137 lines
10 KiB
HTML
137 lines
10 KiB
HTML
<!doctype html>
|
||
<html lang="zh-CN">
|
||
<head>
|
||
<meta charset="utf-8">
|
||
<meta name="viewport" content="width=device-width, initial-scale=1">
|
||
<title>正文智能体正式优化架构</title>
|
||
<style>
|
||
:root {
|
||
color-scheme: light;
|
||
--ink: #182128;
|
||
--muted: #596872;
|
||
--line: #c8d0d6;
|
||
--paper: #f5f7f8;
|
||
--blue: #dcecff;
|
||
--green: #ddf2e5;
|
||
--amber: #fff0c7;
|
||
--violet: #ebe3f8;
|
||
--red: #f8dddd;
|
||
--cyan: #def0f1;
|
||
}
|
||
* { box-sizing: border-box; }
|
||
body { margin: 0; background: var(--paper); color: var(--ink); font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", sans-serif; letter-spacing: 0; }
|
||
main { width: min(1240px, calc(100% - 32px)); margin: 28px auto 48px; }
|
||
h1 { margin: 0 0 8px; font-size: 28px; }
|
||
.lead { margin: 0 0 26px; color: var(--muted); line-height: 1.6; }
|
||
section { margin-top: 30px; }
|
||
h2 { margin: 0 0 12px; font-size: 19px; }
|
||
.lane { display: grid; grid-template-columns: repeat(6, minmax(130px, 1fr)); gap: 12px; align-items: stretch; }
|
||
.node { min-height: 112px; padding: 14px; border: 1px solid var(--line); border-radius: 8px; background: white; position: relative; }
|
||
.node:not(:last-child)::after { content: "→"; position: absolute; right: -13px; top: 42px; z-index: 2; font-size: 18px; color: #53616c; }
|
||
.node strong { display: block; margin-bottom: 7px; font-size: 15px; }
|
||
.node span { color: var(--muted); font-size: 13px; line-height: 1.45; }
|
||
.outline { background: var(--blue); }
|
||
.cards { background: var(--green); }
|
||
.prose { background: var(--amber); }
|
||
.writer { background: var(--violet); }
|
||
.detector { background: var(--red); }
|
||
.decision { background: var(--cyan); }
|
||
.loops { display: grid; grid-template-columns: repeat(3, 1fr); gap: 14px; margin-top: 14px; }
|
||
.loop { padding: 14px 16px; border-left: 4px solid #687d8a; background: white; }
|
||
.loop b { display: block; margin-bottom: 6px; }
|
||
.loop p { margin: 0; color: var(--muted); font-size: 13px; line-height: 1.55; }
|
||
.eval { display: grid; grid-template-columns: repeat(6, 1fr); gap: 12px; }
|
||
.runtime { display: grid; grid-template-columns: 1fr 1.35fr 1fr; gap: 14px; align-items: stretch; }
|
||
.runtime-group { display: grid; gap: 10px; }
|
||
.runtime > .node, .runtime-group .node { min-height: 0; }
|
||
.runtime > .node::after, .runtime-group .node::after { display: none; }
|
||
.orchestrator { background: var(--cyan); border-width: 2px; }
|
||
.vault { background: var(--amber); }
|
||
.closed { background: var(--red); }
|
||
.contract { display: grid; grid-template-columns: repeat(3, 1fr); gap: 14px; margin-top: 14px; }
|
||
.contract article { padding: 15px; border: 1px solid var(--line); border-radius: 8px; background: white; }
|
||
.contract h3 { margin: 0 0 7px; font-size: 15px; }
|
||
.contract p { margin: 0; color: var(--muted); font-size: 13px; line-height: 1.55; }
|
||
.evidence, .orders { display: grid; grid-template-columns: 1fr 1fr; gap: 14px; }
|
||
.evidence article, .orders article { padding: 16px; border: 1px solid var(--line); border-radius: 8px; background: white; }
|
||
.evidence h3, .orders h3 { margin: 0 0 7px; font-size: 15px; }
|
||
.evidence p, .orders p { margin: 0; color: var(--muted); font-size: 13px; line-height: 1.55; }
|
||
@media (max-width: 980px) { .lane, .eval, .runtime { grid-template-columns: 1fr 1fr; } .node::after { display: none; } .loops, .contract, .evidence, .orders { grid-template-columns: 1fr; } }
|
||
@media (max-width: 540px) { .lane, .eval, .runtime { grid-template-columns: 1fr; } }
|
||
</style>
|
||
</head>
|
||
<body>
|
||
<main>
|
||
<h1>正文智能体正式优化架构</h1>
|
||
<p class="lead">卡只负责定位原文，writer 不读取卡语义。Gate A 尚未 real-run；Gate B 通过前只实现离线评测闭环。</p>
|
||
|
||
<section>
|
||
<h2>两个方向不能混用</h2>
|
||
<div class="orders">
|
||
<article><h3>能力验证</h3><p>正文智能体 → 细纲智能体 → 大纲+设定智能体。自底向上隔离变量。</p></article>
|
||
<article><h3>正式生产</h3><p>大纲+设定 → 细纲 → 正文智能体。自顶向下提供创作输入。</p></article>
|
||
</div>
|
||
</section>
|
||
|
||
<section>
|
||
<h2>产品化生产链（Gate B 通过后）</h2>
|
||
<div class="lane">
|
||
<div class="node outline"><strong>1. 已确认细纲</strong><span>硬事件、可调节拍、伏笔、钩子和新设定声明。</span></div>
|
||
<div class="node cards"><strong>2. 检索规划与卡索引</strong><span>按 asOf 重建当前态，带出历史来源指针。</span></div>
|
||
<div class="node prose"><strong>3. 连续四章 + 原文回读</strong><span>事实证据与表现证据分开装配，生成确定性 manifest。</span></div>
|
||
<div class="node writer"><strong>4. 无工具 Writer</strong><span>只接收 WriterContext；可按预算请求补证。</span></div>
|
||
<div class="node detector"><strong>5. Detector 硬门</strong><span>核细纲、事实、证据、时间线和输出合同，最多返修两轮。</span></div>
|
||
<div class="node decision"><strong>6. Shadow 与用户三决策</strong><span>第二阶段才实现。通过 accept_preflight 后 confirm，不属于 Gate A/B 前置范围。</span></div>
|
||
</div>
|
||
<div class="loops">
|
||
<div class="loop"><b>补证环</b><p>Writer 可返回结构化 evidenceRequests；最多 3 次，累计新增上下文不超过检索预算 30%。</p></div>
|
||
<div class="loop"><b>返修环</b><p>Detector 只阻断可验证硬错误；Writer 最多自动返修 2 轮，仍失败则交用户。</p></div>
|
||
<div class="loop"><b>创作自由边界</b><p>硬事件、结果和伏笔不可改；场景可拆并调序，可创造环境、动作、过渡和对白；新增 Canonical 事实必须申报。</p></div>
|
||
</div>
|
||
</section>
|
||
|
||
<section>
|
||
<h2>双证据模型</h2>
|
||
<div class="evidence">
|
||
<article><h3>Fact Evidence</h3><p>正式设定、Canonical 卡、状态台账、已确认正文和细纲声明的新事实，负责“写得对”。没有可靠事实来源的硬结论必须阻断。</p></article>
|
||
<article><h3>Prose Evidence</h3><p>历史原文场景，负责人物声音、动作习惯、能力表现、段落节奏和章末钩子，回答“怎么写得像”。</p></article>
|
||
</div>
|
||
</section>
|
||
|
||
<section>
|
||
<h2>离线评测链</h2>
|
||
<div class="eval">
|
||
<div class="node outline"><strong>A/C 核心比较 + B 负对照</strong><span>A=通用原文检索；C=卡索引后回读原文；同字符预算。B=card-only，仅诊断。全评测集统一分配顺序并平衡位置。</span></div>
|
||
<div class="node writer"><strong>隔离 Writer</strong><span>A/C 只看到统一事实约束和选中原文，看不到卡、检索策略或轨迹。完整上下文只允许原文与派生 hash 不同。</span></div>
|
||
<div class="node detector"><strong>统一 Detector</strong><span>严格真实报告逐项计算高严重度与硬约束覆盖率，禁止常量。</span></div>
|
||
<div class="node writer"><strong>Oracle 盲评</strong><span>五维评分之外，逐 assertionId/constraintId 输出 pass/fail/unknown verdict；绑定 candidate hash 与受控证据，unknown 或 ID 不全即失败。</span></div>
|
||
<div class="node prose"><strong>唯一 GateInputBuilder</strong><span>只比较 detector/judge 同类同 ID verdict，禁止用分数推 verdict；只从 manifest、receipts 和严格报告生成 gate-input.json。</span></div>
|
||
<div class="node decision"><strong>Gate A / Gate B</strong><span>系统/schema/hash/泄漏/judge 异常先判 failed；只有预注册集合自身不足才是 insufficient。B 不进 Gate B 阈值，按新角色比例分层报告。</span></div>
|
||
</div>
|
||
</section>
|
||
|
||
<section>
|
||
<h2>agent-example 离线评测 adapter</h2>
|
||
<div class="runtime">
|
||
<div class="node orchestrator"><strong>可信编排器</strong><span>唯一持有授权、臂映射、顺序表、oracle 和预算。直连固定 Claude CLI 是离线例外，不写 New-API 共享账本，也不是产品运行时方案。</span></div>
|
||
<div class="runtime-group">
|
||
<div class="node writer"><strong>隔离模型进程</strong><span>绝对 CLI 路径 + sha256 + 2.1.211 + 完整 model ID；stdin-only、fresh process、无工具/会话/hooks/plugins/MCP。</span></div>
|
||
<div class="node detector"><strong>隔离 Semantic Detector 进程</strong><span>候选与硬约束 → 严格语义报告。报告绑定 candidate hash，指标由逐项结果计算。</span></div>
|
||
<div class="node outline"><strong>sandbox-exec 边界</strong><span>cwd/HOME/TMPDIR 位于隔离目录；目录外文件读写拒绝，仅允许模型调用所需网络，环境字段白名单。</span></div>
|
||
</div>
|
||
<div class="runtime-group">
|
||
<div class="node vault"><strong>vault lease → /private/tmp</strong><span>用户/创始人批准；先写非敏感 lease，再写 0700/0600 raw。恢复扫描 lease 清孤儿，防 symlink/path escape。</span></div>
|
||
<div class="node decision"><strong>gate-report → passed receipt</strong><span>writer_gate.py 是唯一签发者；同一文件 CAS journal 中先落报告，只有 passed 才签 receipt，细纲入口验完整 hash 链。</span></div>
|
||
<div class="node closed"><strong>任一异常失败关闭</strong><span>timeout、nonzero、schema、budget、model mismatch、invalid judge、unstable judge 都使样本 ok=false，禁止换模或跳臂。</span></div>
|
||
</div>
|
||
</div>
|
||
<div class="contract">
|
||
<article><h3>固定调用面</h3><p>print、bare、完整 model ID、effort、预算、deadline、JSON schema、无工具/会话，空 MCP 固定为 {"mcpServers":{}}。temperature/topP/seed 不作支持声明。</p></article>
|
||
<article><h3>联合成功判定</h3><p>exit=0、is_error=false、正常 terminal reason、完整 model 匹配、usage 可核账且 structured_output 合法才成功；result 不作业务 fallback。仍需一次成功探测。</p></article>
|
||
<article><h3>两阶段硬边界</h3><p>第一阶段实现 receipt 签发器与细纲入口守卫；第二阶段才用已验签 receipt 启动细纲 real-run。拒绝自算 JSON，dry-run 豁免。</p></article>
|
||
</div>
|
||
</section>
|
||
</main>
|
||
</body>
|
||
</html>
|