RECAST

Replication and Extension with Causal AI Statistical Toolkit

RECAST a paper: Replicate, Extend, and Enrich Key Findings with Causal AI

RECAST (v.) — to take an econometrics paper and its replication data, reproduce the original results, extend them with Double/Debiased Machine Learning and Causal Forests, put the findings through a structured AI peer review, and publish a report — autonomously, end to end.

“This paper has been RECASTed.”

Browse RECASTed papers RECAST your own paper

What RECAST stands for

Replication and Extension with Causal AI Statistical Toolkit

RECAST is an autonomous multi-agent pipeline built on Claude Code. It takes a published econometrics paper and its replication data and produces a RECAST — a complete package of replicated results, a DoubleML extension, and a referee-reviewed report — without manual intervention.

How RECASTing works

1 · Read

Claude reads the paper PDF and extracts the identification strategy, key specifications, and data structure into a structured manifest.

2 · Replicate

The original OLS/IV regressions are reproduced and cross-checked coefficient by coefficient against the published tables.

3 · Extend

Following Baiardi & Naghi (2024), DoubleML compares 7 ML methods (Lasso, Trees, Boosting, Forest, Neural Net, Ensemble, Best) with median aggregation across 20+ sample splits. Causal Forest estimates individual-level treatment effects. Both include BLP heterogeneity testing, GATE quintile analysis, and CLAN classification of who benefits most.

4 · Review

Three isolated AI referees — identification, DML methods, robustness — each file an independent report following Berk, Harvey & Hirshleifer (2017) principles: contribution assessment first, essential vs. suggested separation, scientific justification required, implicit bargain enforced. One round is the target.

RECASTed papers

RECASTed papers appear here automatically once published with /publish. Run /recast (DML) or /recast-cf (Causal Forest) to RECAST a paper.

View all RECASTed papers →

Why RECAST?

7 ML methods, compared Following Baiardi & Naghi (2024), RECAST tests every result with Lasso, Trees, Boosting, Random Forest, Neural Networks, Ensemble, and Best (by nuisance MSE). Median aggregation across 20+ sample splits with adjusted standard errors.

Principled peer review Three isolated AI referees follow Berk, Harvey & Hirshleifer (2017): contribution first, essential vs. suggested separation, scientific justification required. Synthesis quality-controls the referees and enforces the implicit R&R bargain.

Heterogeneity, quantified BLP pre-test for heterogeneity, 5-quintile GATE analysis with jointly valid CIs, CLAN classification of who benefits most. Causal forests add individual-level CATEs, feature importances, and calibration diagnostics.

Full audit trail Every revision round is logged in append-only review_history/. Nothing is overwritten. The RECAST is fully reproducible and inspectable.

One command /recast <project> extends with DoubleML. /recast-cf <project> extends with Causal Forest. Both handle everything — analysis, gate, review, final report. /publish <project> adds the paper to this site.