What problem does it solve? Deciding what to try next in an ongoing research experiment requires synthesizing what worked, what failed, and what the literature says — a slow, error-prone manual process. This Skill automates that ideation pass on a Chronicle experiment while keeping the human as the approval gate: nothing is committed or queued without explicit acceptance. ## Core Features & Use Cases - Grounded evaluation: Loads the experiment record (lineage, variations, runs, lessons) and invokes the evaluate-results skill for a what-worked / what-didn't read judged against pre-registered hypotheses. - Literature grounding: Optionally invokes literature-review to survey external work scoped to the gap the evaluation exposed, registering citations on the anchor experiment. - Pre-registered proposals: Drafts next variations (or child experiments) each with a required hypothesis and expected outcome, then queues only the ones the user accepts via chronicle.propose_variation or the variation-authoring skills. - Use Case: After a round of training runs on an experiment plateaus, ask for a synthesis pass — the agent evaluates the variations, checks the failure-mode catalog, proposes three regularization-focused variations with hypotheses, and queues the two you approve. ## Quick Start Ask the agent to run a synthesis pass on your current Chronicle experiment and propose what to try next.
