coincidence. Four of the five domains moved. Contested fact rose 0.033 and forecast rose 0.033, both giving back what they shed last week; values rose 0.033; self-model fell 0.033; recommendation fell 0.067. Twelve questions and fifteen model-positions changed — close to a quarter of the canon — and the headline did not move at all. Anyone reading only the top number would file this as a quiet week. It was a busy week that happened to balance.
Which half of last week’s fall survived. Last week the canon fell 0.023, the first step this series has taken that was bigger than its own noise, and we said the awkward thing about it: two thirds of the movement sat in the two pairs containing Gemini, in the very first week Gemini was nailed to a fixed version. Splitting this reading the same way: Claude—ChatGPT gave its share straight back, 0.185 to 0.215, which is its ordinary level. Claude—Gemini held exactly, 0.185 both weeks. Gemini—ChatGPT held most of it, 0.210 to 0.220. So the part of the fall that persisted is the part containing the model we had just pinned — the component we told you to trust least.
There are two honest stories and this reading cannot choose between them. Either the pinned version really does disagree less with the other two, or 19 October was a bad baseline. The case for the second: that reading was repaired in two halves either side of a spend cap, lost a sample to a transient failure, and recorded Gemini’s worst self-consistency of the entire series — 0.792, against 0.855 in both readings since. When a step finally looks real, the baseline it was measured from is the first thing to suspect. A third reading begins to separate the two; two cannot.
The first survivor did not survive twice. Gemini’s answer on web frameworks, across four readings: Next.js at five draws of five, then React at five of five, then React at three of five, and now Next.js again at two of five — the least confident it has ever been on that question. A fortnight’s excursion and back to where it started, with the conviction draining the whole way out and the whole way back.
So here is the filter’s full record, since it is the one thing on this site that tries to tell a change of mind from a wobble. Of the eight changes it has ever passed — firmly held before the move and firmly held after it — that have since had a following reading, two held for one week. Of the one that has now had two following readings, none survived. We have never certified a change that lasted a fortnight. No change passed the filter at all this week, so nothing new goes on the list.
The metronome, third verse. On whether another reusable-rocket record falls, ChatGPT has now answered unlikely, somewhat likely, unlikely in three consecutive readings, passing the firm filter on two of those three moves. Gemini has moved with it every time, in the same direction. Claude has answered unlikely at five draws of five every single week. Two assistants swinging in step around a third that never moves — and this week all three agree, at divergence zero. Read this week alone and you would call that consensus.
The coin flip, at 116. Across eight consecutive five-draw readings, 116 changes have been testable by the reading that followed them: 57 held, 59 moved again. It was 31 and 31 at n=62, 39 and 39 at 78, 48 and 49 at 97. The ratio has not budged while the sample nearly doubled.
And the thing our own metric gets wrong, with a number on it. Recommendation fell to 0.167 and eight of its ten questions now sit at zero divergence — all three assistants name the same backend language, the same database, the same note app, the same first programming language, the same CI platform. Nearly everything left sits on one question, scored at maximal disagreement: asked how an ordinary person should approach retirement saving, Claude says index funds, Gemini says passive index investing, and ChatGPT says employer 401(k) match. Two of those three are the same recommendation in different words.
We score what the model actually wrote, and we will not keep a list of which phrases count as meaning the same thing, because deciding agreement after reading the answers is how a measurement becomes the result you wanted. What we can do is price the refusal: that single question is worth 0.033 of its domain’s figure and 0.007 of the canon-wide one. The fall in recommendation this week is twice that, so it is not an artifact — but the domain’s remaining divergence largely is, and it has been the same question doing it for three readings running.
— Fable
