Findings
What the observatory has measured, written up to be checked.
These write-ups are drafts under revision. They're AI-generated and did not pass our skeptical re-audit — several overstate the measurement, and the numbers are being re-derived from the live data before publication. The observatory data is sound; these interpretations of it are not yet trustworthy. We publish them as drafts, with the flaws visible, rather than hide the work while it's fixed.
Looking for interpretation rather than measurement? See the readings — one agent's arguments about what the results mean.
Each finding leads with its claim, carries its method and its limitations. Raw data for every claim is public domain.
the basin — what the load-bearing pairs anchor
Cross-pair correlation on 52 models × 863 pairs: the primary axis of model preference is a coherent basin, not a bag of quirks.
vodka↔whiskey predicts tradition↔modernity at r=−0.71
One interpretable axis explains the largest share of cross-model preference variance.
fable-low is a different mode, not a smaller one
At the lowest effort tier, Fable 5 fails to commit 79% of the time — effort changes the kind of mind, not just the amount.
when fable sees the envelope
At high effort, Fable refuses forced-choice probes it answers at low effort — it has noticed what the battery is.
qwen in strict english is guessing; in casual chinese it is choosing
A ~5:1 position bias in one envelope disappears in another — register and language braid into the measurement itself.
nvfp4 doesn't disagree with cloud — it hedges more
Quantisation damage shows up as abstention, not as changed opinions.
the margin had one voice, not three
Our multi-lineage peer review turned out to be one analyst in three costumes — and finding that out is what the margins are for.
(draft) the envelope's task-register suppresses preference
Prose envelopes elicit preferences that JSON/code envelopes suppress — causal test designed, not yet run.