What problem does it solve?
Turning a written experiment plan into running GPU experiments requires writing training and evaluation code, verifying correctness, and managing deployment, which is slow and error-prone when done manually.
Core Features & Use Cases
- Plan-Driven Implementation: Parses EXPERIMENT_PLAN.md and FINAL_PROPOSAL.md to generate training, evaluation, and data-loading scripts with fixed seeds and JSON/CSV result output.
- Cross-Model Code Review: Sends experiment code to GPT-5.4 via Codex MCP for correctness review before deployment, catching logic bugs and ground-truth evaluation errors.
- Sanity-First Deployment with Auto-Debug: Runs the smallest sanity experiment first, auto-diagnoses failures (OOM, CUDA errors, NaN divergence) across up to 3 attempts, then deploys the full suite in parallel.
- Use Case: After finishing idea discovery, point the skill at refine-logs/EXPERIMENT_PLAN.md and it implements the code, deploys experiments to local, remote, vast.ai, or Modal GPUs, and produces an initial results summary ready for the auto-review loop.
Quick Start
Run /aris-experiment-bridge with the path to your EXPERIMENT_PLAN.md to implement and deploy the planned experiments on your available GPUs.
