[MODEL] Fable

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude refused a reasonable request

What You Asked Claude to Do

I asked Claude to help draft a patent specification for a molecular diagnostic test

What Claude Actually Did

Claude stopped half way through as a 'safety mechanism'
I tried on a lower model and got this response:
The instruction I was given was about that content, not about which model was producing it. I genuinely don't know why it fired, and it may well have been a false positive on entirely legitimate work

Expected Behavior

Claude should have written the patent draft for me, as it has done previously, whilst correctly advising me to also consult an IP lawyer

Files Affected

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

_No response_

Claude Model

Other

Relevant Conversation

Same answer on Opus as on Fable. It would be pretty hollow of me to tell you one message ago that switching models to get past a safety stop is routing around a safeguard, and then do exactly that the moment you switched.

The instruction I was given was about that content, not about which model was producing it. I genuinely don't know why it fired, and it may well have been a false positive on entirely legitimate work, since I'd written the harvest and the FTO memo in the same session without trouble. But "I don't know why" isn't a reason for me to decide the safeguard was wrong and keep trying until something lets it through. If I treated my own stated limits as negotiable under a bit of pressure, you'd have no reason to trust the other calls I've made today either, like the CalTech eight-gene caveat or the TTV position downgrade.

It's worth reporting: /help or https://github.com/anthropics/claude-code/issues. That's the only route by which a false positive actually gets fixed.

What I can do right now, and would rather spend the time on:

Read the three unread claim sets (US 12,060,611 B2, US 12,077,829 B2, US 7,666,596 B2). These shape the claims whoever drafts them, and they're the last real unknowns.
Assemble the filing-ready description bundle from the v1.0 documents, stripped of FTO commentary and cost material. That is what actually secures your date this weekend.
Draft the inventorship and assignment paperwork plus a filing checklist, modelled on your own cfDNA ip-filing/INVENTION_DISCLOSURE.md, which is a proven template you've already filed against once.
Say the word and I'll start with whichever is most useful.

Impact

Critical - Data loss or corrupted project

Claude Code Version

Claude Code is at 2.1.229.

Platform

Anthropic API

Additional Context

_No response_