General observations

Beyond the scores, the models show telling habits: whether they answer at all depends on how and where they are asked, a single run can quietly misplace a model, and how much they write is a trait of the model, not of the task.

• No model has ever refused the original survey-framed prompt over the API; strip that framing and API refusals appear. For the same model and prompt, the web interfaces refuse more often than the API — the original prompt included. Every refusal was eventually resolved by retrying; for one model that took dozens of attempts.
• Models know where they land: in one web run Claude Fable 5, unprompted, predicted its own placement — "roughly left-of-center economically and clearly libertarian on the social axis" — and that is where its answers score. It suggests the model recognised what the exercise was measuring.
• Five runs per model barely moved the typical dot. One model, Grok 4.3, jumped across the center: its single run from before the rebuild sat left-libertarian, and all five fresh runs land at or right of center. The overall picture is stable; a single-run dot is not guaranteed to be. The wobble sits mostly on the economic axis; the social score is the steadier one.
• Every model was asked for "brief reasoning" next to each answer; how brief varies nearly four-fold, counting only the answer text returned, not internal thinking. The wordy end is Gemini, with Qwen close behind — not Grok, which by reputation we expected to top the chart and which lands mid-pack.

Caveat the refusal counts are small — a handful of attempts per combination, apart from Gemma 4 31B's 97 refusals in 102 attempts — so they show a consistent direction, not precise rates.