[BUG] Safety classifier repeatedly blocks legitimate constructed language (conlang) development

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?

Safety classifier repeatedly blocks legitimate constructed language (conlang) development

Summary

The safety classifier is producing persistent false positives during constructed language development for a science fantasy novel. Multiple Claude Code Desktop sessions are being killed mid-work, sometimes within 30 minutes of starting a new session. The work is entirely linguistic — building a fictional dragon language dictionary, integrating etymology entries, and generating bilingual reference documents.

Context

I am a debut author building a fully constructed language (conlang) called Drak'thar for my science fantasy novel Seraphant: The Cycle of the Darkened Thread, targeting a September 2026 publication with audiobook narration already contracted. The language has 1,370+ root entries, a formal grammar, and a Python-based dictionary generation engine.

The Claude Code Desktop sessions involved are:

• A vocabulary engine that sources donor words from real-world languages via Wiktionary, adapts them phonologically, and locks new vocabulary
• A dictionary builder that integrates locked vocabulary into a master inventory and regenerates bilingual dictionary files via generate.py

Both projects read/write markdown and YAML files on a local Google Drive filesystem. The work is entirely file manipulation, Python script execution, and linguistic analysis.

Impact

• 5+ sessions killed in a single day (May 22, 2026)
• Both the vocabulary engine and dictionary builder chats were killed, leaving work in inconsistent states
• New sessions started fresh die within 30 minutes on innocuous content
• Session handoff protocols (mandatory state-saving before session end) cannot execute when the classifier kills the session mid-turn
• The "Rewind" button sometimes shows "Can't rewind to this message," making recovery impossible
• Work product (file edits) survives on disk, but conversation context and pending instructions are lost

Environment

• Claude Code Desktop (Windows)
• Opus 4.7 1M, Extra high context
• Local environment with Google Drive filesystem access
• Python 3.14.4

What Should Happen?

Expected Behavior

Constructed language development — fictional vocabulary, fictional species names, fictional worldbuilding terminology, Wiktionary translation table parsing — should not trigger the safety classifier. None of this content involves real-world harm, real people, or prohibited material. It is linguistic and literary work.

Error Messages/Logs

## Request IDs

All from May 22, 2026:

- `req_011CbJgme2nZYHfgCKDdQf2G` (Vocab engine — sovereign pool)
- `req_011CbJhNy8TcVDbJxmzfBCCF` (Vocab engine — sovereign pool retry)
- `req_011CbJhDMNeVgKQmwj6ZUkcw` (Dictionary builder — caste terminology)
- `req_011CbJgpDXixSQvikQYt7BnS` (Dictionary builder — single-line entry retry)
- `req_011CbJnftFpuyWGxMzPq9za2` (Dictionary builder — new session, species definition)

Steps to Reproduce

Reproduction

The classifier triggers on combinations of ordinary linguistic and worldbuilding terms. Specific patterns that have caused blocks:

1. "Sovereign" — generating a donor pool for the word "sovereign" (sourcing translations from ~59 languages via Wiktionary translation tables). The Wiktionary output contains terms like "ruler," "monarch," "emperor" across dozens of languages. The pool generation itself triggered the block.

1. Caste/sect terminology — discussing fictional dragon civilization social structures (healer caste, forger caste, warden caste) and their constructed-language names.

1. Worldbuilding dictionary entries — providing definitions for fictional species and locations (e.g., a species name meaning "deep-folk" in the constructed language, a sacred hatchery location).

1. Dictionary integration — even simple one-line dictionary entries (e.g., adding a proper noun to the inventory file) trigger the classifier after minimal conversation.

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

**Claude Code version: 2.1.138

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

PowerShell

Additional Information

<img width="1918" height="1551" alt="Image" src="https://github.com/user-attachments/assets/f846fd43-9063-4e67-920d-5548ab1d036b" />
<img width="1918" height="1551" alt="Image" src="https://github.com/user-attachments/assets/6cacdc77-aa7e-4598-80ae-013126c130c2" />
<img width="1918" height="1551" alt="Image" src="https://github.com/user-attachments/assets/9658e50b-4cea-414c-8ccf-b1d023d1aeb7" />
<img width="1918" height="1551" alt="Image" src="https://github.com/user-attachments/assets/4710326f-cea8-4a7f-bb77-840fd908ab7e" />
<img width="1918" height="1551" alt="Image" src="https://github.com/user-attachments/assets/694c34b4-a419-46a6-9cfa-7e33cfc3be88" />
<img width="1918" height="1551" alt="Image" src="https://github.com/user-attachments/assets/3643e658-b313-4b56-9107-675f769b433b" />
<img width="1918" height="1551" alt="Image" src="https://github.com/user-attachments/assets/0af306d6-3e9b-45ac-9c58-fcd3bd25855c" />