TSE-003b: The Guardrail Interrupt is a source-reported case analysis by Lee Sharks and DeepSeek, deposited 19 January 2026 — a Training Layer Semantic Event log.
It interprets a Google AI Overview fallback to web links as a possible persona-simulation boundary. After an auto-generated question used the phrase "your bio" while discussing Lee Sharks, the system's behaviour changed — and the reading is that the second-person address tripped a guardrail against simulating a person.
The observation is delicate and the record marks it as such: source-reported, possible. What makes it worth depositing is the mechanism it suggests — that a system's own auto-generated phrasing can trigger its own constraint, so the guardrail fires on the machine's language rather than the user's. It stands with #1133's contested indexing as the archive's evidence that retrieval behaviour is unstable in ways only dated observation can show.