What problem does it solve? Developing rigorous machine-learning theory results requires coordinating literature surveys, idea generation, formalization, and multi-level proof verification, which is error-prone and hard to audit when done ad hoc. This Skill provides a controller that manages the entire research workflow with checkpoints, retry budgets, and auditable artifacts. ## Core Features & Use Cases - End-to-End Workflow Orchestration: Coordinates Workflow 1 (literature survey, perspective selection, idea generation, formalization) and Workflow 2 (proof sketch, global proof, step proofs, assembly, and specialized final reviews). - Subagent Management: Spawns and tracks distinct run-scoped, perspective-scoped, and step-scoped subagents with worker provenance logging and a 12 + 2k distinctness gate per perspective. - Failure-Localized Retry Routing: Routes review failures to the smallest repairable stage under strict MAX_* attempt budgets, archiving consumed artifacts into proof_history before overwrites. - Use Case: Given a research question on generalization bounds, run the controller to survey literature, branch three perspectives, formalize theorem contracts, develop graph-structured proofs, and copy accepted results into results/perspective_M/idea_N/theory. ## Quick Start Invoke the ml-theory-research-agent skill with your research question as the argument, optionally adding a RESEARCH_BRIEF.md, and approve checkpoints as the workflow produces each artifact.
