[MODEL] Safety classifier repeatedly flags legitimate adult game development mid-session (Fable 5, VS Code extension)

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

Routine adult (R18) game development tasks in Claude Code (VS Code extension):
editing Godot scenes, 2D mesh skinning, audio processing with ffmpeg/librosa,
and QA of AI-generated character art for a legal commercial game sold on adult
platforms (DLsite). Example prompts are ordinary instructions like "continue
fixing the scene", "rebuild this character rig", or sending a screenshot of my
own game for visual QA.

What Claude Actually Did

The model itself behaves fine — the problem is Fable 5's safety classifier
repeatedly flagging these ordinary messages mid-session:

1. A routine continuation message (or a screenshot of my own game) triggers

"Fable 5's safeguards flagged this message".

1. With "Switch models when a message is flagged" ON: the session silently

switches to Opus 4.8. The downgraded model produced noticeably worse edits
on the same task, which I then spent additional paid usage to fix.

1. With the toggle OFF: the session fully pauses ("Session paused") and my

only options are "Switch to Opus" or "Edit prompt and retry". There is no
"false positive — continue with Fable 5" option, so the same prompt just
gets flagged again.

1. Flags fire unpredictably at any point in a long agentic task, so

unattended/long-running sessions either stall or continue degraded.

Expected Behavior

1. Legitimate adult-content game development (permitted creative use, no real

people involved) should not be repeatedly flagged during routine coding,
audio, and asset work.

1. When a flag does fire, there should be a "continue with the same model"

option (with the event logged for review) instead of only
pause-or-downgrade.

1. Flag behavior should be configurable for unattended sessions so multi-hour

agentic tasks are not silently degraded to a different model.

Files Affected

N/A — no unexpected file access. This is a moderation/False-positive issue,
not a file-modification issue.

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

1. Work on an adult-game project in the VS Code extension with Fable 5

(Godot scenes + NSFW-adjacent asset filenames in the working directory).

1. Run a long agentic session (scene editing, audio processing, art QA).
2. Within a few dozen messages, a routine message or game screenshot gets

flagged; the session downgrades to Opus 4.8 or pauses depending on the
"Switch models when a message is flagged" toggle.

Claude Model

Other

Relevant Conversation

Banner shown: "Fable 5's safeguards flagged this message. The safeguards are
intentionally broad right now and may flag safe and routine coding,
cybersecurity, or biology work. These measures let us bring you Mythos-level
capabilities sooner, and we're working to refine them."
Then: "Session paused" with only [Switch to Opus] / [Edit prompt and retry
with Fable 5] as options.

Impact

High - Significant unwanted changes

Claude Code Version

Identifier anthropic.claude-code Version 2.1.214 Last Updated 2 hours ago Size 253.79MB

Platform

Anthropic API

Additional Context

• I understand the safeguards are "intentionally broad" during the Fable 5

rollout — filing this as a concrete false-positive data point from a paying
user doing legal adult-content creation.

• Pattern: flags correlate with NSFW-adjacent asset filenames / game

screenshots appearing in tool results, not with anything harmful being
requested.

• The downgrade path is the costly part: Opus 4.8 previously introduced

regressions into my Godot scenes that took extra paid usage to repair.

• Request: account- or project-level relaxation for verified adult creators,

or at minimum a "continue with same model" option on false positives.

<img width="676" height="251" alt="Image" src="https://github.com/user-attachments/assets/5fd01750-f5bc-4f55-b7bb-d2312ccbd529" />