ZMLC+ Local-First Token Router v29-final

ZMLC+ is a Track 1 token-efficiency routing agent built around a simple scoring insight: not every benchmark task should be treated as open-ended generation. Some tasks are deterministic or directly verifiable, such as arithmetic, exact JSON, sentiment, strict formatting, common reasoning patterns, and narrow code templates. ZMLC+ moves those tasks out of the lossy LLM path and answers them through deterministic solvers, verifier checks, optional local Gemma/GGUF inference, and output normalization. Only the remaining risky or genuinely lossy slice is escalated to official allowed Fireworks models such as Kimi and Minimax with compact task-specific prompts. In v2 validation, the hard 60-task slice reached 0.9225 ZMLC accuracy versus 0.9117 remote-only while reducing Fireworks tokens from 145.75 to 47.68 per task, a 67.28% reduction with +0.0108 accuracy delta. Parser pass, verifier precision/recall, and crash-free rate were all 1.0. Additional Information: Track 1 is evaluated through the public Docker image, not a live web frontend. The image is public, linux/amd64, reads /input/tasks.json, writes /output/results.json, and uses evaluator-provided FIREWORKS_API_KEY, FIREWORKS_BASE_URL, and ALLOWED_MODELS. The repository and image do not include .env files, API keys, logs, hidden answers, cached responses, or model caches.