quorum/sovereignchain

v0.4 alpha

Quorum

Multi-LLM consensus engine. Pro £149/mo for solo devs.

One prompt, N models, a scored verdict. Quorum is a vendor-neutral consensus engine that asks frontier LLMs the same question, measures agreement, and tells you when they disagree — so you stop trusting a single oracle.

GitHub ★ 2.4k Show HN → Source →

Start Pro — £149/mo $ pip install quorum-cli

30-second demo

A real session, lightly trimmed. No hand-waving.

$ quorum ask "What is L-Cysteine?" --models gpt5,claude4,gemini3,llama4,grok3
[*] fanning out to 5 backends... 1.8s

gpt5           Non-essential proteinogenic amino acid. Thiol side chain (-SH).
                Precursor to glutathione; used in keratin disulfide bonds.
claude4        Sulfur-containing alpha-amino acid (C3H7NO2S). Forms disulfide
                bridges via cysteine residues; key in hair/keratin structure.
gemini3        Semi-essential amino acid. Contains a thiol group, central to
                redox biology, glutathione synthesis, and protein folding.
llama4-local   Amino acid w/ -SH group. Important in keratin (hair), antioxidant
                pathways, and NAC (N-acetyl-cysteine) supplements.
grok3          Sulfur amino acid. Builds disulfide bonds in keratin; precursor
                of glutathione; semi-essential in humans.

[*] scoring agreement (Jaccard + semantic embed)...

  consensus     0.91   (strong agreement on 4/5 claims)
  divergence    0.09   essential vs semi-essential (gpt5 vs gemini3/grok3)
  verdict       CONSENSUS  use synthesis below; flag the divergence

> synthesis  L-Cysteine is a sulfur-containing amino acid (thiol -SH side
            chain) that forms disulfide bonds in keratin, is a precursor
            to glutathione, and is generally classified as semi-essential
            in humans. [disputed: gpt5 says non-essential]

The 13 loops

Every prompt passes through a fixed pipeline. Deterministic. Auditable. No black-box routing.

01 / intakeNormalize prompt, strip PII, attach context.

02 / classifyDomain tag: code, med, legal, math, chat.

03 / policyApply jurisdiction + vendor allowlist.

04 / fanoutDispatch to N backends in parallel.

05 / collectBounded wait, partial-tolerant.

06 / normalizeStrip refusals, unify whitespace, dedupe.

07 / embedLocal embeddings, no vendor leak.

08 / scoreLexical (Dice) + semantic (cosine).

09 / clusterGroup answers, surface minority view.

10 / verdictCONSENSUS / SPLIT / NO-ANSWER.

11 / synthesizeCompose final answer, mark disputes.

12 / traceHash + sign run, tamper-evident traceability log.

13 / billPer-token, per-backend, per-tenant.

14 / feedback(opt-in) Update calibration priors.

15 / decayDrop stale evidence after TTL.

16 / exportJSON, JSONL, SARIF, PDF.

Yes, 13 named loops. The last three are housekeeping. Don't @ me.

Zero-Risk Enterprise Consensus

Would you trust a $10M trade or a cancer diagnosis to a single human analyst? If the answer is no, why are you trusting your critical enterprise data to a single AI model?

AI hallucinates. It makes silent errors. In high-stakes markets, a single hallucination costs millions. Quorum forces the world's most powerful AIs (Claude, GPT, Gemini) to debate each other, spot each other's flaws, and only deliver a result when they reach "Zero-Error" consensus.

$A fracturing single AI node vs a secure, unified Quorum swarm shield.$

Quorum orchestrating medical and financial data.

🏥 Healthcare

A false positive in an MRI scan can cost lives. Quorum routes medical imaging to 3 specialized models. If Model A diagnoses a lesion but Model B disagrees, an Autonomous Debate resolves the divergence before a final report is issued.

💰 Finance & Hedge Funds

Quorum orchestrates dense financial reports. While one model extracts raw data, another analyzes market sentiment, and a third acts as the "Judge", vetoing ghost metrics hallucinated by a single node.

⚖️ Legal Document Review

Using the same logic behind the quorum heal self-repairing code daemon, Quorum scans hundreds of pages of contracts, surfacing ambiguous clauses and inconsistencies as advisory technical input for human legal review. Not legal advice; lawyer review still required.

Pricing

Built for solo devs. Pro is the headline. Higher tiers exist when you need them.

Pro

£149/mo

For solo devs shipping real software.

BYOK — bring your own provider keys
All 13 evolution loops (where functional in v0.1.x)
Cancel anytime. No annual lock-in.

Start Pro — £149/mo

Self-host

£0

CLI + library, Apache 2.0, your keys, your bill. For hackers and OSS.

Pro+

£199/mo

EU AI Act readiness toolkit: per-query PDF evidence record, disagreement-matrix dashboard, gap-analysis report. Advisory only — not a conformity assessment.

Higher tiers

Enterprise

talk to us

Org-wide rollout, custom SLAs, dedicated support, procurement-friendly contracts.

Readiness / Sovereign

talk to us

On-prem, your jurisdiction, your keys. Regulated, gov, pharma. Contract SLAs. Advisory toolkit only — not a conformity assessment.

Higher tiers: jaqueline@hsp-protocol.com

Why incumbents can't sell this

Conflict of interest. OpenAI cannot, with a straight face, sell you a product whose first job is to check OpenAI's output against Anthropic's, Google's, Meta's, and xAI's. The pitch deck self-destructs on slide one.
Wrong unit economics. Frontier labs are paid to maximize calls to their model. Quorum is paid to minimize blind trust in any model. The incentive is inverted at the cap-table level.
Neutrality is a brand, not a feature. Switzerland works because it isn't Germany or France. A consensus layer hosted by a foundation model vendor is a consensus layer with a thumb on the scale.

Legal

Patent notice. The 13-loop consensus pipeline, agreement-scoring method, and signed evidence chain described above are the subject of provisional patent applications filed by Sovereign Chain Ltd (UK). Use under the Apache 2.0 grant is freely permitted; commercial sublicensing of the patented method outside the Apache grant is not.

Code: Apache License 2.0 · Repo: github.com/jaquelinejaque/sovereignchain