What problem does it solve? The autonomous calibration agent stages its auto-learned proposals to a pending queue instead of writing them directly to the curated knowledge base, and without a human review step those proposals are silently dropped and the knowledge base stops learning from runs. ## Core Features & Use Cases - Queue Inspection: List open or all staged proposals in auto_discovered_pending.json via the review_pending_knowledge.py tool. - Evidence-Based Vetting: Cross-check each proposal's mechanism, experiment results, and do_not_repeat rules against run data, RAG sources, and existing curated entries before deciding. - Promote or Discard: Promote vetted entries into discoveries.json, failed_approaches.json, or experiments.json with optional confidence overrides, and discard phantoms, duplicates, or overbroad rules. - Use Case: After several autonomous calibration rounds, run the curation pass to review what the agent proposed to learn, promote the verified discoveries, and discard misunderstandings so the curated knowledge base stays accurate. ## Quick Start Ask the agent to review the pending knowledge proposals and promote the ones verified against experiment evidence.
