Sem Manuscript Audit

5. Audit design and timing.

• Check unit of analysis, aggregation, pooling of heterogeneous units, censoring, time-window consistency, and outlier rules.
• Flag cross-sectional claims that are dressed up as longitudinal inference.

6. Audit proxies and data sources.

• Check whether QCEW, SVI, matching crosswalks, or risk indices are adequately defined and defended.

7. Audit interpretation.

• Downgrade deterministic or causal language when the design is associative.
• Flag misleading labels such as partial dependence plots when the workflow is not actually PDP-based.
• Flag metacommentary that makes the paper read like a project memo instead of a standalone article.
• Treat phrases like this revision, the revised manuscript, the current repo, the audit found, or the imported outputs as manuscript defects unless the user explicitly wants a reflexive methods note.

8. Output findings in the right format.

• If the user asked for review, present findings ordered by severity with file references.
• If the user asked for revision support, convert findings into tracker entries or a revision plan.

High-Risk Failure Modes

• Model-comparison paragraph contradicts its own table.
• Fit indices are poor but described as strong without caveat.
• Moderation or invariance is claimed but not estimated.
• Timing outcomes ignore right-censoring.
• State and local units are pooled without adequate justification.
• Latent constructs rely on proxies that are not defended.
• Data-availability statement hides which components are actually public.
• References overstate preprints or gray literature.

References

• Read references/audit-checklist.md for the detailed issue list and evidence prompts.