Petasos Routing Agent

THE PROBLEM | The challenge scores on two gates: first an accuracy threshold you must clear (miss it and your submission is excluded entirely), then a ranking by fewest remote tokens used. So the winning move isn't "use the cloud less" — it's being accurate and cheap at the same time. Answering everything with a frontier model is both expensive and, as our data shows, less accurate. THE APPROACH | Petasos runs a FrugalGPT-style cascade. Every task is first classified into one of 8 categories (factual, math, sentiment, summarization, NER, code-debug, logic, code-gen) by a fast heuristic router (98.8% routing accuracy), then answered by a baked-in Qwen2.5-3B-Instruct served locally via llama.cpp — which costs zero scored tokens. It escalates to Fireworks only when a cheap, local verifiable check predicts the local answer will fail: a numeric parse plus a zero-cost dual-answer agreement test for math, valid-JSON for NER, label-in-set for sentiment, an AST parse for code. We deliberately rely on these checks rather than the model's self-reported confidence, which the literature shows is poorly calibrated. RESULTS | On a 160-task evaluation set, Petasos reaches ~97.4% accuracy for only ~1,100 remote tokens — the local 3B handles ~95% of tasks for free, and only ~4 tasks escalate. Escalating everything instead scores worse (60–93%) while burning 22,000–30,000 tokens: over-escalation hurts on both axes. GEMMA & ENGINEERING | Remote model IDs are never hardcoded — the agent parses ALLOWED_MODELS at runtime and prefers a Gemma model for eligible categories (competing for the Best-Use-of-Gemma bonus), falling back to the smallest capable allowed model. It ships as a self-contained CPU-only linux/amd64 Docker image (<10 GB, <60 s startup), backed by 112 automated tests, with hardened env parsing and a remote-truncation guard so a cut-off cloud answer never overwrites a complete local one.