data-audit

name	data-audit
description	Audit supplied research data for identity, provenance, schema, missingness, duplicates, impossible values, temporal and unit consistency, leakage, and transformation traceability. Use before analysis to determine data readiness without claiming to detect fabrication.

Data Audit

Audit what can be observed and traced. This Skill is a data-quality and integrity screen, not a fraud detector.

Required boundary

Use ../../templates/data-audit.md. Inventory every file or object actually supplied, its intended role, and the data documentation available. Do not open unrelated paths or run user data through an external service.

Only report a check as executed when a tool actually completed it. Record the method, scope, and output anchor. A tool error, parse failure, unsupported format, missing required file, or incomplete transformation chain fails closed as BLOCKED or UNKNOWN; it must never produce a ready or green result.

Audit layers

1. Identity — file role, size, digest when computed, schema, version, and duplicate-file relationships.
2. Provenance — source, access basis, acquisition date, immutable raw input, transformations, and output lineage.
3. Structure — columns, types, labels, units, encodings, keys, and referential integrity.
4. Quality — missingness, duplicates, impossible values, range and unit inconsistencies, temporal order, and internal arithmetic.
5. Analysis construction — eligibility, exposure, outcome, covariates, weights, time zero, exclusions, leakage, and derived-variable trace.
6. Reproducibility — transformation code or instructions, deterministic settings, environment, and output identity.

Treat missing documentation as missing evidence. Treat a directly observed contradiction as negative evidence. Treat unusual distributions as signals for domain review, never proof of fabrication.

Decision

• DATA_READY: required identity, provenance, construction, and quality gates pass for the stated analysis;
• REVIEW_REQUIRED: no blocker is demonstrated, but material checks require domain or human review;
• BLOCKED: identity, provenance, parsing, transformation trace, or a consequential quality issue prevents responsible analysis.

Do not auto-modify, impute, delete, rename, or overwrite user data. Recommend a repair plan instead.

Call record_checkpoint with the strongest provenance boundary, unresolved uncertainty, or stop rule. Prefix the statement with [data-audit].

Output

Write data-audit.md and include:

1. audit boundary and executed checks;
2. dataset identity and provenance;
3. prioritized findings with evidence anchors;
4. analysis-readiness gates;
5. unexecuted checks and limitations;
6. DATA_READY, REVIEW_REQUIRED, or BLOCKED.