Breaking

K-Dense-AI Releases 'Science Skills' Toolkit for AI Research Agents

July 3, 2026 at 10:33 EDT

AI Agents
Open Source
Research & Papers

K-Dense-AI has released science-superpowers, an open-source set of skills designed to embed scientific discipline into AI research agents. Centered on pre-registration, it aims to enforce reproducibility and rigor across data analysis and computational science workflows.

Open Source · K-Dense-AI

Science Superpowers: hard-coding the scientific method into AI agents

An MIT-licensed set of ~15 composable skills that builds pre-registration and reproducibility into AI research agents — foreclosing p-hacking, HARKing, and optional stopping through the agent's own working procedures.

~15

core skills, auto-triggered at session start

agent harnesses supported (Cursor, Claude Code, Codex…)

dependencies — only a harness & POSIX shell

The pre-registration workflow — one flow from question to archive

Frame question & survey prior work

→

Pre-register analysis
(the "Iron Law")

→

Execute, verify & red-team review

→

Report & archive

Hypotheses, predictions, and decision rules are committed before results are seen — foreclosing after-the-fact manipulation.

Related benchmark: K-Dense Analyst on BixBench

Bioinformatics accuracy — a +6.3 point lead, shown to scale.

22.9%

GPT-5

29.2%

K-Dense Analyst

Received as progress

Forcibly builds scientific method into agent procedures
Seen as a step toward better reproducibility
Open, free, zero-dependency, harness-agnostic

Open concerns

Context consumption from automatic skill triggering
Dependence on specific harnesses
Reduced flexibility in complex research; real-world evaluation still limited

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.