What problem does it solve? After running a research experiment, deciding how the results affect your claims and ideas is subjective and error-prone. This Skill turns completed experiment results into structured wiki knowledge updates through an independent cross-model verdict process. ## Core Features & Use Cases - Cross-Model Verdict: A Review LLM independently judges whether results support, partially support, refute, or leave a claim inconclusive, then Claude synthesizes both verdicts conservatively. - Automatic Wiki Updates: Updates claim confidence and evidence, idea status (validated/failed), experiment pages, and knowledge graph edges (supports/invalidates) via the research_wiki.py tool. - Four Verdict Paths: Supported raises confidence and validates ideas; partially supported suggests supplementary experiments; not supported records a specific failure_reason as anti-repetition memory; inconclusive triggers debugging guidance. - Use Case: After completing an experiment testing whether a new attention variant improves benchmark scores, run the evaluation to get an impartial verdict, update the claim's confidence from 0.4 to 0.65, mark the linked idea as validated, and add a supports edge to the research graph. ## Quick Start Evaluate the completed experiment with slug attention-variant-v1 and update my research wiki claims, ideas, and graph based on the verdict.
