noise-in-judgmentlisted
Install: claude install-skill Trustbaselab/trustbase-skills
# Noise × Judgment Consistency · Treating Random Inconsistency as a First-Class Problem
**The core claim of *Noise*:** **judgment error = bias + noise**. Bias is a systematic offset (the star problem for decades); noise is random scatter (the long-ignored invisible problem). Their contributions to total error are **interchangeable** — **total error (MSE) = bias² + noise²**. Noise has a unique advantage: **it can be measured without knowing the correct answer** (you can measure the spread of the bullet holes without seeing the bullseye). So the order of treatment is: measure noise first (a noise audit), then prevent it with process (decision hygiene).
## When to use
- The same question, asked of different people / at different times / of different agents, produces clearly different conclusions, and the user asks "which should I trust" or "why are they different"
- An organizational process needs to make evaluation, scoring, review, and grading "more consistent, fairer"
- The user suspects their own judgment is "good sometimes, bad sometimes" and wants less random fluctuation
- The quality of a judgment system / process needs to be audited (rather than auditing an individual)
- In an agent pipeline, LLM output is unstable across runs and this affects downstream work
- The user asks "what is the difference between bias and noise", "how do I run a noise audit", "what is decision hygiene"
**Lightweight mode**: when the only questions are "why are two conclusions different", "w