What problem does it solve? Setting up karenina's verification pipeline requires coordinating many configuration surfaces—answering and parsing models, adapter interfaces, guard stages, replicates, presets, deep judgment, and MCP integration—and misconfiguration leads to cryptic auto-fails or silently skipped rubric evaluation. This Skill organizes the entire VerificationConfig and ModelConfig surface by task so you can configure, run, and debug verification runs correctly. ## Core Features & Use Cases - Model and Adapter Configuration: Define answering and parsing ModelConfigs and choose the right interface (langchain, claude_tool, claude_agent_sdk, openai_endpoint, manual, and more) with a decision table covering MCP and tool support. - Guards, Replicates, and Presets: Understand the four auto-fail guard stages, set replicate counts for variance measurement, and save or load reusable JSON presets with overrides. - Advanced Pipeline Features: Configure prompt assembly, deep judgment (template and rubric), MCP tool-use evaluation, run extension via extend_template/extend_rubric, and progressive save-and-resume sinks for long runs. - Use Case: You want to evaluate a tool-using biomedical agent with 3 replicates, an abstention guard, and deep judgment enabled, then resume the run after an interruption. This Skill gives you the exact config fields, code snippets, and debugging fields (result.metadata.failure) to do it. ## Quick Start Ask the assistant to create a VerificationConfig with a Claude answering model, a Haiku judge model, three replicates, and the abstention guard enabled, then run it with a progressive file sink.
