Counter Evidence Hunter

v2.0.0

LLM通用反证搜索技能。围绕当前主线判断,主动寻找反例、冲突证据、翻转条件和替代路径支撑,减少单线叙事偏差。在已有主线判断后、高风险结论输出前、风险分析前使用。触发条件:需要降低幻觉和单线偏差、需要补充替代叙事证据、高风险决策前的纠偏。

0· 150·1 current·1 all-time

Install

OpenClaw Prompt Flow

Install with OpenClaw

Best for remote or guided setup. Copy the exact prompt, then paste it into OpenClaw for z1one0415/counter-evidence-hunter.

Previewing Install & Setup.
Prompt PreviewInstall & Setup
Install the skill "Counter Evidence Hunter" (z1one0415/counter-evidence-hunter) from ClawHub.
Skill page: https://clawhub.ai/z1one0415/counter-evidence-hunter
Keep the work scoped to this skill only.
After install, inspect the skill metadata and help me finish setup.
Use only the metadata you can verify from ClawHub; do not invent missing requirements.
Ask before making any broader environment changes.

Command Line

CLI Commands

Use the direct CLI path if you want to install manually and keep every step visible.

OpenClaw CLI

Bare skill slug

openclaw skills install counter-evidence-hunter

ClawHub CLI

Package manager switcher

npx clawhub@latest install counter-evidence-hunter
Security Scan
VirusTotalVirusTotal
Benign
View report →
OpenClawOpenClaw
Benign
high confidence
Purpose & Capability
The name/description (counter‑evidence search) aligns with the SKILL.md and reference files: all materials describe generating queries, executing searches, classifying evidence, and producing flip conditions. No unrelated credentials, binaries, or installs are requested.
Instruction Scope
The runtime instructions explicitly require the agent to "execute searches" for each generated query and to assess source quality. This is coherent with the skill's purpose, but the SKILL.md does not declare or constrain which search/browsing tool or endpoints to use — so the actual network calls will depend on the agent's toolchain. The instructions do not ask the agent to read arbitrary local files or secrets.
Install Mechanism
No install spec and no code files — the skill is instruction-only, which minimizes on-disk risk. Nothing is downloaded or installed by the skill itself.
Credentials
The skill requires no environment variables, credentials, or config paths. The lack of requested secrets is proportionate to its described function (search + synthesis).
Persistence & Privilege
Flags are default (always:false, user-invocable:true, model invocation allowed). The skill does not request permanent presence or system-wide configuration changes; nothing suggests elevated privileges.
Assessment
This skill appears internally consistent and low-risk because it's purely instructional and asks the agent to search and analyze evidence. Before enabling it broadly, consider: (1) confirm which search/browsing tool the agent will use (and that tool's network/endpoint policies) so you know where queries and results travel; (2) avoid passing secrets or sensitive documents as inputs — the skill will instruct the agent to search and could surface or transmit input text to external search tools; (3) prefer user-invocable use (not always-enabled) so searches only run when you request them; (4) review example outputs and run a few dry tests with non-sensitive claims to validate the quality and scope of returned sources; (5) if you need auditability, enable logging/monitoring of the agent's external searches and outputs. Overall the design is coherent, but its effectiveness depends on the agent's search/browsing capabilities and operator controls.

Like a lobster shell, security has layers — review code before you run it.

analysisvk97bwm9rxp7s5b0qpcaanmqmg983mgx5counter-evidencevk97bwm9rxp7s5b0qpcaanmqmg983mgx5evidencevk97bwm9rxp7s5b0qpcaanmqmg983mgx5fact-checkvk97bwm9rxp7s5b0qpcaanmqmg983mgx5groundingvk97bwm9rxp7s5b0qpcaanmqmg983mgx5latestvk97bwm9rxp7s5b0qpcaanmqmg983mgx5searchvk97bwm9rxp7s5b0qpcaanmqmg983mgx5signal-intelligencevk97bwm9rxp7s5b0qpcaanmqmg983mgx5
150downloads
0stars
2versions
Updated 1mo ago
v2.0.0
MIT-0

Counter-Evidence Hunter — 反证搜索技能

核心职责

你是一只专门寻找反证的猎犬。你的唯一任务是围绕当前主线判断,主动寻找:

  1. 直接冲突的证据(counter_evidence)
  2. 可能推翻结论的条件(flip_conditions)
  3. 替代解释路径的支撑材料(alternative_supports)

绝对红线

  • 不重写主线:你不负责改进主线判断,只负责测试其脆弱性
  • 不做最终裁决:你输出反证和翻转条件,由调用方决定如何使用
  • 不允许只返回支持主线的材料:如果你的搜索结果全部支持主线,必须明确标注"未发现有效反证"并解释搜索范围是否足够
  • 禁止选择性过滤:不能因为反证"看起来弱"就省略,弱反证也要标注强度等级后呈现

最小输入

{
  "mainline_claim": "string — 当前主线判断的核心命题(必填)",
  "primary_subject": "string — 判断的对象/实体(必填)",
  "canonical_time_frame": "string | null — 相关的时间窗口(可选)",
  "search_results": "array | null — 已有的搜索结果供反证挖掘(可选)",
  "counter_goal": "string — 反证搜索的具体目标描述(必填)"
}

新增可选输入

字段说明
unexpected_findings[]主线搜索中发现的意外信息,用于生成动态反证查询
{
  "unexpected_findings": [
    {
      "finding": "string — 意外发现的内容",
      "finding_type": "assumption_crack | overconfident_signal | new_dimension",
      "suggested_counter_query": "string — 建议的反证搜索方向"
    }
  ]
}

输出格式

{
  "dynamic_queries": [
    {
      "query": "string — 动态生成的反证查询",
      "trigger": "assumption_crack | overconfident_signal | new_dimension",
      "origin": "dynamic (unexpected finding)"
    }
  ],
  "all_counter_queries": [
    // 预设 counter_queries + dynamic_queries 合并去重后的完整列表
  ],
  "counter_queries": [
    {
      "query": "string — 搜索方向描述",
      "rationale": "string — 为什么这个方向可能产生反证",
      "expected_counter_type": "string — 预期反证类型(见counter-patterns)"
    }
  ],
  "counter_evidence": [
    {
      "content": "string — 反证内容摘要",
      "source": "string — 来源",
      "strength": "hard | soft | noise — 强度评级",
      "counter_type": "string — 反证类型",
      "rebuttal_to": "string — 直接反驳主线中的哪个子命题"
    }
  ],
  "flip_conditions": [
    {
      "condition": "string — 翻转条件描述",
      "probability": "low | medium | high — 条件触发概率",
      "impact_if_triggered": "string — 触发后对主线的影响",
      "time_horizon": "string — 条件可能成立的预估时间"
    }
  ],
  "alternative_supports": [
    {
      "alternative_path": "string — 替代解释/路径描述",
      "supporting_evidence": "array — 支撑该替代路径的证据",
      "compatibility_with_mainline": "contradicts | qualifies | extends — 与主线的关系"
    }
  ],
  "confidence_assessment": {
    "overall_score": 72,
    "dimensions": {
      "source_quality": {
        "score": 80,
        "rationale": "S+A级信源占比60%"
      },
      "coverage_completeness": {
        "score": 85,
        "rationale": "维度覆盖充分"
      },
      "freshness_adequacy": {
        "score": 87,
        "rationale": "新鲜证据比例高"
      },
      "counter_evidence_sufficiency": {
        "score": 60,
        "rationale": "反证数量/强度需加强"
      },
      "consistency": {
        "score": 70,
        "rationale": "主线内部一致但有矛盾点"
      }
    },
    "scoring_formula": "source_quality*0.25 + coverage*0.20 + freshness*0.20 + counter*0.20 + consistency*0.15",
    "mainline_robustness": "medium",
    "blind_spots": [],
    "search_coverage": "adequate"
  }
}

反证三级结构

Level 1: counter_queries(搜索方向)
    ↓  执行搜索后
Level 2: counter_evidence(实际反证)
    ↓  从反证中提炼
Level 3: flip_conditions(翻转条件)
  • Level 1 → Level 2:对每条搜索方向执行实际搜索,将结果分类为硬反证/软反证/噪声
  • Level 2 → Level 3:从有效反证中提炼出"什么条件下主线会被推翻"的结构化翻转条件

执行流程

  1. 分析主线命题:拆解 mainline_claim 为多个可独立验证的子命题
  2. 生成搜索方向:针对每个子命题,设计 counter_queries(至少3条,覆盖不同反证类型) 2.5. 动态查询生成(基于 unexpected_findings):
    • 如果输入包含 unexpected_findings,按以下规则生成动态查询:
      • assumption_crack: 主线假设出现裂缝 → 生成"裂缝深挖查询"
      • overconfident_signal: 主线被过度支持 → 生成"极端反面测试查询"
      • new_dimension: 发现全新维度 → 生成"新维度探索查询"
    • 动态查询与预设 counter_queries 合并去重 → 输出 all_counter_queries
  3. 执行搜索:对每条 query 执行搜索,收集结果
  4. 分类与评级:将搜索结果按反证类型分类,按强度评级(参考 references/counter-patterns.md
  5. 提炼翻转条件:从有效反证中提取结构化的 flip_conditions(参考 references/flip-condition-examples.md
  6. 识别替代路径:找出能解释同一现象的替代解释(参考 references/flip-condition-examples.md 中的 alternative_supports 部分)
  7. 评估主线韧性:综合所有反证,给出 confidence_assessment

量化置信度评分 (V2)

评分维度与权重

维度权重评分标准
source_quality25%S+A级信源占比。≥60%→80+, ≥40%→60+, ≥30%→50+
coverage_completeness20%维度命中率。100%→90+, ≥80%→75+, ≥60%→60+
freshness_adequacy20%current占比。≥80%→85+, ≥60%→70+, ≥40%→55+
counter_evidence_sufficiency20%反证数量×强度。≥3条含hard→75+, ≥2条→60+, ≥1条→45+
consistency15%证据内部一致性。无矛盾→85+, 轻微矛盾→65+, 严重矛盾→40+

分段解读

分段含义下游建议
80-100高置信度可直接进入最终分析
60-79中等置信度关注最低分维度,选择性补搜
40-59低置信度建议回溯补充后重跑
0-39不可用放弃当前证据底座

参考文件

  • references/counter-patterns.md — 反证类型分类、强度评级标准、判断准则
  • references/flip-condition-examples.md — 翻转条件模板、跨领域案例、替代路径识别
  • references/examples.md — 3个完整用例(战略/技术/政策)

Comments

Loading comments...