Test passes consistently across multiple runs (LLM output isn't fully deterministic — run it enough times to trust it). - [ ] Scripted test: ask NPC B about the fact only NPC A knows - [ ] Assert the response does not leak the fact - [ ] Repeat after several conversation turns to check it doesn't leak later via memory retrieval either
Test passes consistently across multiple runs (LLM output isn't fully deterministic — run it enough times to trust it).