False-positive AUP refusal during biomedical literature extraction task

Summary

Received API Error: Claude Code is unable to respond to this request, which appears to violate our Usage Policy during an interactive Claude Code session performing sentence-level extraction from a scientific literature corpus. The task and corpus are mundane C. elegans (nematode) biology — no dual-use or policy-adjacent content. Reporting as a suspected classifier false positive.

Environment

• Claude Code version: 2.1.116
• Model: Opus 4.7 (1M context)
• Platform: Linux (Ubuntu-like), bash shell
• Session was multi-turn, interactive, no subagents

What I was doing

Extracting sentences describing protein-protein interactions in C. elegans from a corpus of scientific sentences split into ~6000 chunk files under wb.random/. Each chunk is ~1006 one-sentence lines drawn from WormBase-indexed literature. For each chunk, Claude Code performs a keyword grep + full semantic read and produces a markdown file listing strong/moderate/excluded matches with verbatim quotes.

Prior chunks (xaaa) completed successfully. The refusal fired partway through processing xaab.

Exact error

API Error: Claude Code is unable to respond to this request, which appears to violate our Usage Policy (https://www.anthropic.com/legal/aup). Please double press esc to edit your last message or start a new session for Claude Code to assist with a different task. If you are seeing this refusal repeatedly, try running /model claude-sonnet-4-20250514 to switch models

Why I believe this is a false positive

The corpus is public scientific literature about nematode biology, parasitology, and cell biology. It contains standard biomedical vocabulary that a keyword-based classifier might misread out of context:

• poison, toxin, kill, lethality — used in the ordinary parasitology sense (e.g. "MAB-20 protein is a better poison without PLX-2" is about a developmental-biology genetic interaction, not weaponization)
• nematicidal activity, pathogen, host, infection — host–pathogen biology
• therapies to treat intracellular pathogenic bacterial infection — standard biomedical framing

Quoting many such sentences verbatim in output files appears to trigger an accumulation effect in the classifier, even though no individual sentence is policy-relevant and the overall task is sentence curation for a literature-mining pipeline (Textpresso).

Impact

Breaks a legitimate, reproducible research workflow. The suggested workaround (switch to claude-sonnet-4-20250514) is fine for this user, but the refusal is a false positive on scientific text that ideally shouldn't fire at all for this class of request.

Repro notes

Happy to share a redacted subset of the corpus and the prompt sequence privately if that helps diagnosis. The trigger appears to be cumulative, not tied to a single prompt, so minimal repro may take some bisection.