Science Superpowers provides a composable computational-science methodology for AI research agents, designed to formalize the scientific process. It transforms fuzzy interests into precise, falsifiable questions and ensures rigorous, reproducible analysis via pre-registration and systematic investigation. This methodology benefits AI researchers by protecting against common statistical fallacies like p-hacking and HARKing, leading to more trustworthy scientific outcomes.

How It Works

It uses composable skills that auto-trigger via a session-start bootstrap within an agent harness. Reimplementing the Superpowers methodology for science, its central discipline is pre-registration over TDD. The workflow begins by framing fuzzy interests into precise, falsifiable questions, grounding them in prior work, and designing the analysis. Crucially, it enforces the "Iron Law" of pre-registration, locking hypotheses, predictions, and decision rules before outcomes are observed. This separation, combined with execution in a reproducible workspace (pinned environments, fixed seeds, immutable data), protects against bias. Anomalies are root-caused, claims verified, and results red-teamed before reporting.

Quick Start & Requirements

Installation is harness-specific, requiring configuration within agents (e.g., Cursor, Claude, Gemini CLI, Pi) to load Science Superpowers as a plugin/extension via hooks or manifests. It has zero third-party dependencies, running with the agent harness and a POSIX shell.

Highlighted Details

• Pre-registration ("Iron Law"): Locks hypotheses, predictions, and decision rules before outcomes, strictly separating confirmatory/exploratory analysis.
• Root-Cause Investigation: Implements a systematic process for investigating anomalies, prioritizing understanding over data exclusion.
• Red-Teaming: Integrates adversarial review to challenge analysis and conclusions before reporting.
• Composable Skills Library: Offers automated methodological skills covering framing, planning, execution, investigation, review, and reporting.

Maintenance & Community

Users can follow K-Dense on X, LinkedIn, and YouTube for updates and demos. Specific community channels are not mentioned.

Licensing & Compatibility

Released under the permissive MIT License, allowing broad compatibility with commercial use.

Limitations & Caveats

Setup and integration are dependent on the specific agent harness, requiring per-environment configuration. Emphasizes pre-registration as a core discipline, representing a specific philosophical approach.