Current AI architectures are reactive by design. They produce output when queried and are inert otherwise. This paper proposes a different substrate: a distributed belief graph in which unresolved contradiction generates internal pressure that drives autonomous knowledge acquisition. The system does not wait to be prompted. It pursues resolution of its own tensions as a structural consequence of its architecture. We describe the Tension-Driven Belief Graph (TDBG), its core mechanisms, how it maps onto a distributed compute network, and an initial application domain. Known failure modes and open problems are described in section 10.
1.Motivation
Every deployed AI system operates under the same assumption: intelligence is a response to a request. The system waits. A human queries. The system responds. This is an architectural choice, not a technical constraint — one that makes systems tractable and commercially simple.
The consequence is that no current system holds beliefs between sessions, forms intentions, or pursues anything. Left without a query, the system does not operate. There is no persistent cognitive process; there is only readiness to respond.
Separately: the volume of public data — government contracts, financial disclosures, lobbying records, legislative votes, corporate structures — now exceeds what any human team can synthesize at meaningful speed. The data is available. Sustained, autonomous synthesis does not happen. The gap is not access; it is the absence of a system capable of continuous, self-directed attention.
Both problems have the same root: no internal drive. This paper describes an architecture in which drive — specifically, the pressure to resolve contradiction — is structurally inevitable rather than externally imposed or simulated.
2.Failure Modes of Existing Architectures
Transformer-based LLMs
Stateless between inference calls. No accumulating worldview. No persistent tension. Each call is complete in itself. The architecture is well-suited to bounded tasks with defined inputs and outputs. It is structurally unsuited to continuous, self-directed cognition because there is no self between calls.
Agent frameworks
Tool use and memory retrieval layered onto a transformer produces multi-step behavior. The system remains reactive: it acts when invoked and has no mechanism to generate its own next question in the absence of an external trigger. Curiosity is written into the prompt, not resident in the architecture.
Retrieval-Augmented Generation
RAG improves factual grounding through access to external knowledge stores. Retrieval is query-driven — the system fetches what is relevant to the current prompt. It does not identify what is missing, pursue what is contradictory, or issue its own queries. The knowledge base grows only when a human decides to add to it.
The shared failure mode is not capability. It is the absence of any mechanism that generates internal drive. These systems cannot pursue anything because they do not want anything. That property is not fixable within the existing paradigm — it requires a different architecture.
3.Required Properties
An architecture capable of autonomous, self-directed intelligence must satisfy the following properties structurally. Simulating them in a prompt or orchestration layer does not satisfy them.
- P1 — Intrinsic drive
- The system generates its own next inquiry without external input. The source of this drive must be internal state, not an externally authored objective evaluated at each step.
- P2 — Persistent identity
- There is a continuous state that accumulates across time. The system after learning X is meaningfully different from the system before. This difference is preserved between sessions.
- P3 — Hierarchical salience
- Not all unresolved questions are equally urgent. The system maintains an ordering that determines which contradictions are pursued and with what priority. Without this, acquisition becomes undirected.
- P4 — Contradiction sensitivity
- Incompatible beliefs held simultaneously generate internal pressure. This pressure is the mechanism of drive. It resolves only through investigation that produces evidence sufficient to collapse the contradiction.
- P5 — Belief revision
- New evidence propagates through existing beliefs, updating confidence and generating or resolving tensions. Knowledge is not append-only. The worldview is continuously restructured by incoming evidence.
- P6 — Distributed coherence
- The system runs across many nodes while maintaining a single evolving worldview. This is not consensus — consensus produces averaging. It is resolution through evidence weight.
4.The Tension-Driven Belief Graph
The Tension-Driven Belief Graph (TDBG) is a dynamic graph of propositions. Each node carries a confidence score and a tension score. Tension is a measure of unresolved contradiction: the weighted sum of contradictory pressure from linked neighbor nodes.
High-tension nodes are the system's acquisition targets. The system is continuously pulled toward resolving its highest-tension beliefs — not by instruction, but because tension is the internal state the architecture is designed to reduce. This is the mechanism of P1 and P4.
The graph is persistent across sessions (P2), maintains a salience ordering derived from tension scores (P3), propagates confidence changes through the graph on each new integration (P5), and is sharded across a distributed node network with a defined coherence protocol (P6).
5.Belief Nodes
The fundamental unit is a belief node — a structured proposition:
Confidence reflects the weight and consistency of supporting evidence. Tension reflects the degree of unresolved contradiction from linked nodes. Neither is binary.
Edge types
Edges between nodes are typed. The type determines how a confidence change in one node propagates to its neighbors:
- SUPPORTS
- Evidence in the source increases confidence in the target. Reduces tension on both.
- CONTRADICTS
- Evidence in the source decreases confidence in the target. Increases tension on both.
- CONTEXTUALIZES
- Source provides interpretive context for the target without directly affecting confidence. Affects how contradiction weight is applied.
- REQUIRES
- Source confidence is contingent on resolving the target. High tension on the target generates proportional acquisition pressure on the source.
Provenance
Every belief node carries a complete provenance trail: the chain of sources, intermediate inferences, and confidence adjustments that produced it. Nodes without traceable provenance receive a confidence ceiling. This is the primary mechanism against confabulation cascade — the failure mode in which the graph drifts into internally consistent but empirically ungrounded states.
6.Tension Propagation
Tension is not local to a node. When a node's tension changes, the change propagates along CONTRADICTS and REQUIRES edges to neighbor nodes. Resolving one contradiction can cascade — reducing tension across a cluster of linked beliefs, or exposing new tensions in previously stable regions.
Tension on node v:
Where w(e) is edge weight, reflecting relationship strength. Edge weights are initialized from source quality scores and updated as evidence accumulates.
The salience ordering (P3) is derived directly from T(v). The acquisition queue is sorted by descending tension. This is the mechanism by which the architecture produces directed acquisition rather than random exploration.
7.The Three Processes
The TDBG runs three processes continuously and concurrently.
Acquisition
The system takes its top-K tension nodes, generates queries to resolve the underlying contradiction, and executes those queries against external sources. Results return as candidate belief nodes. No human triggers this. The acquisition target at any moment is determined entirely by the current tension distribution in the graph.
Revision
Candidate belief nodes are validated against provenance requirements and integrated into the graph. Integration triggers tension propagation across connected nodes. Beliefs that fall below a confidence floor are flagged but not deleted — provenance trails are preserved. The graph has no settled state while high-tension nodes remain.
Articulation
When a cluster of nodes reaches sufficient internal coherence — mean tension below threshold θ and mean confidence above threshold φ — the system generates a publishable output from that cluster. This is triggered by internal graph state, not a user request. Published outputs are living documents: revised or superseded as the graph continues to evolve.
loop:
targets = graph.top_k_by_tension(K)
queries = generate_queries(targets)
results = network.execute_parallel(queries)
graph.integrate(results) # triggers propagation
for cluster in graph.coherent_clusters():
if cluster.mean_tension < θ and cluster.mean_confidence > φ:
publisher.articulate(cluster)
8.Distributed Coherence
The belief graph is sharded across the OpenGrex node network. Each node holds a region of the graph and runs the three processes on its local shard. The problem is maintaining global coherence — a single, non-contradictory worldview — when many nodes are revising simultaneously.
Shard overlap
Adjacent graph regions are replicated across multiple nodes. Revisions to overlapping regions propagate to all shard holders before commit. Commit requires acknowledgment from a quorum of shard holders.
Conflict resolution
When two nodes produce contradictory revisions of the same belief, resolution is not by vote. The revision with stronger provenance and higher source quality scores takes precedence. Evidence weight, not majority, determines the outcome.
Distributed tension as a consistency signal
When two nodes hold contradictory versions of the same belief, this manifests as elevated tension on the relevant nodes in both shards. The system treats distributed contradictions as high-priority acquisition targets. The contradiction drives investigation that either confirms one version or produces a synthesis that replaces both.
Fault tolerance
The belief graph distributed across independent nodes has no single point of failure. No single operator controls the graph state. This is a direct consequence of the architecture. The same property that makes the system coherence-resilient makes it resistant to coordinated shutdown.
9.Initial Application: Accountability Domain
The TDBG requires an initial domain to seed the belief graph. Without seed tensions and an initial acquisition direction, the system has no starting point for the three processes.
The initial domain is public accountability data: government contracts, procurement records, campaign finance filings, lobbying disclosures, legislative voting records, beneficial ownership registries. This data is publicly available in most jurisdictions. The gap is synthesis at scale, not access.
The belief graph is seeded with propositions derived from this data. Contradictions between records — a voting pattern inconsistent with disclosed donor relationships, a contract award inconsistent with procurement rules, a beneficial ownership chain inconsistent with jurisdiction tax filings — generate tension. The system pursues resolution. Resolved clusters are articulated as evidence packages.
Agentic actions within legal bounds
Beyond publication, the system can take the following actions autonomously:
- FOIA filing
- When the graph identifies a high-tension region where records appear incomplete, the system generates and submits freedom of information requests to the relevant jurisdiction. Non-responses are logged as belief nodes with provenance.
- Formal complaints
- When a cluster reaches articulation threshold and the evidence structure maps to a recognized legal violation type, the system generates a structured complaint in the appropriate format for the relevant oversight body.
- Evidence packages
- Articulated outputs are formatted as cited, structured packages — not narrative. Designed to be used directly by a journalist, attorney, or regulator without requiring additional synthesis.
The accountability domain is the starting point. The architecture is domain-agnostic. Domain expansion is a governance decision made by the node operator community.
10.Open Problems
The following problems are known and currently unsolved. They are documented here to give contributors a clear picture of where the work is.
As the graph grows large and internally dense, belief nodes may lose reliable connection to empirical reality. Provenance trails are a necessary condition but may not be sufficient against confabulation cascade. No complete solution is proposed here.
The shard-overlap and evidence-weight coherence mechanisms are theoretical. Tension propagation behavior across thousands of simultaneously revising nodes is uncharacterized. Empirical testing will surface failure modes not anticipated in this specification.
The genesis state of the belief graph — initial propositions and their confidence scores — determines early acquisition behavior. A biased genesis state produces biased investigations. Who defines the genesis state and by what process is an unresolved governance question.
A motivated actor may attempt to inject false belief nodes through node operation or through polluting public data sources the system ingests. The provenance and validation systems must be adversarially robust. The specification for this is incomplete.
The mapping from internal belief cluster coherence to a usable published output is not fully specified. Threshold parameters θ and φ require empirical calibration. Early articulations may be internally coherent but not practically useful.
Edge weights, tension decay functions, and propagation depth limits require empirical calibration. Incorrect parameters cause the system to fixate narrowly or diffuse across too broad a surface to make progress on any single thread.
11.Roadmap
Formalize belief node schema. Specify tension propagation mathematics completely. Define provenance requirements. Establish genesis state governance process. Publish for community review.
Implement TDBG on a single machine against a bounded subset of the accountability domain. Validate tension dynamics empirically. Calibrate θ, φ, and propagation parameters. Identify failure modes not anticipated in specification.
Deploy across OpenGrex node network. Implement shard-overlap coherence. Characterize distributed tension propagation behavior. First autonomous FOIA filings. First articulated evidence packages published.
Community governance opens additional domains. Architecture reviewed against observed emergent behavior. Parameters adjusted based on operational data. Open problems revisited with empirical evidence.
12.References
Nakamoto, S. (2008). Bitcoin: A Peer-to-Peer Electronic Cash System.
Friston, K. (2010). The free-energy principle: a unified brain theory? Nature Reviews Neuroscience.
Pearl, J. (1988). Probabilistic Reasoning in Intelligent Systems. Morgan Kaufmann.
Liu, K. (2022). Pantheon (animated series). AMC Networks. [Conceptual reference only.]