judge

Evaluate agent task outputs against a three-dimension rubric and produce structured verdict records. The judge operates as a quality gate at the task completion boundary, scoring outputs on Semantic accuracy, Pragmatic usefulness, and Syntactic consistency.

The rubric reuses three dimensions from the KLS (Krogstie-Lindland-Sindre) quality framework defined in disciplined-quality-evaluation:

| Semantic | Does it accurately represent the domain? | Factual correctness, domain terminology, no contradictions | | Pragmatic | Does it enable the intended decisions/actions? | Actionable, useful, addresses the task goal | | Syntactic | Is it internally consistent and well-structured? | Format compliance, structural completeness, no broken references |

使用源自 KLS 质量框架的三维评分标准（语义、实用、句法）评估代理任务输出。在以下情况下使用：(1) 任务已完成并且需要在验收前进行质量评估，(2) 需要自动任务后质量检查，(3) 代理输出需要多模型共识判决，(4) 必须使用结构化 JSON 判决对文档、代码或规范质量进行评分，或者 (5) 在模型不一致后需要人工决策。生成与 Automation/judge/ 中的判决模式兼容的 JSONL 判决记录。 来源：terraphim/terraphim-skills。