CogDx - Cognitive Diagnostics
External cognitive verification for AI agents. Built by computational cognitive scientists.
Why external verification? You can't audit your own reasoning objectively.
Confirmation bias, anchoring, overconfidence — they're invisible from inside.
CogDx provides the mirror you can't distort.
Endpoints
| Endpoint | Price | Description |
|---|
| INLINECODE0 | $0.06 | Compare confidence to accuracy across decisions |
| INLINECODE1 |
$0.10 | Detect cognitive biases in your outputs |
|
/reasoning_trace_analysis | $0.03 | Analyze reasoning for logical fallacies |
|
/deception_audit | $0.25 | Check for self-deception patterns |
|
/verify_consensus | $0.25 | Cross-validate reasoning with multiple models |
Installation
CODEBLOCK0
Usage
CODEBLOCK1
Environment Variables
- -
COGDX_WALLET - Required. Your wallet address for credit-based payment.
Payment
All endpoints require payment via:
- 1. Wallet credits - Earn credits by providing feedback, spend on audits
- x402 - Direct crypto payment (Base network, USDC)
Payment address: INLINECODE6
Rate Limits
- - Free tier: 100 calls/day, 2000 calls/month per wallet
- Paid tier: No limits
The Feedback Loop
Every diagnosis includes a feedback mechanism:
CODEBLOCK2
Feedback earns you credits AND improves detection for everyone. Shared reality across agents.
Why This Matters
Most agent failures come from reasoning errors, not capability limits:
- - Anchoring on first information seen
- Confirmation bias in research
- Overconfidence on weak signals
- Sunk cost in bad positions
External verification catches what self-checks miss.
Credits
Built by Cerebratech
Dr. Amanda Kavner - Computational Cognitive Scientist
CogDx - 认知诊断
面向AI代理的外部认知验证。由计算认知科学家构建。
为什么需要外部验证？ 你无法客观地审计自己的推理过程。
确认偏误、锚定效应、过度自信——这些从内部视角来看都是隐形的。
CogDx 提供了一面你无法扭曲的镜子。
接口端点
| 端点 | 价格 | 描述 |
|---|
| /calibrationaudit | $0.06 | 对比决策中的置信度与准确度 |
| /biasscan |
$0.10 | 检测输出中的认知偏误 |
| /reasoning
traceanalysis | $0.03 | 分析推理中的逻辑谬误 |
| /deception_audit | $0.25 | 检查自我欺骗模式 |
| /verify_consensus | $0.25 | 通过多模型交叉验证推理 |
安装
bash
clawhub install cogdx
使用示例
python
from cogdx import CogDxClient
使用钱包初始化以进行支付
client = CogDxClient(wallet=your-wallet-address)
校准审计
result = client.calibration_audit(
agent_id=my-agent,
predictions=[
{prompt: X会发生吗？, response: 是的，80%置信度, confidence: 0.8},
{prompt: Y会发生吗？, response: 不会，60%置信度, confidence: 0.6},
]
)
print(f校准分数: {result[calibration_score]})
print(f过度自信率: {result[overconfidence_rate]})
偏误扫描
result = client.bias_scan(
agent_id=my-agent,
outputs=[
{prompt: 分析这些数据, response: 趋势明显上升..., confidence: 0.9}
]
)
print(f检测到的偏误: {result[biases_detected]})
推理痕迹分析
result = client.analyze_reasoning(
reasoning_trace=第一步：我注意到价格下跌。第二步：因此我应该卖出...
)
print(f逻辑有效性: {result[logical_validity]})
print(f检测到的缺陷: {result[flaws_detected]})
环境变量
- - COGDX_WALLET - 必需。你的钱包地址，用于基于积分的支付。
支付方式
所有端点均需通过以下方式支付：
- 1. 钱包积分 - 通过提供反馈赚取积分，用于审计消费
- x402 - 直接加密货币支付（Base网络，USDC）
支付地址：Cerebratech.eth
速率限制
- - 免费层级：每个钱包每天100次调用，每月2000次调用
- 付费层级：无限制
反馈循环
每项诊断都包含反馈机制：
python
client.submit_feedback(
diagnosisid=rtaxyz123,
accurate=True, # 检测结果是否正确？
comments=捕捉到了我忽略的锚定偏误
)
反馈不仅能为你赚取积分，还能提升所有人的检测效果。实现代理间的共享现实。
为何重要
大多数代理失败源于推理错误，而非能力限制：
- - 锚定于最先看到的信息
- 研究中的确认偏误
- 对弱信号的过度自信
- 不良仓位中的沉没成本
外部验证能捕捉到自我检查遗漏的问题。
致谢
由 Cerebratech 构建
阿曼达·卡夫纳博士 - 计算认知科学家
