What problem does it solve? Running machine learning experiments requires coordinating code changes, compute backends, run IDs, and result collection, and doing this manually leads to lost runs, duplicated IDs, and untraceable results. ## Core Features & Use Cases - Backend Detection and Sanity Checks: Detects the configured local CUDA, MPS, CPU, SSH, Vast.ai, or Modal backend and runs a bounded sanity command before any full training launch. - Attributable Run Management: Reserves stable run IDs from canonical state, launches training as DSH background jobs, and records both job ID and run ID for full traceability. - Result Collection and Routing: Collects terminal job results, recomputes declared metrics from raw outputs, appends factual records to research/RESULTS.jsonl, and hands evidence to experiment-run for routing. - Use Case: During a screening phase of an autonomous research project, use this Skill to implement the selected candidate, launch a seeded training run on the configured GPU backend, and route the completed comparison according to the precommitted decision rule. ## Quick Start Use the experiment-bridge skill to implement the current candidate and launch the planned screening run as a background job on the configured backend.
