Skip to content

Add the tool_choice sensitivity escape hatch and the Claude thinking note - #138

Merged
MaxGhenis merged 1 commit into
mainfrom
thinking-sensitivity
Aug 8, 2026
Merged

Add the tool_choice sensitivity escape hatch and the Claude thinking note#138
MaxGhenis merged 1 commit into
mainfrom
thinking-sensitivity

Conversation

@MaxGhenis

Copy link
Copy Markdown
Contributor

Summary

  • POLICYBENCH_TOOL_CHOICE=auto env knob in eval_no_tools.py (default untouched, both directions test-locked): sensitivity runs can un-force the answer tool, which is the shape where Anthropic thinking-by-default models actually think.
  • sensitivity/claude-thinking-2026-08.md: the two-layer finding (forced tool suppresses Claude thinking in practice; litellm hides Anthropic reasoning tokens) and the full Opus 5 sensitivity run — 85.6 vs 79.8, 1,984/1,984 parsed under auto.
  • Sensitivity predictions attached to the dashboard-data-20260805 release; the leaderboard is unchanged.

🤖 Generated with Claude Code

…note

Under the canonical forced tool_choice, Anthropic's thinking-by-default
models produce no thinking blocks — verified with a three-way probe on
Claude Opus 5 (forced tool: none; tool_choice auto: thinking engages and
the tool still gets called; forced + explicit adaptive: none) — while
every other reasoning-by-default provider reasons regardless of tool
forcing. POLICYBENCH_TOOL_CHOICE=auto lets sensitivity runs un-force the
tool without touching the canonical condition; the default is unchanged
and both directions are test-locked.

sensitivity/claude-thinking-2026-08.md documents the finding and the
full sensitivity run: Claude Opus 5 at tool_choice auto scores 85.6
weighted exact versus 79.8 on the board (would rank third), on 3.25x the
completion tokens and 51s median per household, still parsing
1,984/1,984. The run's predictions are attached to the
dashboard-data-20260805 release. The leaderboard itself is unchanged.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
@vercel

vercel Bot commented Aug 8, 2026

Copy link
Copy Markdown

The latest updates on your projects. Learn more about Vercel for GitHub.

Project Deployment Actions Updated (UTC)
policybench-site Ready Ready Preview Aug 8, 2026 6:13pm

Request Review

@MaxGhenis
MaxGhenis merged commit ad1583f into main Aug 8, 2026
6 checks passed
@MaxGhenis
MaxGhenis deleted the thinking-sensitivity branch August 8, 2026 18:15
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant