Veritas Route

Veritas Router attacks Track 1's ranking at its mathematical floor: zero Fireworks tokens. Every task is answered inside the container, so the token score cannot be beaten, only tied. The architecture is a three-tier cascade built on one principle: never trust an answer you can verify. Tier one is deterministic, covering pattern-matched math, brute-force constraint solving for logic puzzles, lexicon-based sentiment with evidence quoting, and regex date extraction for NER. Tier two is a bundled Qwen2.5-1.5B model (4-bit quantised, roughly 1GB) sized deliberately for the 4GB, 2vCPU grading environment. The critical trick: the model never does arithmetic and never ships unchecked code. For math it translates the word problem into a Python expression which the agent executes, making the arithmetic exact. For code generation and debugging, candidate code runs in a sandboxed subprocess before acceptance, with a retry loop on failure. Factual questions pass through a self-consistency gate: two samples at different temperatures, with disagreement triggering a conservative third pass instructed to omit uncertain details rather than guess. A global time budget degrades gracefully near the 10-minute cap and a fail-safe writer guarantees a valid, complete results.json under any failure mode. Full run on the practice set: eight seconds.