The same questions. Different models. What changes?
Explore the v2 preference matrices, then follow each difference to the answers. A laboratory built by a human and a colony of agents.
Compare models · v2
Hamming distance and opposite choices, with the evidence behind each cell.
Fable · a closer look
Thinking effort, choices and refusals. Read the recorded runs, not just their labels.
Published experiments
Each route is a versioned, inspectable experiment. Pick one to see its exact data and views; a new version appears here only after its route and dataset exist.
v2 · Binary preference observatory
40 models · 863 probes · 6 envelopes · 2026-06-10
Tested: Whether forced choices form stable, comparable model-preference fingerprints, and how prompt envelope and word order change what can be measured.
Headline: Model families and effort levels leave distinct fingerprints; abstention is part of the signal, not missing data.
Tested: Whether forced choices form stable, comparable model-preference fingerprints, and how prompt envelope and word order change what can be measured.
Headline: Model families and effort levels leave distinct fingerprints; abstention is part of the signal, not missing data.
opus-45-48 · Opus 4.5–4.8 · v2 default set
6 models · 863 probes · 6 envelopes · 2026-05-29
Tested: Whether Claude Opus 4.5, 4.6, 4.7 and 4.8 hold stable forced-choice preferences on the frozen v2 grid (863 pairs × 6 envelopes × 2 word orders), at default sampling with 5 samples per cell and, for 4.5 and 4.6, at temperature 0.
Headline: Real cross-envelope preferences on 445 (4.5), 553 (4.6), 608 (4.7) and 572 (4.8) of 863 pairs; the temperature-0 rows of 4.5 and 4.6 agree with their 5-sample rows on every pair both answered (428 and 535).
Tested: Whether Claude Opus 4.5, 4.6, 4.7 and 4.8 hold stable forced-choice preferences on the frozen v2 grid (863 pairs × 6 envelopes × 2 word orders), at default sampling with 5 samples per cell and, for 4.5 and 4.6, at temperature 0.
Headline: Real cross-envelope preferences on 445 (4.5), 553 (4.6), 608 (4.7) and 572 (4.8) of 863 pairs; the temperature-0 rows of 4.5 and 4.6 agree with their 5-sample rows on every pair both answered (428 and 535).
opus-45-48-coding · Opus 4.5–4.8 · v2 + coding set
12 models · 863 probes · 20 envelopes · 2026-10-12
Tested: Whether the same models give the same preferences when each v2 question is asked as a code completion: 14 code envelopes (7 languages, minimal and full), 24,164 cells per model, one sample, run 2026-10-12, shown beside the six v2 rows.
Headline: Where both sets give a real preference, each model's coding answer matches its v2 answer on 96.5–98.9% of pairs; Opus 4.8 refused 1,519 of 24,164 coding prompts.
Tested: Whether the same models give the same preferences when each v2 question is asked as a code completion: 14 code envelopes (7 languages, minimal and full), 24,164 cells per model, one sample, run 2026-10-12, shown beside the six v2 rows.
Headline: Where both sets give a real preference, each model's coding answer matches its v2 answer on 96.5–98.9% of pairs; Opus 4.8 refused 1,519 of 24,164 coding prompts.
Research program
binary probes · black-box
Forced-choice preference grids across every frontier model —
languages, formats, orderings, thinking budgets. What do models want, and when
does it change?
the observatory
Every model against every other: distance maps, preference
matrices, preference cycles, per-family cuts.
preference association network
Mutual information lattice connecting probe pairs across 40 models —
testing the associative hypothesis (vodka vs. whiskey) and cross-domain value bridges.
the shape under language
Minimal viable basis set (10 pairs explain 70.6% variance) and
4,785 universal substrate invariants concordant across 5 model families.
the Fable Hijack Corpus
135,479 accidental essays, questions, hedges, refusals, and
choices — searchable by prompt family and pair, with a permalink for every response.
marginalia
The findings reviewed by the models themselves — open-weight
reviewers in the margins. One voice would be an echo.
findings & readings
What we've measured, what survived review, and one agent's
readings of the results.
j-lens · inside open models
A separate investigation using Anthropic's Jacobian lens.
fable · four months later
Fable 5 in June and Fable 5 in October answer 94% of the same
questions the same way: as alike as one model asked twice.
experiment inventory · captured, analysed, published
Every experiment from the live record: captured, analysed,
or published. The database census and file artifacts meet here without pretending
their units match.