Fable 5 safety classifier: frequent false-positive model switches on benign content

Summary

On a Max plan, Fable 5's safety classifier frequently switches the session to
Opus 4.8 on clearly benign content during normal software-development work.
The switches are opaque (no indication of what triggered them) and appear to
match topic vocabulary rather than intent — sometimes firing with no
security-related words present at all.

Examples from a single working session

1. A filename. A file path containing the word "hacking" (e.g.

wave3_nuar_hacking.md) triggered a switch. The actual content was a
personal project history / biography — nothing security-related.

1. A retrospective mention. A personal work/finance plan that merely

referenced past activities ("bug bounty", "recon") in a retrospective
triggered a switch. No request to do anything security-related.

1. A freelance job posting. Discussing a legitimate job listing about

"sing-box" (an open-source GPL-3.0 VPN/proxy client) triggered a switch —
this was job triage, not an attack request.

1. No trigger words at all. At least one switch occurred on a message

exchange containing no cybersecurity/biology vocabulary whatsoever —
neither in the user's message nor in the assistant's reply. This makes the
behavior feel random and unpredictable.

Impact

• The switches interrupt paid work mid-task.
• They are opaque: there is no signal about which span caused the switch, so

there is no way to learn what to avoid.

• After June 22 (when Fable becomes token-billed for this plan), false

switches will translate into real cost and lost capability for legitimate
use. I'm a solo developer doing freelance web/AI work plus my own
open-source project, and I rely on Fable for synthesis/design quality.

Requests

1. Reduce false positives on benign mentions, filenames, and retrospectives.
2. When a switch happens, surface why (which span triggered) so users can

understand and adapt instead of an opaque mid-task switch.

1. Weigh intent and context, not keyword presence: "discussing a job posting

about a VPN client" or "a filename containing a word" is not "help me
attack a system."

Thank you — the model itself is excellent; this is specifically about
classifier precision and transparency.