Noise Audit Skill

This skill implements the "Statistical Jury" method to quantify Decision Noise. It requires the agent to simulate multiple independent judges evaluating the same material to detect variance.

Workflow

1. •Isolate the Input: Identify the specific decision, report, or text to be audited.
2. •Spawn Judges: You must launch THREE (3) separate task sub-agents (type: noise_judge).
   • •Instruction to Judges: "Evaluate the following text for Decision Quality on a scale of 1-10. Provide a brief justification."
   • •Independence: Ensure each judge runs in isolation (the task tool handles this by default).
3. •Collect Scores: Retrieve the three numerical scores (e.g., 6, 8, 7).
4. •Calculate Statistics:
   • •Average Quality Score: (Score1 + Score2 + Score3) / 3
   • •Noise Score (Variance): Calculate the standard deviation or simple range (Max - Min).
     • •Formula: sqrt( sum( (x - mean)^2 ) / N )
     • •Interpretation: High variance (> 1.5) indicates high Noise (inconsistent judgment).
5. •Output Artifact:
   • •Generate a "Noise Variance" table (Confusion Matrix style) showing the divergence.

Python Script: Variance Calculation (Reference)

import math

def calculate_noise_metrics(scores):
    n = len(scores)
    if n == 0: return 0, 0
    
    mean = sum(scores) / n
    variance = sum([(x - mean) ** 2 for x in scores]) / n
    std_dev = math.sqrt(variance)
    
    return {
        "mean": mean,
        "std_dev": std_dev,
        "range": max(scores) - min(scores)
    }