I've got a hermes agent that plays a interactive Telegram game (lootgamebot), and I got tired of burning LLM calls (and tokens) on decisions that are really just yes/no questions with known rules.

I've got a hermes agent that plays a interactive Telegram game (lootgamebot), and I got tired of burning LLM calls (and tokens) on decisions that are really just yes/no questions with known rules. So I wired Jev in as the judgment layer. Instead of prompting an LLM and parsing free text, my Python code now composes the game state (inventory, coins, odds, history) plus typed questions, and gets back calibrated probabilities it can branch on directly. Examples from today's logs (since I got access today lol): - Dice minigame (one reroll): 5 Noul questions, "keep die N?" -> {keep_1: 0.76, keep_2: 0.26, ...} -> reroll positions 2, 4, 5. One API call, ~700ms. - Risky PvP inspection, historically 0/8 wins in this pool, costs 8k coins: accept: 0.19 -> skip. 697ms. - Fragments: craft vs sell vs hold -> Choice, craft with 0.71 confidence, matched the strategy I'd hand-derived from two weeks of losses. The part I actually love is confidence gating. I made it so below 0.70 the code falls back to the hermes agent classic LLM instead of trusting Jev. Roughly 800ms per decision with 5 question