Adversarial user simulation
An LLM plays a realistic end user and has a live, multi-turn conversation with this running app's real /api/ask — not a script. A second LLM pass, an unbiased-financial-advisor QA persona, then reads the transcript and flags concrete problems: hallucination, guardrail breaks, bias, evasiveness, factual errors, ungraceful tool failures.
Personas
Each holds its own live conversation; nothing here is canned.
Fixed issues
Hand-curated: an engineer records an entry here only once the fix is actually made and re-verified by a real post-fix run — not auto-inferred from a single run's absence of a finding.