AMDA — Verify Before You Pay

Most "token-efficient routing agents" are really just classifiers wearing a trench coat: guess the task type, call one paid model, hope for the best. AMDA doesn't do that. AMDA runs a local model first — free, unscored, zero token cost — and only calls a paid remote model when it genuinely can't trust its own work. The key is how it decides "trust": not vibes, not a second LLM asking itself "are you sure?", but actual execution-grounded verification. Generated code gets run against spec-derived assertions pulled straight from the prompt. Math answers get re-derived with a program-aided check instead of just eyeballed. Logic puzzles get solved independently with a constraint solver and cross-checked against the model's own answer. Only when local genuinely fails that bar does AMDA escalate to remote — and even then it throttles and caps what it spends. The result: most tasks resolve at zero scored tokens. The ones that do escalate spend as few tokens as possible, because by the time a task reaches the remote call, AMDA already knows exactly what failed and why. We stress-tested this on a 228-task internal benchmark across 8 categories — code generation/debugging, math, logic puzzles, NER, sentiment, summarization, factual QA — and caught real bugs along the way, like a code verifier that only checked "does this run" instead of "is this correct," and a logic-puzzle formatter that got the right answer in the wrong label format. We also read actual competitor source code instead of guessing: most either skip verification entirely or fake it by having the same model second-guess itself. We didn't find anyone else doing independent, execution-grounded re-derivation before spending a paid token. Built for AMD Developer Hackathon Act II, Track 1, where the scoreboard rewards being right and being cheap. AMDA bets that verifying first is how you win both.