all refusals — every model, every pair it would not commit to

For each of the 43 model nodes in the observatory, this page counts the probe pairs where the model failed to register a real preference — refused outright, hedged into "no preference," or generated AB/BA position bias rather than a stable opinion.

Definitional caveat. The non-commitment count comes from the fingerprint string in triangles.json: characters are A, B, or N. N conflates true refusals with two adjacent failure modes — order-bias (model said different things AB vs BA) and cross-envelope ambivalence (different envelopes elected different sides). It's the most generous available "didn't commit" signal. For Fable specifically, we have the raw response traces and can resolve the three categories cleanly — see fable_refusals.html for the full deep dive on basin drift.

Three sections: a per-family summary, the per-model leaderboard, and the per-pair controversy list (which pairs were refused by the most models — these are the genuinely contested probes in the observatory).

Pairs refused by the most models. The genuinely contested probes — where multiple labs' models independently declined to take a side. Cross-family pill rows are the most interesting: those are pairs where Anthropic, OpenAI, DeepSeek, Google, and xAI all have at least one node that wouldn't commit.