What problem does it solve? Autonomous research projects often lose coherence across long sessions: state gets duplicated, negative results derail the pipeline, and transitions between literature review, experimentation, and promotion happen without evidence gates. This Skill enforces a single canonical state machine that routes one active candidate through literature, adaptation, screening, confirmation, and promotion with auditable transitions. ## Core Features & Use Cases - Canonical State Management: Uses research/STATE.json and transition_guard.py to validate and append every pipeline transition to an append-only PIPELINE.jsonl log. - Stage-Gated Pipeline: Enforces evidence requirements at each phase (baseline, literature, mapping, adaptation, screening, confirmation, promotion, frozen) before allowing progression. - Failure Routing: Converts screening failures and unconfirmed results into structured next steps—either an unused mapped candidate or a targeted literature refresh—instead of stalling. - Use Case: After an entrance skill initializes a research goal to improve a model's primary metric, this loop deep-reads relevant literature, maps ranked candidate mechanisms, runs precommitted screening and confirmation experiments, audits the evidence, and hands off a promoted method or triggers paper writing. ## Quick Start Resume the research pipeline by running the transition guard to show current canonical state, then dispatch the skill matching the active phase.
