ARIS (Auto-claude-code-research-in-sleep)

Why this matters

Large-scale ML projects waste time on repetitive loops: read papers, generate ideas, run pilots, get reviews, fix drafts, rerun experiments. ARIS reframes that cycle by shipping a portable, skill-based methodology (not a heavy platform) so LLM agents can autonomously carry out many of those steps overnight and produce reproducible artifacts (plans, experiments, audits, LaTeX + PDF).

What Sets It Apart

• Markdown-native skills: every workflow is composed from plain SKILL.md files, so behavior is transparent and editable. No heavy framework, database, or daemon required; the same skills run in Claude Code, Codex CLI, Cursor, Trae, or adapted local model environments.
• Cross-model adversarial review: ARIS orchestrates an executor (Claude Code or other) and a separate reviewer (Codex / GPT-5.4 or alternatives) to avoid self-play blind spots—practical for stress-testing claims and code before costly experiments.
• End-to-end research pipelines: one-command flows for idea discovery, experiment bridge (implement+deploy), autonomous review loops (sleep & wake to results), paper writing (claims→figures→LaTeX→compile) and rebuttal drafting, with optional persistent Research Wiki and meta-optimize to improve the harness itself.
• Practical assurance & audit gates: optional —effort: max|beast triggers mandatory audits (proof-checker, claim-audit, citation-audit) and a verifier gate so submission-ready output has machine-checkable evidence trails.

Key capabilities (so what?)

• Automates hypothesis→pilot→analysis loops: useful when iterative pilots and quick ablations are needed; reduces manual orchestration overhead so teams can validate many variants rapidly.
• Portable reviewer backends: supports Codex MCP, Oracle/GPT-5.4 Pro, Gemini, local models (GLM, MiniMax, Kimi), and falls back gracefully—so you can run it with or without specific cloud APIs.
• Lightweight install & update model: project-local symlink installer and a smart updater let teams adopt ARIS skills per-project without global collisions.

Who it's for and tradeoffs

Great fit if you are an ML researcher or small team who wants to accelerate iteration cycles: you already write experiments and papers locally, can provide a GPU or use on-demand rentals (Vast.ai/ModelScope), and are comfortable granting an agent limited permissions to run and monitor experiments. ARIS is optimized for research workflows (idea generation, ablations, writing), not for production ML deployment pipelines.

Look elsewhere if you need a managed cloud service, enterprise-grade governance out of the box, or a GUI-first product: ARIS favors transparency (Markdown skills) and automation scripts over centralized dashboards and enterprise policy controls. Also, some reviewer quality and behavior depend on the chosen LLM backends—expect to tune reviewer routing and prompts for unusual domains.

Quick decision points

• Turn on research-wiki to accumulate paper/idea/experiment memory across runs.
• Use —reviewer: oracle-pro or Codex MCP for the strictest audits if available; otherwise ARIS works with many alternative model combos.
• Use the assurance gate (—assurance: submission / —effort: beast) when preparing a real conference submission to enforce proof and citation audits.