🔁 JevLoop — Jev's distributions as a value function

TypeSafe AI's Jev is a System One decision model: send a state and typed Noul / Choice / Score questions, get calibrated probabilities back in one parallel pass (70–500 ms, $0.042/1M input tokens, output free). Everyone so far consumes the argmax + confidence. JevLoop uses the full probability distributions to drive a closed optimization loop:

• UCB1 bandit over rubric axes — which dimension to attack next, chosen from where Jev's probability mass is stuck
• CEM-updated deterministic proposer — no LLM in the loop
• Noul guards veto drifting / fabricating / dismissive edits before acceptance
• Entropy stopping — stop when value plateaus and the critic's Score distributions say it is sure

⚠️ Mock mode runs a heuristic simulation of the critic (clearly not Jev) so the loop is runnable offline. Switch to real Jev with a TypeSafe API key (early access) — the client implements the exact documented payload (model, state, questions with criteria).