What problem does it solve? Research experiments lose critical context when failures are discarded without recording why they happened, what was tried, and what to change next. This Skill preserves the full reasoning trail of an LLM research project—failed runs, rejected ideas, predictions versus actual results—so nothing is lost to memory. ## Core Features & Use Cases - Living Lab Notebook: Maintains a single chronological LAB_NOTEBOOK.md with phase-tagged entries, decision journals (predicted outcome, confidence, assumptions), and an auto-generated tag index for reverse lookup. - Reproducible Failure Postmortems: Auto-drafts a POSTMORTEM.md per failed run with a 3-5 hypothesis space derived from events.jsonl and error.txt, plus a reproduce.sh script and uv.lock snapshot satisfying a 7-tuple reproducibility checklist. - Rejected Ideas & Lessons DB: Archives rejected ideas with reasons and revisit conditions in 03_REJECTED_IDEAS.md, and appends generalizable lessons to a cross-project ~/.research-lessons.json database. - Use Case: During Phase 6 of an auto-research workflow, a training run fails with CUDA OOM; the Skill auto-generates a postmortem with hypothesis drafts (memory, gradient checkpointing, config drift), a blameless-framed decision section, and a frozen reproduce script. ## Quick Start Ask the agent to use the research.lab.notebook skill to record a failed experiment run and draft its postmortem for your current research project.
