ROUTEZERO

RLSym is a hybrid local-first agent for the Hybrid Token-Efficient Routing Agent track. It reads /input/tasks.json, classifies each prompt into one of eight task categories, and selects the cheapest reliable execution path for that task. Closed-form work such as arithmetic, logic, sentiment, entity extraction, and common code repairs is handled by deterministic engines at zero scored tokens. Tasks that require interpretation use a baked Qwen2.5-3B-Instruct model served through llama.cpp inside the container. Local outputs are validated for correctness, structure, sentence counts, bullet limits, JSON shape, entity grounding, and code syntax. A failed local result is retried with a stricter representation before Fireworks is used as an emergency fallback. The same repository ships the evaluation infrastructure that gated every build: a seeded pool generator across all eight categories, an automated judge (numeric tolerance, entity coverage, executable code tests, format gates), and a one-command gate script. No build was promoted without passing the organizer set, a regression pool, and two fresh-seed pools it had never seen - roughly 4,000 logged task executions across 25 configurations. The final linux/amd64 image was tested under the official 2 vCPU and 4 GB limits. It passed the organizer-authored retired set 10/10, completed 384 generated tasks at 100% accuracy with zero blanks, and returned all 64 task IDs during two complete remote-outage tests without hanging. The exact image pulled back from GHCR matched the locally tested binary and passed the retired set again. RLSym's design principle is simple: deterministic computation when certainty is available, local inference when interpretation is required, and remote inference only when both cheaper paths cannot produce a validated answer.