The research engine behind Goldman Forge

Building AI that learns to understand humans — not the other way around.

Goldman Global Research Labs is our open-source R&D arm. We research human-perspective, decentralized, dual-LLM-audited AI. Everything we publish is AGPL.

Star on GitHub See how it works

1 · Raw frame — what the contributor's device sees

[frame 0421 · 09:47:12 · openglasses-cam]
1920×1080 · 2.1 MB · motion-triggered
scene: bus depot, lane 4 — vehicle pass

What we’re building

Open-source research projects. Each ships code, not slideware.

● LIVE

Hive.AGI

A human-perspective knowledge symbiosis network. Capture what matters → fuse into structured entries → publish as content-addressed Seed Packages.

Python · MiniMax M3 · DeepSeek V4 Flash · IPFS

Explore the project

○ IN PLANNING

Project 2

Coming. Follow the repo for the next research line.

○ IN PLANNING

Project 3

Coming. Follow the repo for the next research line.

How it works

One pipeline. Four stages. Every step is open and inspectable.

1. CAPTURE
human
defines what matters
2. GENERATE
MiniMax M3
writes the wiki entry
3. AUDIT
DeepSeek V4
reviews & auto-corrects hallucinations
4. PUBLISH
IPFS
content-addressed exchangeable by CID

Capture — A contributor (phone, glasses, sensor, or note) records a moment they judge worth keeping. The human defines the signal.

Generate — MiniMax M3 turns raw capture into a structured wiki entry: description, analysis, related links.

Audit — DeepSeek V4 Flash independently reviews for hallucination, schema, and bias. Auto-corrects or quarantines. No model marks its own homework.

Publish — The entry becomes a Seed Package, content-addressed on IPFS. Anyone can verify integrity by recomputing the hash.

P2.5 (automatic peer-to-peer sync between contributors) is in design. Today, packages are publish / verify / resolve.

The difference

We’re building something corporate AI labs structurally can’t.

Dimension	Corporate AI labs	Goldman Global Research Labs
Data	Web-scraped, scale over signal	Human-curated, contributor defines value
Quality	Model evaluates itself	Independent generator + auditor (different providers)
Distribution	Centralized API, vendor lock-in	IPFS content addressing, peer-to-peer
License	Closed or "open-weight" with usage restrictions	AGPL-3.0 — genuine copyleft, SaaS-safe
Sovereignty	US/UK-centric	Australian-built, AU-hosted
Privacy	Policy-level, often bypassed in practice	Code-enforced PII stripping (no bypass exists)

Genuinely open

Every line of code is AGPL-3.0. If a corporation deploys our work as a service, they must share their improvements back — or buy a commercial license that funds the contributors.

Every Seed Package is CC-BY-NC-SA-4.0. Contributors keep their copyright. Non-commercial sharing is free; commercial use funds the network.

This isn’t “open-core” theater. The full pipeline — generator, auditor, IPFS exchange, Obsidian plugin — is in the repo.

Browse the repo Contributor guide

Built to be trusted

Australian-built. Sydney-based team, AU data hosting, ready for the 2026 Privacy Act.
Privacy by code. PII stripping (faces, plates) runs before any LLM call. There is no --skip-blur flag — we checked.
Real-world signal. Goldman Forge’s live client deployments across transport, travel, and finance feed Labs genuine edge cases.
Reproducible. Every commit runs CI. Mock mode lets anyone verify the pipeline without API keys.

Want to build with us?

Contributors & researchers

Star, fork, or open a PR.

github.com/CTO-goldmanglobal/gg-hiveagi

Enterprise R&D & collaboration

Talk to us about commercial licensing, custom Seed Packages, or research partnerships.

cto@goldmanglobal.com.au

Hive.AGI is research software. The audit step catches some hallucinations, not all. Validate outputs before any production use. AGPL-3.0 license · Privacy policy