AaronandWork/dsh-evidence-gate ↗★ 3
@aaronandwork/dsh-evidence-gate
Evidence-first conclusion gate for DeepSeek Harness (DSH): 3-tier evidence hierarchy policy, a model tool that verdicts claims (VERIFIED/REFERENCED/UNVERIFIED), durable per-session tool-activity cross-checking (arguments + output anchors, restart-surviving), per-session switch + persisted ledger, and a session-scoped UI toggle.
安装
npx -p @deepseek-ai/dsh dsh plugin --profile web add github:AaronandWork/dsh-evidence-gate说明文档
阅读完整 README ↗Usage
| Command | Effect |
|---|---|
/evidence-gate on | enable the gate for this session (policy text + verdicts active on the next model step; state persists across restarts) |
/evidence-gate off | disable for this session (the gate tool stays registered but returns ⚪ 门未开启 and counts nothing) |
/evidence-gate status | one-line per-session state + ledger + cross-check/采信 hits + index size |
/evidence-gate report | full session report: verdict groups, by-basis tally, blocked/converted/采信, recent claims |
The composer's shield chip toggles the same per-session state; its tooltip shows this session's numbers only. The agent can also be asked to toggle/report — it calls evidence_gate_switch (its description restricts on/off to explicit user request).
Reading the numbers
统计 Stats: 实证 verified 12 · 引用 cited 3 · 拦截 blocked 5 · 无效 invalid 0 · 转化 converted 3 · 核验命中 cross-checked 9 · 采信 floor-passed 1（索引 index 214）
reads as: 12 claims passed (9 corroborated by the evidence index, 1 admitted on the existence floor); 5 guesses were caught and downgraded; 3 of those came back verified after the agent actually ran the verification; the durable index currently holds 214 tool-activity entries.
Localization / 双语
The web UI (chip + cards) is bilingual: it follows the GUI language setting (Settings → Language), falls back to the browser language, then to Chinese. Switching the GUI language re-renders the gate UI live. Host-side tool/command output is compact-bilingual — verdict banners (【已实证 Verified】), action labels and the report header carry English glosses so both audiences can scan them, while instruction sentences stay Chinese (the model relays them in the conversation language). 工具与命令输出为紧凑双语：判定横幅、动作短语与报告表头带英文对照，指令句保持中文（由模型按对话语言转述）。
使用 / Usage
| 命令 | 效果 |
|---|---|
/evidence-gate on | 仅为当前会话开启（下一模型步生效；状态跨重启保留） |
/evidence-gate off | 当前会话关闭（门工具保持注册但返回 ⚪ 门未开启，不计数） |
/evidence-gate status | 一行本会话状态 + 台账 + 核验命中/采信 + 索引大小 |
/evidence-gate report | 本会话完整报告：判定分组、basis 分布、拦截/转化/采信、最近声称 |
输入框旁的盾形开关切换同一份按会话状态；悬停只显示本会话数字。也可以直接让 agent 开关/汇报——它会调用 evidence_gate_switch。
数字读法：统计 Stats: 实证 verified 12 · 引用 cited 3 · 拦截 blocked 5 · 无效 invalid 0 · 转化 converted 3 · 核验命中 cross-checked 9 · 采信 floor-passed 1（索引 index 214） —— 12 条结论过门（9 条有索引佐证、1 条按存在性底线采信）；5 个猜测被拦下；其中 3 个在 agent 真正跑完验证后翻案；持久索引现有 214 条工具活动。
双语：web UI（开关 + 卡片）中英双语，跟随 GUI 语言设置实时切换；宿主侧工具/命令输出为紧凑双语（判定横幅、动作短语与报告表头带英文对照）。