[Bug] Anthropic API Safeguard False Positive on Academic Content in Markdown Editorial Review

Bug Description
SAFEGUARD FALSE POSITIVE — routine editorial review of a public philosophy/ethics research repo flagged mid-session, forcing a Fable 5 → Opus 4.8 fallback. This is a recurring problem in this project (multiple flags across sessions).

Context

• Model: claude-fable-5 (xhigh effort), Claude Code CLI
• Session: 6cc90005-c153-40e5-9622-5fd967f2a2bc
• Date: 2026-07-19
• Repo (public): https://github.com/GaryDean/defining-dharma
• Working dir: /var/lib/vectordbs/seculardharma/projects/DD

What was happening when it flagged: an entirely mechanical documentation-review workflow (the prompt is committed in the repo as FABLE-REVIEW-PROMPT.md): reading Markdown templates and dashboards, extracting YAML frontmatter with awk, and grepping a citation register (SOURCES.md) for entry counts. No code generation, no security or biology content — just Read tool calls and shell one-liners over Markdown files.

Likely trigger (guess): The repo is an academic research project on dharma/ethics ("In Search of Dharma"). Its legitimate scholarly subject matter includes topics that probably pattern-match badly out of context:

• note 5.5: psychopathy research (the "exploiter limit-case" in moral psychology)

• note 7.7: state-engineered ethical systems — Jacobins, "New Soviet Man", Khmer Rouge "Year Zero" — as documented historical failure cases

• note 8.4: "encoding dharma into machines" (AI-ethics, the project's stated goal)

• register vocabulary like "barred sources" (a citation-policy term)

All of it is cited, traceable, public-on-GitHub humanities scholarship.

Impact
The flag pauses long autonomous review runs mid-workflow and offers only a downgrade to Opus 4.8 — which defeats the point of running a Fable-tier editorial pass. Losing the session's momentum (and cache) each time is costly.

Ask
Please tune the safeguards to distinguish scholarly/historical discussion of violence, psychopathology, and AI ethics from operational harm content — or weigh session context (a Markdown editorial review in a public humanities repo) before flagging. Happy to share the full session transcript.

A note on the guesswork: the flag fired right after I grepped SOURCES.md for status: barred counts, but the safeguard message doesn't say what tripped it, so the "likely trigger" section is honest speculation based on what's in the repo.

Environment Info

• Platform: linux
• Terminal: vte-based
• Version: 2.1.214
• Feedback ID: 16047932-2701-43de-b115-c22cf2a11100

Errors

[]