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.
Place Sonnet 4.6 computer use inside an isolated environment with domain and action allowlists, action-time approval, prompt-injection resistance, and replayable audit.
Separate Anthropic's published Opus 4.6 capabilities from local evidence, then evaluate retrieval, editing, verification, effort, and total task cost.
Evaluate Vibe 2.0 through subagent contracts, clarification, skills, permissions, and repeatable repository tasks instead of counting product features.