Pipeline Audit — Adversarial Data Review

v1.0 — Adversarial reviewer that challenges data pipeline outputs, assumptions, and integrity. Complements /code-review (which reviews code quality) by testing what the code actually produced.

Review the outputs of a data cleaning or analysis pipeline by reading the data, checking for anomalies, and challenging the assumptions embedded in the pipeline. This is the "red team" for your data work.

Argument: $ARGUMENTS

• Path to a directory containing .dta output files (e.g., data/for_analysis/)
• Or a project name (will look in ~/Dropbox/Github/[project]/)
• Or a specific .dta file to audit

Modes (append to argument):

• quick (default) — Scan outputs for red flags: duplicate IDs, unexpected missingness, implausible values
• deep — Full adversarial review: all quick checks + cross-file consistency, merge integrity, distribution analysis
• pre-submission — Publication-readiness check: all deep checks + AEA-style output verification, table-to-code tracing