critic

What it does

Scores a finished Data2Story article against five rubric dimensions (visual_design, narrative_pacing, data_method_transparency, claim_data_alignment, insight_value), using on-page evidence and traceability from verifier.json, and emits critic.json containing pass/fail plus targeted send-back instructions. It verifies each load-bearing claim/asset against its traceability chain and applies caveat-survival, honest-accuracy, and third-party-attribution caps. It does NOT rewrite content and runs at Stage 6.5 after Auditor and verify.py have produced verifier.json; re-runs each revision round. Triggered when a built index.html plus verifier.json exist, or when you determine whether the article is actually good.

How it works

• Read index.html (the finished article prose) and verifier.json, analyst.json, editor.json, designer.json, detective.json. Rubric is from references/rubric.json.
• For each dimension, read how the article handles that aspect and verify load-bearing claims by tracing data lineage in verifier.json to code lines, data_tables, or source URLs. If evidence is not resolvable, score cannot exceed 3 for transparency and claim-data alignment.
• Check for the Runnable verifier enhancement: ensure in-page run capability and reproducible notebook exist to support data_method_transparency.
• Apply caveat_survival: detect material caveats from analyst/detective/editor inputs and ensure they survive into visible prose, otherwise cap the relevant dimensions at 3.
• Score each dimension on a 1-7 scale anchored at 3, citing concrete evidence.
• Determine overall pass/fail: every dimension must be >=4, and at least one dimension >=5 (signature move). If all dims are >=4 but none reach 5, prepare a flagship-lift send-back targeting a single dimension and a concrete honest-axis move. For each failing dimension, provide a surgical send-back with role, exact section/asset to change, minimal fix, and why.
• Write critic.json with the required shape including overall and five dimensions, ensuring the 'pass' logic and tier rules are followed.

When to use it

• Triggered when a built index.html and verifier.json exist, or when judging whether the article is actually good.

What it can touch

• Uses the allowed tools Bash, Read, Write, Grep to inspect index.html and verifier.json and craft critic.json. It references the rubrics and chart references as evidence anchors but does not modify source assets.

Caveats

• It does not rewrite content and relies on stage-6.5 gating; it requires verifier.json produced previously by verify.py. It enforces the five-dimension evaluation with explicit evidence and surgical fixes, and returns critic.json with the round information and a single round assessment per cycle.