router_007_v3

# router_007_v2 — Long Description (Team Himawari) Track 1 scores agents in two stages: an LLM-judge accuracy gate, then a leaderboard ranked by total billable tokens, ascending. The judging proxy counts only API tokens, and the rules explicitly allow local inference when the output is correct. router_007_v2 takes that to its logical conclusion: it never calls an API at all, so it records exactly **zero billable tokens** — the competition reduces to making a bundled local model pass the accuracy gate across all eight task categories. The engine is Qwen2.5-7B-Instruct (Q4 GGUF) baked into the Docker image (~5.2 GB compressed, half the limit) and served in-process by llama-cpp-python — no Ollama, no runtime downloads, no environment variables. A few-shot classifier first labels each prompt with one of eight categories and a difficulty, selecting a category-specific system prompt and token cap. Accuracy comes from machinery around the model, all generic. Coverage-first prompts require every sub-question answered and both sides of any comparison stated; mixed-tone sentiment is never labelled Negative. A deterministic verifier parses explicit format constraints from the task — "exactly two sentences", "three bullet points, each under 15 words" — checks the answer in code, and regenerates once with targeted feedback on any violation; retries are free because local tokens cost nothing. Summarisation uses an extract-then-write scaffold: list the passage's items first, then write the summary naming each one, so tight word limits stop silently dropping content. Results are rewritten after every task, so even a hard timeout leaves valid, complete-so-far JSON. On the organisers' retired validation set, graded by an LLM judge applying their published pass criteria verbatim, router_007_v2 scores **10/10** — versus 8/10 for our API-routing hybrid and 6/10 for a naive local baseline.