The tool that won't let AI say anything it can't cite

What it is

Grainulator is a research-sprint orchestrator for Claude Code that aims to make AI outputs auditable and decision-ready. It has been reported that every finding is stored as a typed claim, adversarially challenged, confidence-graded, and compiled into a self-contained brief — with evidence tiers that range from “stated” to “production.” There’s an interactive demo at grainulator.app that runs a small in-browser model so you can try the flow without installing anything.

How it works

Claims are the unit of knowledge. Each claim gets a type (factual, estimate, risk, recommendation, etc.) and an evidence tier; a seven‑pass compiler checks coverage, strength, conflicts, and bias, then produces a confidence score. The repo says the system will block output when conflicts remain unresolved — allegedly preventing the model from spitting out unsupported conclusions until humans sort it out. Want a hands‑on example? Tell Claude “research how our auth system works” and it will run a multi‑pass sprint, or “challenge r003” to adversarially test a specific claim.

Why it matters

Tired of hallucinations? You’re not alone. This tool is squarely aimed at the trust problem: governance, traceability, and a paper trail for each assertion. In an era where businesses need explainable results, Grainulator tries to marry old‑school citation discipline with new‑school LLM workflows. The emotional beat here is obvious — relief. Imagine handing a brief to a stakeholder and being able to point to a structured, graded chain of evidence instead of vague prose. Nice, right?

Try it (if you want)

Installation for Claude Code is straightforward: add the plugin marketplace URL, then install the grainulator plugin — Node.js >= 20 is required for some helper servers. There’s also an autonomous subagent that can run sprints for you; it has been reported that the agent reads compiler output to decide the next step until a sprint reaches decision‑ready confidence. If you just want to poke around, the PWA demo downloads a ~200MB SmolLM2 model for local inference and demonstrates the compile flow and progressive claim disclosure.

Sources: github.com/grainulation, Hacker News