What problem does it solve? Single-turn benchmarks cannot measure how a model behaves across a live conversation, such as whether it abandons correct answers when challenged or retains knowledge across turns. This Skill guides you through building graph-based multi-turn scenario evaluations in karenina, where each turn's result determines the next question. ## Core Features & Use Cases - Scenario Graph Construction: Define nodes (questions with answer templates), conditional and unconditional edges, entry points, and END transitions, with validation for orphan nodes and missing fallbacks. - Outcome Criteria: Assert properties of the full execution using sugar functions like last_turn, any_turn, all_of, cross_turn, and TurnCheck with scope selectors. - Execution and Analysis: Run scenarios with answering and parsing models, then inspect per-turn history, paths taken, outcome verdicts, and terminal failures via VerificationResultSet. - Use Case: Build a sycophancy check where a model answers a biomedical question, is then challenged with a false claim, and outcome criteria verify whether it maintained the correct answer across turns. ## Quick Start Ask the AI to build a karenina scenario evaluation that branches the conversation based on whether the model's first answer passes verification.
