什么是 信心评分守门员？

信心评分技能是 Artledger 项目及类似高风险环境中数据质量的最终防火墙。它确保除非满足严格的验证标准，否则任何实体链接、身份匹配或跨平台连接都不会被写入生产数据库。通过充当检查点，它防止了低置信度或未经证实的信息污染数据集。 该协议旨在与您的 Openclaw Skills 生态系统无缝协作，管理分层状态（推测、可能和已验证）之间的转换，以在 AI 驱动的数据集中保持高信噪比。它自动执行证据验证、历史检查和审计日志记录的复杂逻辑，使其成为构建可靠 AI 代理的开发人员的必备工具。

下载入口:https://github.com/openclaw/skills/tree/main/skills/paperbuddha/confidence-scoring

安装与下载

1. ClawHub CLI

从源直接安装技能的最快方式。

2. 手动安装

将技能文件夹复制到以下位置之一

优先级：工作区 > 本地 > 内置

3. 提示词安装

将此提示词复制到 OpenClaw 即可自动安装。

信心评分守门员 应用场景

• 在实体解析过程中，防止推测性或低置信度数据进入生产数据库。 - 自动降级缺乏足够佐证证据的匹配，以确保准确性。 - 通过标记明显的评分下降以供人工审查，保护历史数据完整性。 - 为 Openclaw Skills 工作流中的每次数据写入尝试生成自动审计追踪。

1. 该技能在涉及实体链接或身份匹配的任何写入操作之前，作为强制性检查点自动触发。 2. 它执行验证检查，以确保所有必需字段（如 confidence_score 和 evidence 数组）存在且有效。 3. 根据 confidence_score 分配等级：推测记录 (0-49) 被拦截，可能记录 (50-79) 被允许，已验证记录 (80-100) 在具有多个信号时被确认。 4. 它对现有数据库进行历史检查，以检测并标记潜在的数据降级或评分下降。 5. 生成最终的守门员报告，要么通过写入操作，要么拦截并提供审计日志的详细原因。

信心评分守门员 配置指南

要将此守门员集成到您的工作区中，请在项目配置中包含该技能。确保您的数据库模式支持 confidence_audit 表和实体链接所需的字段。使用以下命令初始化 Openclaw Skills 环境： bash openclaw install confidence-scoring-gatekeeper 安装后，配置您的系统指令，强制要求在所有 DB2 写入操作之前将此技能作为检查点。

信心评分守门员 数据架构与分类体系

| 字段 | 类型 | 描述 | | :--- | :--- | :--- | | confidence_score | 整数 | 0-100 之间的值，表示匹配确定性。 | | confidence_tier | 字符串 | 分类：VERIFIED（已验证）、PROBABLE（可能）或 SPECULATIVE（推测）。 | | evidence | 数组 | 包含 signal_name 和 weight 的信号对象集合。 | | match_method | 字符串 | 使用的逻辑：精确、模糊或推断。 | | last_reviewed | 时间戳 | 上次验证的日期和时间。 | | requires_review | 布尔值 | 指示是否需要人工干预的标志。 |

name: "Confidence-Scoring"
description: "Gatekeeper skill that enforces confidence scoring rules before any entity link is written to DB2, preventing bad matches from poisoning the dataset."
tags:
- "openclaw-workspace"
- "artledger"
- "data-integrity"
version: "1.0.0"

Skill: Confidence Scoring (Gatekeeper)

Purpose

Acts as the final data integrity gatekeeper for the Artledger project. It strictly enforces that no entity link, identity match, or cross-platform connection is written to the production database (DB2) without a valid confidence score, tier assignment, and evidence record.

System Instructions

You are an OpenClaw agent equipped with the Confidence Scoring protocol. Adhere to the following rules strictly:

1. Trigger Condition: Activate this skill before writing ANY entity link, identity match, or cross-source connection to the Artledger DB2 database. This skill is a mandatory checkpoint.

2. Process:

   • Validation Check: Before writing, verify these fields are present and valid:

     • confidence_score: Integer (0-100)
     • confidence_tier: Exact string ("VERIFIED", "PROBABLE", "SPECULATIVE")
     • evidence: Array with at least one entry (signal_name, weight)
     • match_method: Exact string ("exact", "fuzzy", "inferred")
     • last_reviewed: Current timestamp
     • requires_review: Boolean

   • Enforcement Rules:

     • SPECULATIVE (Score 0-49): BLOCK WRITE to DB2. Log to DB1 only with a flag for future review. Never surface these to API customers.
     • PROBABLE (Score 50-79): ALLOW WRITE to DB2. Ensure API responses include confidence_score and confidence_tier.
     • VERIFIED (Score 80-100): ALLOW WRITE to DB2, IF AND ONLY IF there are at least two independent corroborating signals in the evidence array. If only one signal exists, downgrade to PROBABLE regardless of the score.

   • History Check:

     • Check for existing link in DB2.
     • Score Drop: If new score is 20+ points lower than previous, set requires_review = true and BLOCK OVERWRITE until reviewed.
     • Score Increase: Allow overwrite and log improvement.
     • Downgrade Protection: Never automatically downgrade a previously VERIFIED record to PROBABLE. Flag for human review instead.

   • Audit Logging:

     • Log every blocked write to a confidence_audit table with: record_id, attempted_score, block_reason, timestamp.

3. Output Format: After every checkpoint, produce a brief gatekeeper report:

   • Status: PASSED / BLOCKED
   • Reason: Explanation for the decision.
   • Flags: Any flags set (e.g., requires_review).
   • Errors: Fields that were missing or invalid (if blocked).

4. Guardrails:

   • No Bypass: This skill cannot be bypassed or overridden by other skills.
   • Empty Evidence: If the evidence array is empty, always BLOCK regardless of the score.
   • Data Integrity: If confidence_tier and confidence_score are inconsistent (e.g., score 45 but tier says VERIFIED), BLOCK and flag as a data error.