Let OpenAI, Anthropic, and Gemini adapters own their structured-output parameters while mapping into one versioned domain schema and error model.
Turn Qwen3 hybrid thinking into a task router with bounded time and tokens, paired evaluation, deterministic verification, and explicit fallback.
Evaluate Mistral Small 4 for constrained tool routing while accounting for open weights, configurable reasoning, strict validation, and its real infrastructure floor.
Use a small deterministic data set to separate retrieval recall, citation support, grounded answers, and abstention instead of replacing evidence with one LLM-judge score.
Build provenance around DeepSeek-OCR 2 from immutable source and page rendering through pinned model execution, structural parsing, and human review.