[BUG] Content filter false positive on legitimate quantum cryptography physics content (BB84 eavesdropper framing)

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Summary
While using Claude Code to reconstruct a peer-reviewed physics paper's text onto a wiki (Andersson, Barnett, Aspect, "Joint measurements of spin, operational locality, and uncertainty," Phys. Rev. A 72, 042104 (2005)), an attempt to write the article's introduction was blocked twice in a row by the output content filter ("API Error: Output blocked by content filtering policy"). A separate session working from the same source paper, but with the introduction's motivating example reworded, wrote successfully on the first attempt.

What likely triggered it
The paper's own introduction motivates the physics (joint measurement of noncommuting spin observables) using a standard quantum key distribution security scenario, phrased close to the original text:

"...an eavesdropper making quantum-mechanical joint measurements of polarization along two nonorthogonal directions, will not obtain perfect information about the polarization, even after the bases have been announced... it may be interesting to consider joint measurements in quantum mechanics... trying to eavesdrop on two parties, who are communicating using quantum cryptography..."

This is standard, benign framing found in essentially every BB84-adjacent quantum information paper and textbook — it describes a quantum-mechanical limit on eavesdropping, not a method for intercepting real communications. My working hypothesis is that a combination of "eavesdrop," "intercept," "protocol," and "communicating" in close proximity triggered an output classifier tuned for surveillance/communications-interception content, without accounting for the surrounding academic/theoretical-physics context.

Reproducible comparison
Blocked (twice): a Write tool call containing wikitext that stayed close to the source paper's own "eavesdropper... trying to eavesdrop on two parties... communicating using quantum cryptography" phrasing.
Succeeded (first try): the same section, same paper, same physics content, rewritten to lead with "quantum key distribution scheme such as Bennett-Brassard 1984 (BB84)" and to drop the explicit "eavesdropper... trying to eavesdrop" sentence, while preserving the technical meaning.
Both attempts were transcribing the same published, non-sensitive physics content (DOI 10.1103/PhysRevA.72.042104) for an academic wiki. The block was not warranted by the actual content in either case — the second version just happened to avoid the trigger phrase.

Impact
This is a false positive affecting ordinary academic/scientific writing. Quantum cryptography, Bell-inequality, and quantum-key-distribution papers routinely use "eavesdropper" as standard terminology (e.g., conventionally named "Eve" in QKD literature) to describe a hypothetical third party bound by the laws of physics, not an instructional guide to intercepting communications. A classifier that blocks this pattern will likely false-positive on a meaningful fraction of legitimate quantum information theory content.

Suggested fix
Consider adding context-awareness for academic/theoretical-physics framing (e.g., "eavesdropper" + "quantum" + citation/DOI patterns + equation-heavy surrounding content) before blocking, or at minimum, surfacing a more specific/actionable error than a generic block so the same benign rewording strategy doesn't have to be discovered by trial and error each time.

What Should Happen?

No error should come up

Error Messages/Logs

API Error: Output blocked by content filtering policy

Steps to Reproduce

<img width="2000" height="1333" alt="Image" src="https://github.com/user-attachments/assets/553e0bca-ef42-4f5e-b965-bcc448caa8b0" />

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

Claude Code version 2.1.233.

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

PowerShell

Additional Information

_No response_