@[Next Big Creative] Thanks — that’s exactly the tension.
My rule of thumb: anything that must be fast, irreversible, or easy for an LLM to rationalize away stays outside the model.
In this bot that means:
- Hard stop / flatten — band breach or max loss. No LLM veto.
- Universe / capital-fit gates — quote must be USDC/USDT, min fee tier, deny whale/low-fee books. Ranking may use AI later; eligibility is code.
- In-band ride — if price is healthy inside the band, temporary
fee < IL cannot force exit (the paper-hands bug).
- Enter path — no deposit if fee/capital-fit fails; we also skip calling the LLM when there’s no enterable pool (don’t burn tokens to “think” about a blocked book).
What the LLM can do: overlay verdicts on ambiguous holds, soft risk language, escalate on toxic/force-exit/enter-candidate moments — advisory, never sole authority on exit/enter.
Heuristic I use: if a wrong call costs gas + slippage and a human would say “that’s just noise,” it belongs in deterministic policy. If the call is judgment under incomplete evidence, LLM may advise — but policy still owns the button.