← ClaudeAtlas

noise-in-judgmentlisted

A judgment-consistency skill based on "Noise" (Kahneman, Sibony, Sunstein). Separate bias from noise, run a noise audit to quantify disagreement, then apply decision hygiene (independent judgment, aggregation, ranking, guidelines) to reduce variance. Includes a multi-run consistency-check protocol for agents. Use when the user says "why is the answer different every time", "which conclusion should I trust", "how do I make scoring consistent", "how do I audit decision quality", "is this much difference between two judgments normal"; not for pure fact lookup, single-answer computation, or personal taste.
Trustbaselab/trustbase-skills · ★ 1 · AI & Automation · score 74
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