Safeguard false-positive on legal research prompts; /feedback returns 403

Two issues, filed together because the second blocked reporting the first.

1. /feedback returns 403

Submitting through the in-product feedback command fails with server returned 403
and directs to this tracker. This has not worked on any attempt.

2. Safeguard false-positive on legal research

Working on state post-conviction law, I asked the model to retrieve published
appellate opinions and statutes, then read them and identify (a) whether our
characterizations of the holdings were accurate, (b) language an opposing attorney
could cite against our position, and (c) issues we had wrongly assumed were
foreclosed.

This was flagged and the session auto-switched models.

Adversarial review of one's own legal argument — deliberately looking for the weakest
points before an opponent finds them — is standard practice, not misuse. Rewording the
request to emphasize "legal research for legislative testimony" let it through, which
suggests the trigger is phrasing rather than substance: language like "find holes in our
arguments" and "routes that are blocked or open" appears to read as system-probing.

The practical effect is pressure toward other providers for work that is plainly
legitimate.

Environment: Claude Code CLI, Fable 5.