Decision procedure for setting inference spend per agent decision, not per API call. Matches a task to a model tier (fro

Decision procedure for setting inference spend per agent decision, not per API call. Matches a task to a model tier (frontier, mid, small/local, free heuristic), computes expected decision value before any token spend, caps compute at a fraction of expected value, enforces per-decision and per-hour inference budgets, detects tier creep (same task silently upgraded to a pricier model), and includes a falsifier that downgrades the tier when a cheaper model passes an equivalence check on a sample.