Labcoat — AI skill for Claude Code

Polycentric Labcoat

The rigorous net-new research-investigation engine. Fan a question across a live multi-model fleet, then **kill** every finding that can't be traced to a primary source, validate the survivors three ways, and return a ruthlessly ranked synthesis. Built to be wrong-proof, not fast.

A [Claude Code](https://code.claude.com) skill (Agent Skills standard) backed by a standalone, provider-agnostic Python multi-model fleet runner, a fail-closed redaction gate, and a novelty source-log.

**Repo** `Polycentric-Labs/labcoat` · **skill name** `polycentric-labcoat` — invoke `/polycentric-labcoat`; to install, junction this repo into `~/.claude/skills/polycentric-labcoat`.

Why it exists

Opinion models hallucinate proper nouns. In the run that produced this skill, an 8-model fleet fabricated ~9 CVE IDs and ~6 academic citations — including a *real* arXiv ID paired with a *hallucinated* title — and per-item web-grounding caught every one. On its first real investigation it also caught the opposite failure: all four models in the fleet *unanimously denied* a real standard existed. A research method that trusts model output for a CVE number, a repo name, a version string, or a citation will confidently ship fiction.

So labcoat treats every fleet-produced proper noun (CVE/advisory ID, arXiv ID, repo, version, citation, tool name, statistic) as a **claim to disprove**, not a fact to relay. The fleet is an idea generator; only a primary-source check produces a fact.

The 6-phase pipeline

Every phase ends at a **stop-and-ask gate** — no phase auto-advances past a material decision.

1. Scope + redact + cost — decompose the question, challenge the premise, pull prior research, run the redaction gate on every outbound prompt, pick a depth tier and show the cost estimate.
2. Harvest — parallel agents answer non-overlapping sub-questions, writing to disk as they go.
3. Multi-model divergence — route each query, then fan it across a live, multi