Framework Documentation

RECAST is built as a set of Claude Code slash commands and internal skills. All orchestration is done by AI agents communicating through the filesystem — no custom Python runtime is required to operate the framework.

Methodological foundations

The RECAST pipeline is grounded in two key academic references:

  • DML methodology: Baiardi & Naghi (2024), “The value added of machine learning to causal inference: evidence from revisited studies,” Econometrics Journal. The pipeline’s DML extension — 7 ML methods, adaptive K-fold cross-fitting, median aggregation across sample splits, B&N adjusted standard errors, BLP/GATE/CLAN heterogeneity analysis — is directly based on their approach and replication code.

  • Referee process: Berk, Harvey & Hirshleifer (2017), “How to Write an Effective Referee Report and Improve the Scientific Review Process,” Journal of Economic Perspectives. The review loop follows their principles: contribution assessment before criticism, essential vs. suggested separation with scientific justification required, cost-benefit weighting, the implicit revise-and-resubmit bargain, and the synthesis referee as quality control against signal-jamming.

Pipeline Architecture

Visual diagram of the full pipeline — from paper ingestion to published report, including both DML and Causal Forest paths.

Quick Start

Install, scaffold a project, add your files, and launch the pipeline in four steps.

Pipeline Stages

Reference for each of the six analysis notebooks, the Advisor Gate, the Review Loop, and the Final Referee.

Slash Commands

Usage and rules for /recast, /recast-cf, /init, /stage, /review, /final, and /publish.

Skills Reference

Internal sub-agent skills: orchestrator, notebook runner, gate validators, referees, synthesis, and revision agent.

Project Structure

The folder layout created by /init, immutability rules, and what to read when the pipeline finishes.