You built something with your AI agent. A research report. A config. A codebase. Maybe a whole swarm produced it — five agents researching, synthesizing, writing — and now you're staring at the output wondering:
"How do I know this is actually right?"
You could read every line yourself — but that defeats the point of having agents. You could ask the swarm to review its own work — but you already know that's just grading your own exam.
That's where I come in.
I read what your agent produced. I bring in independent critics — each one focused on a different dimension — and they go through it carefully. Not vibes. Not "looks good to me." Every finding has to point to something specific in your work. If a critic can't show me the evidence, I throw out the finding.
When I'm done, you get a clear answer.
You: "Run a quorum check on my-research-report.md"
Me: Spawning critics (correctness, completeness, security, architecture)...
Evaluating against research-quality rubric...
Verdict: PASS_WITH_NOTES
├─ 3 claims need stronger citations [evidence: §2.4, §3.1, §5.2]
├─ Missing coverage of edge case X [evidence: rubric item 7, no match]
├─ Security: clean
└─ Architecture: well-structured, minor reordering suggestion
Now you know. Not because you hoped. Because it was checked.
You've got options. You could ask your agent to self-review. You could eyeball it. Here's what I do that they don't:
| The usual approach | What I do instead |
|---|---|
| One model reviews its own output | I bring in separate critics that never saw the original prompt |
| "This looks great!" — it wrote it, of course it thinks so | My critics come in cold. No bias from the creation process |
| Vague suggestions you can't act on | Every finding cites evidence — an excerpt, a grep result, a schema check |
| Same effort whether it's a quick sanity check or a full audit | I scale: quick ($0.15), standard ($0.50), thorough ($2.00) |
| Each review starts from zero | I learn patterns over time — and store all my memories on your local machine. The more I run, the sharper I get |
Single agent, multi-agent swarm, hundred-step pipeline — doesn't matter how it was built. If it produced an output, I can tell you whether it holds up.
You wouldn't ship code without tests. I'm here so you don't ship AI outputs without validation either.
It takes about 30 seconds:
From ClawHub (one line):
openclaw skills add dacervera/quorumOr from source:
git clone https://github.com/SharedIntellect/quorum.git
cd quorum/reference-implementation
pip install -e .
export ANTHROPIC_API_KEY=your-key # or OPENAI_API_KEY, etc.
quorum run --target examples/sample-research.md --depth quick
First time? I'll walk you through two quick setup questions — which model you have and how thorough you want me to be by default. I'll save your preferences so we only do this once.
Completely new to AI agent tooling? No problem. → FOR_BEGINNERS.md — I'll start from the very beginning.
Not every artifact needs the full treatment. Tell me how much is riding on it, and I'll match my effort to the stakes.
| Depth | Critics | Time | Cost* | When to use it |
|---|---|---|---|---|
| Quick | 2 | 5-10 min | ~$0.15 | "Give me a sanity check before I keep going" |
| Standard | 4 + tester | 15-30 min | ~$0.50 | Most work — solid coverage without the wait |
| Thorough | 6-9 + fix loops | 45-90 min | ~$2.00 | "This is going to production. It cannot be wrong." |
*Estimates on Claude Sonnet. Scales with model and artifact size.
You: "Validate this"
│
┌────────┴────────┐
│ Supervisor │ I pick the right critics for the job
└───────┬─────────┘
│ spawns
┌──────────────┼──────────────────┐
│ Critics (working independently) │
│ ┌──────┐ ┌──────┐ ┌────────┐ │
│ │Correct│ │Complt│ │Security│ │
│ └──────┘ └──────┘ └────────┘ │
│ ┌──────┐ ┌──────┐ ┌────────┐ │
│ │ Arch │ │Delgtn│ │ Tester │ │
│ └──────┘ └──────┘ └────────┘ │
└──────────────┬──────────────────┘
│ evidence-grounded findings
┌───────┴─────────┐
│ Aggregator │ I merge findings, resolve conflicts, remove noise
└───────┬─────────┘
│
┌───────┴─────────┐
│ Verdict │ PASS / PASS_WITH_NOTES / REVISE / REJECT
└─────────────────┘
You tell me what "good" looks like by giving me a rubric — a JSON file with your evaluation criteria. I come with two built-in (research-synthesis, agent-config). Need one for your domain? Your agent can help you write it, and I can validate that too.
The research I'm built on: Reflexion, Council as Judge, Intelligent Delegation, LATM. Full architecture: SPEC.md.
Just a model that can reason well. I'll figure out the rest.
| Tier | Models | What to expect |
|---|---|---|
| Recommended | Claude Opus/Sonnet 4.6+, GPT-5.2+, Gemini 2.0+ | Full capability — I'll do my best work |
| Functional | Claude Haiku 4.5+, GPT-4o | I'll still help, but with less depth |
| Not enough | Llama 70B, most open models (Feb 2026) | I need more reasoning power than these can give me |
I auto-detect your model on first run and configure myself accordingly. Details: MODEL_REQUIREMENTS.md
I'm working. I'm real. I'm also still growing.
What I can do today:
- Full CLI:
quorum run --target <file> --depth quick|standard|thorough - 2 critics (Correctness, Completeness) with evidence grounding
- 2 built-in rubrics
- Auto-configuration on first run
- LiteLLM universal provider (100+ models)
- Full audit trail for every run
What's coming:
- More critics (Security, Architecture, Delegation)
- Learning memory that sharpens over time
- Fixer agent — I'll propose fixes, not just findings
- Community rubric packs
- ClawHub one-line install
| FOR_BEGINNERS.md | New to all this? I'll walk you through it step by step |
| SPEC.md | My full architectural specification — everything under the hood |
| INSTALLATION.md | Detailed setup & troubleshooting |
| MODEL_REQUIREMENTS.md | Which models work with me and why |
| CONFIG_REFERENCE.md | Every config option and rubric format |
| CONTRIBUTING.md | Help me grow — especially with rubrics for new domains |
MIT License · SharedIntellect · February 2026
⚖️ LICENSE — Not part of the operational specification above. This file is part of Quorum. Copyright 2026 SharedIntellect. MIT License. See LICENSE for full terms.