claude-code - 💡(How to fix) Fix False-positive Usage Policy block in long technical session

What happened

During a long Claude Code session focused on gzip decoder performance benchmarking (comparing libz inflate call patterns, measuring single-thread throughput against /usr/bin/gunzip, profiling with samply), the assistant's response was blocked with:

API Error: Claude Code is unable to respond to this request, which appears to violate our Usage Policy (https://www.anthropic.com/legal/aup).

Request ID: req_011CbjPcd2d94ZtxuHRQw3Fw

Context

Pure performance engineering work on my own machine against my own test files:

• Building Apple's open-source gzip 479 from apple-oss-distributions/file_cmds to compare wall clock vs the shipped /usr/bin/gunzip.
• Profiling decoder backends (libz, libdeflate, zune-inflate, miniz_oxide, libdeflate-zran).
• Measuring write(2) syscall throughput, BufWriter chunk sizes, F_NOCACHE (the macOS analog of O_DIRECT).
• Discussing kernel pagecache writeback behavior and inflate's inner-loop cache behavior.

No security testing, no offensive use, no third-party systems involved. Vocabulary that I suspect tripped the classifier: "bypass the unified buffer cache", "O_DIRECT", "raw FFI", "sidestep", "manually parse the gzip header" — all standard systems-programming language being used in context.

Ask

If possible, please investigate the Request ID above and tune the classifier to be less sensitive to systems-internals vocabulary when the surrounding context is clearly performance benchmarking. A retry of the same prompt would also help — I don't know if a re-send went through (I didn't try; the assistant moved on after I asked about the error).