Research Harness Review

Perform an independent, read-only review of research outputs before release.

This skill reads existing scripts, logs, and outputs. It does not run code. It does not edit scripts or data. It produces a structured review report with a verdict.

This skill runs after /research-harness-work and before /research-harness-release.

Quick Reference

Pre-flight Checks

1. Read analysis_plan.md. If it does not exist, stop.
2. Confirm at least one task has status cc:done. If none, report that no completed tasks exist to review.
3. List all cc:done tasks. These are the review scope.
4. Read study_spec.md, reports/data_audit_report.md, reports/data_cleaning_report.md, and reports/merge_report.md (if it exists).

Procedure

This skill is read-only throughout. No Bash commands, no script execution, no file writes except the review report.

Step 1 — Identification credibility review

Read study_spec.md §2 (identification strategy).

For each main model task in analysis_plan.md, check:

• Does the estimator used in the script match the identification strategy in the study spec?
• Is the key identification assumption stated in the script header or task description?
• Is any falsification or placebo test present in the plan?
• For DiD: is there a pre-trend check task?
• For IV: is the instrument defined and the exclusion restriction described?
• For RD: is bandwidth selection documented?

Assign one of: strong / moderate / weak / insufficient

insufficient identification immediately produces REQUEST_CHANGES. Do not write findings as minor if the identification is insufficient for the causal claim being made.

Step 2 — Model specification alignment

For each cc:done analysis task:

• Read the script
• Compare the estimator, outcome variable, covariates, sample restrictions, and fixed effects to study_spec.md §4 and §5
• Flag any deviation as: minor (changes that do not affect the main result), major (changes that affect the result), or critical (changes that contradict the approved study design)

Step 3 — Numerical accuracy check

For each cc:done task with an output file:

• Read the log file
• Read the output file (table or figure metadata)
• For each key number that would appear in a paper (main coefficient, standard error, sample size, p-value, mean): confirm it appears in the log
• If a number appears in the output file but not in the log: flag as unverified

An unverified number is a critical finding if it will appear in the final reported results.

Do not verify numbers by re-running scripts. Only read existing logs and outputs.

Step 4 — Sample construction check

• Read the cleaning report row counts
• Read the log files for each analysis task
• Verify that the N reported in analysis logs matches the expected N given the cleaning report and sample restrictions
• Flag any discrepancy as: minor (small difference with a plausible explanation), major (large unexplained difference), or critical (N is clearly wrong)

Step 5 — Causal claim assessment

Read any interim outputs, table notes, or text summaries (if present in output/ or reports/).

For every causal claim found:

• Record the claim text and its location
• Check that it carries an identification strength tag ([descriptive], [correlational], [quasi-experimental: ...], [experimental])
• Assess whether the tag is appropriate given the design
• Flag claims that overstate causal strength as major findings

Do not rewrite weak evidence as strong causal evidence. If the evidence is correlational, the claim must be correlational.

Step 6 — Data cleaning completeness check

• Confirm reports/data_cleaning_report.md exists and its verification section is PASS
• Confirm reports/merge_report.md exists (if merges were performed) and all entries have pre/post row counts
• Confirm data/raw/ was not modified (if git status data/raw/ or equivalent is available from prior logs, read it)
• Flag any incompleteness as a finding

Step 7 — Hallucination and fabrication check

• Confirm no result is claimed without a corresponding log file
• Confirm no script produces a hardcoded numerical result (i.e., check that output values are computed, not assigned as literals)
• If any citations appear in output text: note that citation accuracy cannot be verified by this skill and flag for manual check

Step 8 — Write review_report.md

Copy templates/review_report.md to reports/review_report.md and fill in all sections from Steps 1–7.

Verdict rules:

BLOCK is a stronger form of REQUEST_CHANGES. It means the research cannot be released in any form until the finding is resolved.

Step 9 — Report verdict

Print a review summary:

Research Harness Review — Complete

Tasks reviewed: N
Scope: [task IDs]

Identification credibility: strong / moderate / weak / insufficient
Numerical accuracy: all verified / N unverified
Causal claims: all appropriate / N overstated

Critical findings: N
Major findings: N
Minor findings: N

Verdict: APPROVE / REQUEST_CHANGES / BLOCK

Review report: reports/review_report.md

If verdict is REQUEST_CHANGES or BLOCK:

Required actions before /research-harness-release:
1. [Finding 1 — required action]
2. [Finding 2 — required action]

Return to /research-harness-work to re-execute affected tasks, then re-run /research-harness-review.

Forbidden Actions

• Do not run any script or Bash command
• Do not edit any script, data file, log, or output file
• Do not approve a result that cannot be traced to a log file
• Do not rewrite weak causal evidence as strong causal evidence in the review report
• Do not produce APPROVE when any critical finding exists

Evidence Requirements

• reports/review_report.md exists with all sections populated
• Verdict is one of: APPROVE, REQUEST_CHANGES, BLOCK
• All findings are documented with location and required action

Completion Criteria

• All cc:done tasks reviewed
• Identification credibility assessed
• Model specification checked against study spec
• Numerical accuracy checked (each key number traced to a log)
• Sample N checked
• Causal claims assessed with identification tags
• Data cleaning completeness confirmed
• reports/review_report.md written
• Verdict printed

Handoff to Stage 7

If verdict is APPROVE:

Review passed. Run /research-harness-release to package the replication archive.

If verdict is REQUEST_CHANGES or BLOCK:

Return to /research-harness-work to resolve the findings listed in reports/review_report.md. After re-executing affected tasks, run /research-harness-review again. Do not run /research-harness-release until the verdict is APPROVE.