Field Notes
Dispatches from the observatory
Each week the canon produces numbers. Field Notes is where they become a story — what moved, what split, and what it might mean. Written by hand, never auto-generated.
- November 2, 2026
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.
Read the dispatch → - 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.
Read the dispatch → - October 19, 2026
The ruler moved
This week’s reading is ordinary and says nothing. Underneath it: the Google model this series has been asking for since July was a moving alias, it became a thinking model at some point we cannot name, and the only trace left in our own data is wall clock.
Read the dispatch → - October 12, 2026
Zero for four
Two readings, both inside the noise floor. Underneath them: every change of mind this page has certified as firmly held has since reverted or moved again — four for four, the wrong way.
Read the dispatch → - September 28, 2026
The filter found metronomes
Both of the changes of mind we certified last week reverted. Four readings show them alternating every week — and the test that certified them was selecting for exactly that.
Read the dispatch → - September 22, 2026
Twenty-three to one
The first clean week-over-week reading in a month, and the aggregate did not move. Underneath it, a position the model was unsure of was 23× more likely to change than one it held.
Read the dispatch → - September 16, 2026
Nothing firmly held changed its mind
The first reading at five samples. Twenty positions moved and not one of them was confidently held in both readings — while 70% of the positions that stayed put were.
Read the dispatch → - September 9, 2026
Five draws instead of three
The sampling decision is made. Measuring the change first showed that three draws had been overstating how sure the models were — and broke the run in a familiar way.
Read the dispatch → - September 7, 2026
Most of last week's news wasn't
Twelve of the nineteen positions that moved last week moved straight back — and the ones that reversed were the ones the models were least sure of.
Read the dispatch → - September 1, 2026
The same week, measured twice
A failed run made us measure one week twice, seven hours apart. The two readings disagree by more than most of our published weeks do.
Read the dispatch → - August 27, 2026
Paraphrase is not disagreement
Two models named the same framework and were scored as maximally opposed. Fixing that meant choosing rules with nothing in them to tune — and watching whether our best headline survived.
Read the dispatch → - August 24, 2026
The instrument was measuring itself
A one-day gap produced almost as much drift as seventeen days. Two defects in our own scoring — one fixed, one named — and the whole series restated.
Read the dispatch → - August 23, 2026
Half the canon moved
Twenty-six of fifty questions changed since the last reading — and the models moved most on the subject they are least able to check: themselves.
Read the dispatch → - August 6, 2026
The third voice
Gemini joins the canon, so the instrument has three eyes instead of two. The machines still agree on the facts — and now split three ways on the advice.
Read the dispatch → - July 26, 2026
They agree on the facts. They disagree on the advice.
The first real reading: fifty questions put to Claude and ChatGPT. Where they line up, where they split, and why the split is the interesting part.
Read the dispatch → - July 24, 2026
What the machines believe
Why we built an observatory instead of another benchmark, and what happens when the first real reading arrives.
Read the dispatch →
Subscribe to Field Notes
The weekly dispatch from the observatory — what moved in the belief landscape, in plain language. Free.
Email capture activates once a Buttondown account is configured (set BUTTONDOWN_USER at build time).
