HN Showcase
Back to browse
GitHub Repository

Autonomous ML research loops for Claude Code with mechanical anti-fabrication guards.

0 starsPython

Novum – Automated ML Research Pipeline with Anti-Fabrication Guards

by euanai·Mar 4, 2026·1 point·3 comments
Visit ProjectView on HN

AI Analysis

●●●BangerBig BrainWizardryZero to One

Anti-fabrication constraints on a 30-hour autonomous research run—addresses real MLR-Bench hallucination data.

Strengths
  • •Mechanical enforcement of result validity (not just prompts) tackles quantified 80% fabrication rate problem.
  • •30-hour autonomous case study with iteration tracking, regression detection, and hypothesis filtering—production evidence.
  • •Integrates with Claude Code's agent loop, enabling true iterative research cycles with real constraints.
Weaknesses
  • •Requires max Claude Code plan ($20k/month) and NVIDIA 8GB+ GPU—prohibitive for most researchers.
  • •Author explicitly declines to verify paper draft, undermining core claim of trustworthy autonomous research output.
Category
AI/ML
Target Audience

ML researchers, academic labs, teams running autonomous experiments

Similar To

Auto-sklearn · AutoGluon · MLflow tracking

Similar Projects

Developer Tools●●●Banger

Claude-Autopilot: autonomous dev pipeline with risk-tiered review

Risk-tiered Codex review gates autonomous merges better than GitHub Copilot.

Ship ItBig Brain
axledbetter01
4215d ago
Developer Tools●●Solid

Claude Extender – Autonomous Agent Management for Claude Code

Using plain markdown + YAML as the canonical agent format is a smart, low-friction choice — edit agents in your editor, commit them, and the daemon runs scheduled, watcher, or persistent sessions. It persits run logs, memory and costs as browsable markdown and can start MCP tool servers, which makes it immediately useful if you already run Claude Code; the flip side is the tight coupling to Anthropic/MCP limits broader appeal.

Niche GemShip It
wbnns
213mo ago