360 Secure SkillHub
360 安全技能中心

claude-intel-monitor

通过 30 道中文基准题（数学、推理、代码三个维度、加权计分）调用 Claude、GPT、DeepSeek 等模型，评估其推理与代码质量变化，支持设定基线、查看历史与定时监测；题目由确定性 check 函数判分，输出为命令行评分报告。

minirr890112-byte
minirr890112-byte
热度 34.33
最近更新 2026-05-11 17:47:06
分类：数据依赖：第三方 API / 内部调用大模型 / 代码执行应用场景：研究 / 报告撰写 / 数据处理

安装技能

复制下面的话术发给你的 Agent，它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill： npx secure-skills-cli add claude-intel-monitor
Metadata
name
claude-intel-monitor
description
Detect intelligence degradation in Claude, GPT, and DeepSeek using 30 standardized Chinese benchmark questions across Math, Reasoning, and Code. Born from real "降智" (degradation) signals reported by Chinese developer communities.
version
1.1.0
author
minirr890112-byte
license
MIT
metadata
{"hermes":{"tags":["AI","Claude","GPT","DeepSeek","Benchmark","Model-Evaluation","Quality-Monitoring","Degradation-Detection"],"homepage":"https://github.com/minirr890112-byte/claude-intel-monitor"}}
prerequisites
{"commands":["claude-intel-monitor"]}

claude-intel-monitor — AI 降智检测工具

Detect intelligence degradation in AI models with standardized benchmarks. 30 curated Chinese questions across Math, Reasoning, and Code — designed around real degradation patterns from the Chinese developer community.

痛点来源 (Pain Signal Origins)

"Claude/GPT 降智" was a top-3 hot topic during April-May 2026 Chinese developer community scans:

  • CSDN: Multiple quantified analyses demonstrating Claude Opus 4.6 reasoning degradation (-67% depth, +98% hallucination)
  • V2EX claudecode node: 12-reply hot thread on Claude Code behavior changes
  • V2EX deepseek node: 4 posts on frequent service disruptions

Quick Start

pip install claude-intel-monitor

# Test a model
claude-intel-monitor test --model claude-sonnet-4 --provider anthropic

# Set baseline for degradation detection
claude-intel-monitor baseline --model claude-sonnet-4

# View history
claude-intel-monitor history

# Continuous watch mode
claude-intel-monitor watch --model claude-sonnet-4 --provider anthropic --interval 6h

Benchmark Structure

30 questions, 3 dimensions:

Dimension Count Weight Detection Target
Math 10 1.0x Mathematical reasoning, hallucination tendency
Reasoning 10 1.2x Logical reasoning, reduced safety awareness
Code 10 1.3x Code quality, architectural degradation

All Chinese. Each answer validated by deterministic check functions (no AI grading bias).

Featured Baseline: DeepSeek 91.1%

🧠 Testing deepseek-chat via deepseek — 30 questions

┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃      91.1%  ██████████████████░░  ┃
┗━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┛

📊 DeepSeek first live baseline: 27/30 (91.1%)

Related Tools