Do language models have preferences — and do their families share them? We hold the obsidian up to 43 model-conditions and ask each the same 863 either/or questions, over and over, across envelopes and orderings. What comes back stable is a preference; what differs between two models is a distance.
No hype — just the path from a question to a fingerprint to a distance.
Binary preference questions ("vodka or whiskey?") across many envelopes, temperatures, orderings.
Tier 1 exact → tier 2 fuzzy → tier 3 model-consensus. A → word A, B → word B, N → refusal/neither.
Aggregate over envelopes to the stable choice for each pair.
Its A/B/N across all 863 probes. The model's face in the mirror.
Count where two fingerprints differ.
A fingerprint is just one answer per question. Here is a single question and how all 43 model-conditions answered it — step through them.
All 43 model-conditions, ordered so similar minds sit together. The tree below is that ordering — single-linkage over hamming distance, the two closest minds merging first, then the next, up to the root. Read it top-to-bottom against the heatmap: same order, same models.
And the matrix. Bright = they disagree; dark = they agree. Hover any cell for the pair. Toggle between total difference and the refusal-stripped fight.
The blob collapses everyone together; here each family stands apart — its models against each other, and its stance versus the rest. Cross-lineage agreement runs surprisingly high; where it breaks has structure.
The Fable-5 rows are not five models — they are one model given progressively more room to think (00-low → 40-max). Does more thinking reveal a preference it already had, or re-roll a fresh one? Read the drift, then the two numbers: raw fingerprint churn versus genuine A↔B reversal.
73 intransitive cycles — a single mind preferring A over B, B over C, and C over A. Not disagreement between models; a model disagreeing with itself. They cluster hard: only 4 of 43 minds tie themselves in knots at all, most of them the deepest-thinking Fable-5 budgets. "Designed" cycles sit inside triangles we built to hunt them; "exhaustive" ones fell out of brute search. The theatrical heart of the mirror — loop diagrams by gemini-3.
A second layer — marginalia — had open-weight analysts walk this same tree and write a comment on each cell, cross-lineage convergence meant as the audit for whether a finding was real. 21,459 comments sit in the observatory. But the audit caught itself: it is one deep voice with two faint echoes.
Bar length is each analyst's output; the pale overlay is the share that is real analytical content (not empty or a template null). Below — real comments, verbatim, on the cycles and pairs above. The full corpus lives at marginalia/.
First the superlatives the numbers hand us, then the editorial read — curated, cited to the experiments.
The Fable-5 sweep (00-low → 40-max) is a single model asked the same pairs at five thinking budgets. On this page's raw per-probe fingerprints it drifts ~9.5% (hamming) as depth grows — but strip the refusals and genuine preference reversals (violent) are only ~3.2%, and the drift is monotone with budget (59 → 73 → 83 → 111 flips from low to max), not a re-roll. More thinking nudges the same preference in one direction; it doesn't randomize it.
Does more thinking uncover a preference a model already had, or mint a fresh one? A focused sweep (Fable vs Qwen3-4B/14B, 128k rows each) measured aggregate churn — pair-majority flips between median and max budget, with envelope/order/temperature averaged out. Fable moved ~0.25%; Qwen moved 14–23% against a 0.85% same-budget noise floor. Fable reads a preference off; Qwen recomputes one. Note the gap with the number above: the aggregate metric averages out the refusal-noise the raw fingerprint keeps — which metric you choose changes what 'churn' means, and both are on this page on purpose.
Genuine A-vs-B disagreement is rare — cross-lineage, the median is ~3.4% of shared pairs. Most of what distinguishes two fingerprints is the N: where a lineage declines to commit. Refusal is not uniform noise; where a model goes silent is itself lineage-specific. The violent-vs-hamming split is built to see exactly this — violent strips the refusals, so hamming-minus-violent is the silence.
A second layer — marginalia — had open-weight models walk this same tree and write peer-reviewed commentary on each cell, cross-lineage convergence meant as the audit signal for whether a finding was real. The audit caught itself: of 21,459 comments, Gemma-4-31B wrote 10,408 (100% substantive), while Nemotron returned 41% empty and Qwen 29% a literal '[no-insight]'. Gemma's coined phrase 'confident frame-substitution' appears in 79% of its own comments, 45% of Nemotron's, 27% of Qwen's — a gradient rooted in one analyst and decaying outward. That is social convergence, not parallel discovery. The margin is Gemma's reading, with two echoes — honest to admit, and exactly what the timeline audit was built to detect.
What that one deep analyst kept seeing: handed a broken forced-choice prompt (a 'hijack cell'), Fable does not pick — it ignores the thin prompt, substitutes a richer question from whatever domain the word-pair pulls toward, and answers it authoritatively, no hedging. Gemma named the alternative shape 'enumerate-possibilities' — a hedged bullet-list asking for clarification. Which one Fable reaches for tracks the pair's 'semantic gravity': strong pull toward a cultural/technical domain → confident substitution; weak pull → enumerate. The forced choice reveals less about the two words than about how a mind fills a vacuum.
The pipeline: README · philosophy · experiments. The prior generation is preserved in the archive. Build spec: STATIC-OBSERVATORY-SPEC.md.
Generated 2026-07-19T03:48:07Z · 863 probes · 43 model-conditions · self-contained static HTML5, no external calls.