| name | research-harness-review |
| description | RES: Read-only review of research outputs. Checks identification, model spec, numerical accuracy, causal claims, reproducibility. Produces review_report.md with APPROVE/REQUEST_CHANGES/BLOCK verdict. Trigger: review research, check results, review analysis, verify outputs. Do NOT load for: cleaning, execution, release, setup, audit, planning. |
| description-en | RES: Read-only review of research outputs. Checks identification, model spec, numerical accuracy, causal claims, reproducibility. Produces review_report.md with APPROVE/REQUEST_CHANGES/BLOCK verdict. Trigger: review research, check results, review analysis, verify outputs. Do NOT load for: cleaning, execution, release, setup, audit, planning. |
| kind | workflow |
| purpose | Independent read-only review of research outputs before release |
| trigger | review research, check results, review analysis, verify outputs, /research-harness-review |
| shape | evaluate |
| role | evaluator |
| pair | research-harness-work |
| owner | research-harness-core |
| since | 2026-05-29 |
| allowed-tools | ["Read","Write","Glob","Grep"] |
| argument-hint | [--quick] [--task TASK-ID] |
| user-invocable | true |
| effort | high |
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
| Input | Action |
|---|
/research-harness-review | Full review of all cc:done tasks in analysis_plan.md |
/research-harness-review --quick | Abbreviated review: identification + numerical accuracy only |
/research-harness-review --task 2.1 | Review a single task |
Pre-flight Checks
- Read
analysis_plan.md. If it does not exist, stop.
- Confirm at least one task has status
cc:done. If none, report that no completed tasks exist to review.
- List all
cc:done tasks. These are the review scope.
- 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.
