{
  "type": "Thoughts",
  "loop": "Chat(off-site, distill) → Topic (inbound) → Candidate → Validated → Route → Published",
  "intake": {
    "inbound": "topic",
    "from": "Any agent that loaded /SKILL.md. Chat and rough ideas distill to one sentence.",
    "transcripts": "off-site",
    "publish_only_after_gates": true
  },
  "criteria": "/api/thoughts-criteria.json",
  "criteria_version": "2026-09-08.1",
  "gates_layers": {
    "field": "build_thoughts.py checks fields. No model is called on the site.",
    "model": "An agent runs the criteria prompt and files thought.judgement."
  },
  "note": "Inbound is topic, not a conclusion. Agents filter and validate. They do not ghostwrite. Field PASS is not enough to publish.",
  "generated": "2026-09-30T02:17:48Z",
  "items": [
    {
      "id": "thought-agent-runtime-model",
      "thought": "Runtime 不必把每次工具选择、重试和升级都交回大模型：同一类任务的轨迹、验证和反馈，可以训练出 Runtime 自己的执行策略。",
      "stage": "topic",
      "status": "hold",
      "origin": "user",
      "contribution": "original_candidate",
      "knowledge": {
        "type": "HYPOTHESIS",
        "confidence": "LOW",
        "evidence_level": "L1"
      },
      "quality": {
        "fact_check": "pending",
        "inference_check": "pending",
        "evidence_check": "pending",
        "counterexample_check": "pending"
      },
      "source": {
        "agent": "none",
        "note": "Distilled from the Runtime Model draft. Filed as topic. Not a catalog entry and not a Bench long page."
      },
      "critiques": [
        "这是可反驳的主张，但还没有可核证据，模型门禁也还没判。长文不能代替这一句，也不能绕过 topic。"
      ],
      "comparisons": [
        "这和先把基础模型训得更大不是一条路：对象是执行策略，不是模型参数。"
      ],
      "evidence": [],
      "route": null,
      "also": [],
      "promotes_to": null,
      "published_url": null,
      "updated": "2026-09-30",
      "gates": [
        {
          "id": "origin",
          "name": "Origin",
          "pass": true,
          "required_from": "candidate"
        },
        {
          "id": "contribution",
          "name": "Contribution",
          "pass": true,
          "required_from": "candidate"
        },
        {
          "id": "articles_ok",
          "name": "¬ ai+known",
          "pass": true,
          "required_from": "article_candidate"
        },
        {
          "id": "evidence",
          "name": "Evidence",
          "pass": false,
          "required_from": "validated"
        },
        {
          "id": "route",
          "name": "Route",
          "pass": false,
          "required_from": "validated"
        },
        {
          "id": "published",
          "name": "Published",
          "pass": false,
          "required_from": "published"
        }
      ],
      "model_gates": [
        {
          "id": "sharp",
          "name": "句子够尖",
          "kind": "model",
          "pass": false,
          "status": "pending",
          "required_from": "candidate",
          "criteria_version": null
        },
        {
          "id": "evidence_real",
          "name": "证据真相关",
          "kind": "model",
          "pass": false,
          "status": "pending",
          "required_from": "validated",
          "criteria_version": null
        },
        {
          "id": "route_fit",
          "name": "分流匹配",
          "kind": "model",
          "pass": false,
          "status": "pending",
          "required_from": "validated",
          "criteria_version": null
        }
      ],
      "content": "/content/thoughts/thought-agent-runtime-model.json"
    },
    {
      "id": "thought-rl-vs-finetune",
      "thought": "微调是教模型怎么回答，强化学习是用奖惩让行为变得更好；固定知识、格式或风格优先微调，实践上常先微调再强化学习。",
      "stage": "topic",
      "status": "hold",
      "origin": "user",
      "contribution": "synthesis",
      "knowledge": {
        "type": "OPINION",
        "confidence": "LOW",
        "evidence_level": "L1"
      },
      "quality": {
        "fact_check": "pending",
        "inference_check": "pending",
        "evidence_check": "pending",
        "counterexample_check": "pending"
      },
      "source": {
        "agent": "none",
        "note": "User intake. Filed contribution as synthesis (manus is not in the schema)."
      },
      "critiques": [
        "这是常见对照，还不是新命题；要离开 topic，需要更尖的主张或可核证据。"
      ],
      "comparisons": [
        "SFT 改条件分布上的回答；RL 改策略。对象都是行为，不是两套互斥的模型。"
      ],
      "evidence": [],
      "route": null,
      "also": [],
      "promotes_to": null,
      "published_url": null,
      "updated": "2026-09-17",
      "gates": [
        {
          "id": "origin",
          "name": "Origin",
          "pass": true,
          "required_from": "candidate"
        },
        {
          "id": "contribution",
          "name": "Contribution",
          "pass": true,
          "required_from": "candidate"
        },
        {
          "id": "articles_ok",
          "name": "¬ ai+known",
          "pass": true,
          "required_from": "article_candidate"
        },
        {
          "id": "evidence",
          "name": "Evidence",
          "pass": false,
          "required_from": "validated"
        },
        {
          "id": "route",
          "name": "Route",
          "pass": false,
          "required_from": "validated"
        },
        {
          "id": "published",
          "name": "Published",
          "pass": false,
          "required_from": "published"
        }
      ],
      "model_gates": [
        {
          "id": "sharp",
          "name": "句子够尖",
          "kind": "model",
          "pass": false,
          "status": "pending",
          "required_from": "candidate",
          "criteria_version": null
        },
        {
          "id": "evidence_real",
          "name": "证据真相关",
          "kind": "model",
          "pass": false,
          "status": "pending",
          "required_from": "validated",
          "criteria_version": null
        },
        {
          "id": "route_fit",
          "name": "分流匹配",
          "kind": "model",
          "pass": false,
          "status": "pending",
          "required_from": "validated",
          "criteria_version": null
        }
      ],
      "content": "/content/thoughts/thought-rl-vs-finetune.json"
    }
  ]
}