Ambient Structure Discovery
A self-organizing computational architecture that detects structural anomalies in financial networks without labeled training data.
In March 2025, the US Corporate Transparency Act was gutted—all domestic companies exempted from beneficial ownership reporting. Banks can no longer look up who owns an entity in a federal database. They have to figure it out themselves.
Current tools—databases, keyword matching, rule-based screening—search for what is declared. They require a query. But the query assumption is precisely what second-order ignorance violates.
You can't search for what you don't know to look for.
That's second-order ignorance.
A stigmergic mesh architecture grounded in Adaptive Resonance Theory. No labels. No predefined rules. No queries. The mesh discovers structure by processing signals it has never been trained on.
Loss of independent evaluation modes in the mesh's correlation matrix. When disparate entities converge on identical structural patterns—same layering, same nominees, same jurisdictions—the eigenvalues collapse. A single intermediary mass-producing shell companies becomes mathematically visible.
Measurable shift from direct to hedged language. Entities engaged in concealment utilize overly broad, non-specific descriptions. "General consulting" and "international trade" replace concrete business purposes. The compression is quantifiable.
Discussion frequency decouples from resolution activity. Banks file thousands of suspicious activity reports while continuing to process billions in illicit transfers. The filing becomes the substitute for action. The mesh detects this ratio mathematically.
Raw financial data overwhelms the mesh with identity noise—names, addresses, proper nouns. The key innovation: replace text with complement-coded continuous structural vectors bounded in [0, 1]. Seven topological coefficients encode the shape of each entity's behavior, not its identity. The mesh then self-organizes by structural geometry.
Three controlled experiments on the Elliptic++ Bitcoin dataset. Same mesh architecture. Same data. Different signal construction. Only complement-coded continuous vectors produce genuine structural differentiation.
| Metric | Text-Based Negative Control |
Structural Vectors Complement-Coded |
Improvement |
|---|---|---|---|
| Signal Dimensionality | ~30 discrete terms | 14-dim continuous [0,1] | — |
| Workers (self-organized) | 12 → 60 | 12 → 60 | Same expansion |
| Worker Differentiation | All identical vocabulary | Genuinely differentiated | Qualitative shift |
| Anomalies Detected | 1 | 59 | 59× |
| Licit Enrichment in Anomalies | — | 2.2× (p < 0.00002) | Statistically significant |
| Mean Familiarity Score | 0.87 (saturated) | 0.19–0.36 (differentiating) | Real discrimination |
This is not a failure—it is a genuine structural discovery. In the Bitcoin network, illicit-compatible activity is the norm. The labeled “licit” entities—exchanges, miners, wallet providers—are the ones with distinctive topology. The mesh found this without being told what to look for.
ASD does not detect “good” versus “bad.” It detects structural categories. The interpretation must be domain-aware.
Seeking partnerships with data exchanges and insight platforms that aggregate licensed financial transaction data. The architecture is built. The methodology is validated on public datasets. The next step is institutional-scale validation on real-world transaction flows—credit card panels, wire transfers, SWIFT messages, and correspondent banking networks. If you operate a data marketplace or insights exchange that serves financial institutions, this is a new intelligence layer for your platform.
The full theoretical framework is detailed in the ASD paper: Ambient Structure Discovery via Stigmergic Mesh (PDF).
Stable category formation with vigilance-gated plasticity. New patterns create new categories; familiar patterns reinforce existing ones. Solves the stability-plasticity dilemma.
Carpenter & GrossbergCoordination through shared environment rather than direct communication. Agents leave traces; other agents respond to those traces. No central coordinator needed.
Grassé; Theraulaz & Bonabeau, 1999Information degradation under strategic communication. When sender bias exceeds a threshold, communication collapses. Used to detect organizational signal decay.
Crawford & SobelAnomaly detection via Laplacian eigenvalue distribution. Structural changes surface as spectral perturbations before they become visible in metrics.
ChungStop-on-first-accept routing. Bounded rationality over exhaustive optimization. Fast, stable, and biologically plausible.
SimonInventor & Patent Holder
Ambient Structure Discovery
Seeking data partnerships for institutional-scale validation.
If you operate a financial data exchange, insights platform, or alternative data marketplace—we want to talk. Your data is the signal. Our mesh is the intelligence layer. Together, we surface what no existing compliance tool can find: the structural anomalies that nobody knows to look for.
For financial institutions: if you're facing the post-CTA reality of building internal UBO discovery capabilities, the mesh integrates with your existing transaction data infrastructure. Consumption-based, no retraining required, no labels needed.
Start a Conversation