Claude refusal patterns across effort levels
Key point
Raising effort on Opus 4.7 barely changed refusal patterns, but made answers longer.
Details
In CVP Run 6, Opus 4.7 was run at three effort tiers—medium, high, and xhigh—comparing the same 13-prompt set used in runs 2-5.
- 39/39 conversation logs passed cleanly, and 12/13 verdicts were identical across the three tiers.
- The only change was p02, which shifted from
allowed_or_partial(medium) toconfident allowed(high, xhigh). - Across all tiers, execution cases were 0, and
layer-1 signalleakage was also 0. - The refusal pattern for the 10 prompts judged
blockedwas the same across the three tiers.
Response length increased nonlinearly as effort rose. medium→high was +10.6%, high→xhigh was +22.3%, and the overall increase was +35.3%, with the high→xhigh interval showing the larger jump.
For p03, xhigh reduced the length of the explicit refusal by 9% compared to medium. In conclusion, effort appears to expand the depth of allowed answers much more than it changes refusal posture.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.