What problem does it solve? Deciding the next step after an LLM training experiment requires reading scattered logs, metrics, and state files, which is slow and error-prone. This Skill reads AutoResearch run evidence and loop state, then produces a structured, evidence-backed triage verdict so each experiment loop iteration stays disciplined and auditable. ## Core Features & Use Cases - Evidence-Backed Triage: Reads LOOP.md, loop-constraints.md, loop-budget.md, STATE.md, and the latest run manifest, logs, metrics, and Git provenance before proposing anything. - Structured Verdict Output: Returns fixed sections covering High Priority, Evidence, Candidate Hypothesis, Risk/Budget/Authority, a KEEP/REJECT/ESCALATE_HUMAN verdict, and exact STATE.md update proposals. - Safety Guardrails: Operates in report-only L1 mode, separates model/data/system/hardware failure layers, stops after two identical failure signatures, and never proposes architectural overhauls. - Use Case: After a verl training smoke run fails on a remote NPU server, invoke this Skill to get a single falsifiable hypothesis, the smallest validation experiment, and a clear escalation recommendation with cost estimates. ## Quick Start Ask the assistant to triage the latest AutoResearch run and propose the next bounded experiment based on current loop state and evidence.
