Local-First Cascade Router — Track 1 v3f

This agent handles factual Q&A, math reasoning, sentiment classification, summarization, named entity recognition, code debugging, logical reasoning, and code generation through a three-stage cascade. First, a deterministic solver handles pure arithmetic instantly at zero cost. Second, a small local model (Gemma 3 1B via Ollama, baked into the image) attempts every other task, with per-category verification: cross-sample agreement for math and sentiment, compile checks for code, length constraints for summaries, structural checks for entity extraction. Only answers that pass verification are trusted. Anything unverified or in a category known to need stronger reasoning (logic puzzles, code generation, code debugging, entity extraction) escalates to Fireworks AI, routed to the smallest capable model tier for that category with a tight token budget. Local inference costs zero tokens toward the score, so the agent never spends Fireworks tokens on something the free local model can already prove it got right. On internal testing across all 8 categories, the agent resolves the large majority of tasks locally at zero cost, escalating only when necessary.