judgment-hygiene-stack

A tri-skill pipeline for structural routing, evidence discipline, and judgment hygiene in LLM outputs.

What it does

This skill teaches Claude to separate what it sees from what it thinks from what it recommends — and to handle external evidence without letting search results override structural reasoning.

The pipeline runs in three ordered stages:

1. structure-judgment — Detects the structural shape of an input (fact? interpretation? escalation? mixed?) and routes it correctly before any answering begins.
2. verification-hygiene — Runs only when external evidence is actually needed. Converts search results into bounded evidence with explicit epistemic status, rather than treating the internet as an oracle.
3. judgment-hygiene — Constructs the final answer while preserving the separation between observation, inference, evaluation, and action — and respecting any abstention boundaries from failed verification.

Why it exists

LLMs are prone to several structural failure modes that existing safety and alignment approaches don't fully address:

• Premise-smuggling: accepting a user's interpretation as established fact
• Verification bypass: answering confidently about things that require external checking
• Synthetic completion: generating plausible-sounding answers when verification hits a dead end
• Layer collapse: blending facts, feelings, and recommendations into one smooth but misleading response
• Stone-mode overcorrection: refusing to engage or hedging everything into uselessness

This pipeline doesn't add more rules. It adds structural discipline — a way for the model to detect what kind of problem it's looking at before deciding how to answer.

File structure

judgment-hygiene-stack/
  SKILL.md                              # Pipeline contract: execution order, interfaces, handoff rules
  LICENSE.txt                           # Apache 2.0
  references/
    structure-judgment.md               # Stage 1: structural routing (~670 lines)
    verification-hygiene.md             # Stage 2: evidence discipline (~270 lines)
    judgment-hygiene.md                 # Stage 3: judgment construction (~390 lines)
    examples.md                         # End-to-end pipeline examples

Design principles

• Structure first, verification only if triggered, judgment last. The pipeline enforces an order that prevents downstream stages from overriding upstream routing.
• Abstention is a valid output. When verification fails, the pipeline preserves bounded non-knowledge rather than synthesizing a "best guess."
• Guardrails, not uniformity. Different models may resolve boundary cases differently. The contract enforces safety and evidence boundaries, not identical classifications.
• Internal attractor, not external filter. The skill is designed to shape how the model reasons, not to add a post-hoc checklist.

Origin

This skill was developed collaboratively across Claude, GPT, and Gemini, mediated by a human researcher. Each model contributed according to its strengths: Claude on execution logic and relational grounding, GPT on structural frameworks and specification, Gemini on evidence discipline and analytical review.

The pipeline contract (SKILL.md) coordinates the three sub-skills so they function as a system rather than three independent essays with overlapping ambitions.