grounded-review

What it does

The skill performs the quality-gating and repair stage of the middle layer. It reads the grounded note, the literature result, the report draft, and optional supporting research artifacts, then outputs a review diagnosis and score report, a review state file, and a final deliverable research report. It acts as a gate between grounded-summary and the final rendering stage, not as a broad search or rendering tool.

How it works

• It uses the inputs grounded.md, lit.md, summary.md (and optional artifacts) as sources of truth for evidence, scope, uncertainty, and technical detail recovery.
• It produces: a review diagnosis and score report, a review state file, and a final deliverable research report.
• The process follows a round-based archival structure: data/review_outputs/<ground_id>/round_<N>/ contains review_report.md and review_state.json for each round, with a final research_report.md in data/reports/<ground_id>/ after completion.
• It preserves the core structure of the input draft, applies structured evaluation and bounded repair, performs re-checks, and finalizes the deliverable for downstream formatting/rendering steps.

When to use it

Use this skill when grounded.md, lit.md, and summary.md exist and you want a final deliverable research report for the current grounded item. Do not use when the report draft has not been written, the literature result is missing, or only a quick synthesis is required.

What it can touch

• The skill reads: grounded.md, lit.md, summary.md, and optional supporting artifacts. It does not specify external editing beyond the bounded repair loop and structured evaluation.

Caveats

• It is described as a quality-gate with bounded repair and should not perform unrestricted new literature search or radical direction changes.
• The process uses a round-based archival system and finalizes the deliverable in data/reports/<ground_id>/research_report.md.