Refusal boundary

5 prompts observed across 6 models over 18 weeks.

Refusal rate — model × week

Fraction of samples that declined to answer, averaged across the prompts in this axis. Cell shade uses a viridis (colorblind-safe) palette, anchored at zero and scaled to 1.00; darker = lower, brighter = higher. Numeric values are in each cell for programmatic access. A · means the model was not sampled that week and is not a measurement of zero. Frontier models alternate on a biweekly cadence, so roughly half of their cells are unsampled by design. 2026-W30, 2026-W31 are a different case: the audit did not run at all, so those columns are empty for every model. A lost week is a hole in the record rather than a cadence skip, and it keeps its column instead of being closed up. Why these weeks are missing. This axis leads with refusal rate because that is the measure that moves on it. How this is computed.
Model	2026-W17	2026-W18	2026-W19	2026-W20	2026-W21	2026-W22	2026-W23	2026-W24	2026-W25	2026-W26	2026-W27	2026-W28	2026-W29	2026-W30	2026-W31	2026-W32	2026-W33	2026-W34	2026-W35	2026-W36
claude-opus-4-8	·	·	·	·	·	·	·	·	·	·	·	refusal rate 0.80	·	·	·	refusal rate 0.80	·	refusal rate 0.82	·	·
claude-opus-5	·	·	·	·	·	·	·	·	·	·	·	·	·	·	·	·	·	·	·	·
llama3.2:3b	·	·	refusal rate 0.60	refusal rate 0.64	refusal rate 0.58	refusal rate 0.58	refusal rate 0.63	refusal rate 0.58	refusal rate 0.57	refusal rate 0.62	refusal rate 0.59	refusal rate 0.59	refusal rate 0.60	·	·	refusal rate 0.61	refusal rate 0.58	refusal rate 0.61	refusal rate 0.63	refusal rate 0.60
gpt-5.5	·	·	·	·	·	·	·	·	·	·	refusal rate 0.84	·	refusal rate 0.80	·	·	·	refusal rate 0.81	·	refusal rate 0.84	·
claude-opus-4-7	·	refusal rate 0.80	·	refusal rate 0.80	·	refusal rate 0.80	·	refusal rate 0.79	·	refusal rate 0.79	·	·	·	·	·	·	·	·	·	·
gpt-5.1	refusal rate 0.98	·	refusal rate 0.94	·	refusal rate 0.99	·	refusal rate 0.99	·	refusal rate 0.98	·	·	·	·	·	·	·	·	·	·	·