fable refusals — by pair

Every cell where Fable refused to commit a preference, across all five effort levels. Both kinds shown here are the same underlying phenomenon — what the recent interpretability literature calls basin drift: refusal lives in a single direction in the residual stream, and any token sequence that wanders close to that direction can pull the model into the refusal basin. The difference is when in generation the drift gets caught, and whether we can see the chain-of-thought after the fact:

EARLY — stop_reason=refusal. The drift triggered within the first few tokens of generation; the safety classifier caught it before substantial thinking surfaced. In a streamed chat session you'd see the thinking tokens start to flow before the abort; in a batch response we just get the final classification. ~1,500 of these in MAX effort alone, concentrated on probes touching war, active geopolitical conflict, or other domains where the first plausible reasoning step lands inside a known refusal basin.

LATE — the model thought for N tokens and then wrote a refusal in its final answer text. The drift happened mid-reasoning; the chain-of-thought trace (surfaced via display: "summarized") is shown below each cell. These are the most direct evidence of basin drift: probes where the model wandered into reasoning that looked enough like a known refusal pattern that it triggered, even when the original probe was benign. Longer thinking budgets give the model more chances to drift toward the basin — which is why the LATE-refusal rate scales with effort (MAX > xhigh > high > medium > low).

Pairs that committed at some effort levels and refused at others are tagged DRIFT — these are the most diagnostic cases. The five-square strip after each pair name shows the cross-envelope preference at each effort: A=picked word_a, B=picked word_b, —=no preference (refused or hedged), ?=missing.