Skip to content
Back to skills

Llm As A Verifier

ASecurity

可选 LLM-as-a-Verifier。测试已绿且有多个合格 diff 时做 best-of-N 排序;不是 ship 门禁,不替代 test/review/VM。

  • 188 stars
  • 0 votes
  • 0 copies
  • 0 views
  • Added September 22, 2026
ai-agentspythonbashgitapi

Works with

  • api

Security analysis

A100/100

Pro scans all 4 files and shows the line behind each finding

Scanned September 22, 2026

npx -y skills add WenJunDuan/Rlues --skill llm-as-a-verifier --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Llm As A Verifier?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Llm As A Verifier
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/wenjunduan-llm-as-a-verifier-rlues/badge)](https://www.skillsdirectory.com/skills/wenjunduan-llm-as-a-verifier-rlues)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: llm-as-a-verifier
description: 可选 LLM-as-a-Verifier。测试已绿且有多个合格 diff 时做 best-of-N 排序;不是 ship 门禁,不替代 test/review/VM。
---

# /llm-as-a-verifier — 可选排序器 (v9.9.9)

9.9.8 把 LaaV 留作 opt-in 槽([design B](https://arxiv.org/abs/2607.05391)、[llm-as-a-verifier](https://github.com/llm-as-a-verifier/llm-as-a-verifier))。本版提供可调用 skill 与脚本,**默认关闭**,不进入 PACE 默认链,不新建 stage。

## 何时用

同时满足:

1. `_index.laav_enabled = true`(`/athena-preferences` 写入),或用户本轮显式要求排序
2. 相关测试已绿
3. 面前有 **≥2 个合格候选**(独立 worktree / 补丁 / 实现变体)
4. 发生在 runtime-verify 之后、一次独立 review **之前**

缺任一条件 → skip,写明原因。不要为了用 LaaV 去生成多余候选。

## 怎么用

```bash
python3 ~/.agents/skills/llm-as-a-verifier/scripts/rank.py \
  --packet .ai_state/sprints/<slug>/review-packet.md \
  --candidate a.diff --candidate b.diff \
  --output .ai_state/.runtime/verifier_rank.json
```

- 后端必须能返回 score token 的 **logprobs**(OpenAI-compatible / Gemini / DeepSeek)。用 `ATHENA_LAAV_BASE_URL` + `ATHENA_LAAV_API_KEY`(或该 URL 对应的官方 key env)。
- Claude / 无 logprobs 的自定义网关 → 脚本 exit 0 且 `status: skipped`,**不得**用 1–5 离散分冒充 LaaV。
- 评分标准 = 当前 `review-packet.md` 的 Done checks,禁止另造「感觉更好」。
- 输出只排序。`reviewer` 可以读 `verifier_rank`,**不能**把它写成 VERDICT。delivery-gate / 测试 / VM 实跑仍是法律。

## 硬边界

1. 不替换 review、不替换 VM、不为 LaaV 新建 PACE stage。
2. 不改全局 `ANTHROPIC_BASE_URL`;独立端口/独立 base URL,fresh install 默认关。
3. 未校准前只记 ignored `.runtime` telemetry,不作为选实现的强制依据。
4. 秘密不进 hash 输入、不进对话。

细则与校准记录见 [playbook](references/playbook.md)。模板:[verifier_rank.json](templates/verifier_rank.json)。

Files in this skill

  • SKILL.md2.1 KB
  • references/playbook.md2.2 KB
  • scripts/rank.py7.1 KB
  • templates/verifier_rank.json377 B

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…