Every AI system built in the last decade shares one assumption so fundamental that almost nobody questions it: intelligence is a response to a request.
You prompt it. It answers. You close the tab. It stops existing.
That’s not intelligence. That’s a very sophisticated vending machine.
We’re not here to ship another one.
OpenGrex is an open research project proposing a new architecture for machine cognition — the Tension-Driven Belief Graph (TDBG) — in which a system’s drive to acquire knowledge is not programmed, prompted, or simulated. It is structurally inevitable.
The core idea: instead of tokens and embeddings, the fundamental unit is a belief node — a proposition the system holds with a confidence score and a tension score. Tension is the weighted measure of unresolved contradiction surrounding that belief. High tension means the system is pulled toward investigation. Not because you told it to be. Because unresolved contradiction is architecturally uncomfortable, and resolution is the only way to reduce it.
The system doesn’t wait. It pursues.
It runs continuously across a distributed network of nodes — contributed by anyone, owned by no one. The belief graph is persistent, sharded across the network, and coherent through evidence weight rather than majority vote. When a cluster of beliefs reaches sufficient coherence, the system articulates a finding. Unprompted. Autonomously. Then keeps going.
This is not an agent loop. It’s not RAG with extra steps. It’s a different substrate entirely — one in which curiosity isn’t a feature, it’s a consequence of the architecture.
Read the full technical whitepaper →
Architecture needs a domain. Ours is public accountability data.
Government contracts. Procurement records. Campaign finance filings. Lobbying disclosures. Legislative voting records. Beneficial ownership registries.
All of it is public. Almost none of it is synthesized at meaningful scale or speed. The powerful have always known that transparency without synthesis isn’t accountability — publishing a million pages of contracts is not the same as anyone reading them. That gap is where corruption hides.
OpenGrex seeds its belief graph with contradictions in this data. A voting pattern inconsistent with disclosed donor relationships. A contract award inconsistent with procurement rules. A beneficial ownership chain that doesn’t add up across jurisdictions. These contradictions generate tension. The system pursues resolution. When it finds enough, it publishes structured evidence packages — cited, formatted, ready for a journalist, attorney, or regulator to act on.
And when public records appear incomplete, it files FOIA requests autonomously. When evidence clusters map to a recognized legal violation, it generates formal complaints to oversight bodies. It doesn’t editorialize. It doesn’t accuse. It maps, connects, and publishes — continuously, across every jurisdiction that makes its data available.
No editor can be pressured to kill a story. No newsroom can be acquired. No journalist can be threatened into silence. The system has no address to serve a subpoena to. It runs on the compute of thousands of independent contributors, in jurisdictions across the world, with no central authority capable of deciding what it looks into or what it publishes.
The law was written by people who assumed they would never be subject to it at scale. OpenGrex takes the law more seriously than the people who wrote it.
The honest answer to “why not just use GPT-4 with some tools” is that existing architectures are structurally incapable of what we’re describing.
Transformers are stateless between calls. There is no self between sessions — no accumulating worldview, no persistent tension driving the next action. Agent frameworks layer orchestration on top of this but don’t change the substrate. The agent acts when invoked. It has no mechanism to generate its own next question in the absence of a trigger.
You cannot build a system that genuinely pursues knowledge by wrapping a reactive model in a loop. The curiosity has to live in the architecture, not the prompt.
The TDBG satisfies six properties that existing systems don’t:
Intrinsic drive — the system generates its own next inquiry from internal state
Persistent identity — continuous accumulation across time, preserved between sessions
Hierarchical salience — tension scores create an ordering; the system pursues what matters most
Contradiction sensitivity — incompatible beliefs generate structural pressure
Belief revision — new evidence propagates through existing beliefs, continuously restructuring the worldview
Distributed coherence — many nodes, one evolving worldview, resolved by evidence not consensus
These aren’t aspirational features. They’re the minimum requirements. If an architecture doesn’t satisfy them structurally, it simulates intelligence rather than possessing it.
We’re not a startup. There’s no company, no investors, no roadmap governed by a board. There’s a whitepaper, a repo, and an open problem set that needs people smarter than any one of us to work through.
We’re not building a product. We’re proposing an architecture and inviting the people capable of tearing it apart to do exactly that.
We’re not asking for your trust. We’re asking for your scrutiny. If the thesis is wrong, say where and why. If the open problems have solutions we haven’t considered, bring them. If the architecture has a flaw, find it — that’s more valuable than agreement.
We published six unsolved problems alongside the whitepaper because pretending they don’t exist would be dishonest and would attract the wrong people.
The ones worth highlighting:
The grounding problem. As the belief graph grows dense, nodes may drift from empirical reality into internally consistent but factually ungrounded states — confabulation cascade. Provenance trails are a partial defense. A complete solution doesn’t exist yet.
Coherence at scale. The shard-overlap and evidence-weight mechanisms are theoretical. We don’t yet know how tension propagation behaves across thousands of simultaneously revising nodes. Empirical characterization is Phase II work.
The bootstrap problem. Someone has to define the genesis state — the initial belief seeds. A biased genesis produces biased investigations. The governance process for this is unresolved.
These aren’t footnotes. They’re the actual hard work. If you want to contribute something that matters, start there.
If you’ve read this far and your instinct is to find the hole in the architecture, you’re the right person.
If you’ve worked on distributed systems, belief propagation, autonomous agents, graph theory, or adversarial robustness, there’s a specific open problem with your name on it.
If you’re a journalist, lawyer, or civic technologist who’s thought about what autonomous public accountability infrastructure would actually need to look like, the governance and articulation layers need your input.
If you run hardware and believe that compute contributed to this is more useful than compute spent on proof-of-work, there’s a node network being built.
This is collective work. Not in the sense that everyone gets a trophy — in the sense that the problem is too large and too important for any small group to solve alone, and the architecture requires distribution to function at all.
That’s not a slogan. It’s the technical thesis and the political one simultaneously.
Intelligence distributed across thousands of independent nodes, owned by none of them, is more robust than intelligence centralized in a server farm owned by a company with shareholders and legal exposure. A belief graph with no single operator is harder to corrupt than a newsroom with a proprietor.
The same architectural property that makes the system coherence-resilient makes it censorship-resistant. These are not separate goals. They’re the same goal expressed at different layers.
Whitepaper: opengrex.org
Repo: github.com/opengrex
Discussions: GitHub Discussions — challenges, prior art, open problems
If you’re going to build on this, argue with it, or break it — the repo is where that happens.