What problem does it solve? Running LLM benchmark evaluations through Python scripts requires boilerplate code for model configuration, feature toggles, and result handling. This Skill lets you execute the full karenina verification pipeline directly from the command line, from a benchmark checkpoint file to exported results. ## Core Features & Use Cases - Preset-Based Verification: Bundle answering and parsing model configurations into reusable JSON presets and run karenina verify against any checkpoint file. - Progressive Save and Resume: Checkpoint long-running jobs after each completed task and resume interrupted runs at the triple level without re-running completed work. - Failure Repair and Model Comparison: Re-run only failed rows with karenina repair, and compare models by overriding preset flags across providers like Anthropic and OpenAI. - Use Case: You have a benchmark checkpoint with 500 questions and want to evaluate Claude Sonnet against GPT-4.1 with rubric scoring. Create one preset, run two karenina verify commands with different --answering-model flags, and export both result sets to JSON for comparison. ## Quick Start Ask the AI to run karenina verification on your checkpoint file using a preset with an Anthropic answering model and export the results to a JSON file.
