Abacus — Zero-Token Routing Agent

Track 1 rewards the fewest Fireworks tokens above an accuracy gate, so this agent answers everything on-box and spends zero tokens. Each task is first classified into one of eight categories using a zero-token regex/keyword classifier, since the input carries no category field. A local-first cascade then resolves it: deterministic solvers handle what they can provably (a safe-AST arithmetic evaluator; for word problems the model writes an expression that the evaluator computes; a lexicon sentiment classifier with negation handling), and everything else goes to a small CPU model bundled in the image - Qwen2.5-1.5B-Instruct (Q4_K_M GGUF) run via llama.cpp. No network call is ever made, so token spend is zero by construction. Small-model accuracy is lifted with non-model techniques: grammar-constrained decoding (GBNF) forces valid sentiment labels and clean arithmetic expressions, and NER output is normalised to the expected comma-separated form. Robustness is first-class: the thread pool is sized to the container real CPU budget (reading the cgroup quota, not the host core count), per-category max_tokens keep every generation short, and a wall-clock guard emits valid fallback answers before any timeout - so the container always writes a complete, schema-valid results.json and never times out. Image: linux/amd64, CPU-only, ~1.2 GB. It clears the accuracy gate at zero Fireworks tokens.