JSON output schema

Both the rubric and synthesis prompts instruct the model to return JSON matching the shapes below. The parser is tolerant of fenced JSON (```json ... ```) and surrounding whitespace; prose-wrapped JSON is not repaired and falls through as unstructured.

Review object (per reviewer)

{
  "summary": "string, 1-3 sentences",
  "findings": [
    {
      "id": "string, stable slug within this review",
      "severity": "blocker | major | minor | nit",
      "category": "correctness | design | security | performance | clarity | scope | other",
      "location": "string, optional",
      "title": "string, one line",
      "body": "string, markdown allowed",
      "suggestion": "string, optional"
    }
  ],
  "open_questions": ["string"],
  "verdict": "ship | iterate | reject"
}

Rules:

summary, findings, verdict are required. open_questions may be an empty array but must be present.
findings may be empty. Reviewers should not invent findings to fill space.
severity is one of the four enums; no custom values.
id must be unique within a single review.

Synthesis object

{
  "summary": "string, 2-4 sentences",
  "findings": [
    {
      "title": "string",
      "body": "string, markdown allowed",
      "severity": "blocker | major | minor | nit",
      "tag": "consensus | contested | singleton",
      "sources": ["reviewer_id", "..."],
      "disagreement": "string, required when tag = contested"
    }
  ],
  "verdict": "ship | iterate | reject",
  "verdict_rationale": "string"
}

Tagging rules are enforced by the synthesis prompt, not by code:

consensus — raised by 2+ reviewers with substantively the same point.
contested — raised by one and disputed by another; disagreement must summarize the split.
singleton — raised by exactly one reviewer, not contradicted; kept.

Parser tolerance

Order of attempts when reading reviewer stdout:

Strip leading/trailing whitespace and ```json / ``` fences.
json_decode strict.
On failure, the output is stored as unstructured (raw text) and still shown to the human at checkpoint 2.

The parser does NOT attempt to repair JSON, ask a model to fix it, or heuristically extract fields. Fix the prompt instead.