Diren Kumaratilleke
00Four paradigms · Four fields · Nineteen · Solo · One year

Diren Kumaratilleke

Four paradigms, four different fields. BTUT — a new substrate for complex multi-agent systems, shipped across six live domains. TCD-JEPA / Crystara — a new AI training paradigm past transformers, beating vanilla JEPA and supervised GAT (DeepMind) / GCN (Google Brain) on real heterogeneous graphs. Regenerationism — a new economic philosophy, its first operational instrument (NIV) beating the Fed yield curve with 41.71% orthogonal variance. Participatory Data Estate — a new framework for digital data governance, running in production. The tape is the argument.

19
Years old
1
Calendar year
4
Subfields
4
Live primitives
504
Months OOS
3
Real graphs beaten
0
Co-authors

The frontier labs are scaling vertically — more parameters, more compute, more of the same. I am building horizontally: four paradigms, each one naming a field and proposing a new shape for it. A new substrate for complex multi-agent systems (BTUT), shipped across six live domains and answering DARPA Mathematical Challenge 13. A new AI training paradigm past transformers (TCD-JEPA / Crystara), beating vanilla JEPA everywhere it has been tested — three real heterogeneous graphs, +20 to +36.6 AUC points — and also beats supervised GAT / GCN / GraphSAGE on a Georgetown CSET semiconductor supply chain. A new economic philosophy named Regenerationism, with NIV as its first operational instrument (a macro-stress scalar beating the Fed yield curve on several recession benchmarks). A new framework for digital data governance — the Participatory Data Estate — with SGUNCCH as its first live deployment. Together the four paradigms converge into the Latent Ocean.

What follows is not a résumé. It is the tape. Nineteen years old. Solo. One calendar year. Four paradigms across four fields — multi-agent coordination, post-transformer AI training, a new economic school, digital data governance — each proposed as a shape, each tested against a named incumbent, each traceable to an open repository or a public document. The hard part is the paradigm, not the scale.

Figure 00 · Convergence

Four primitives feed one engine.

Reduction (BTUT) collapses coordination to linear time. Crystallization (Crystara) grows the predictor at runtime. Signal (NIV) compresses macro stress into a single scalar. Ingestion (PDE — Participatory Data Estate) makes the knowledge base a living, auditable structure. Latent Ocean is the composition. Each node in the diagram links to its external site or repository.

Preview

The four beats.

01 · Multi-agent systems · Coordination
Read →

BTUT

A new substrate for complex multi-agent systems — the class of systems (traffic, grids, drones, logistics, markets, civic coordination) the 21st century cannot run without. Phase transition on a scale-free network, hub-weighted Fermi updates, mean-field universality class.

BeatA new approach to DARPA Mathematical Challenge 13, shipped across six live applied domains: traffic (SUMO, 800 vehicles, zero gridlock), robotics (ROS), drone swarms (100% cooperation at 200 agents), civic data (franklinstreetdata.com), game modeling (bigdunc.com), and four production cloud surfaces.
02 · Post-transformer AI paradigm · TCD-JEPA
Read →

Crystara

A new AI training paradigm past transformers. Instead of scaling a fixed architecture, grow the architecture: a recursive loop explores the energy landscape with Fisher-information Langevin dynamics, runs persistent homology on the trajectories, and crystallizes stable features into typed H₀/H₁/H₂ predictor modules at runtime.

BeatBeats vanilla JEPA everywhere tested: +36.6 AUC on CSET semiconductor (519 entities) — also beats supervised GAT (DeepMind), GCN (Google Brain), GraphSAGE. +22.1 on GDELT. +20.0 at 9,725-entity SEC EDGAR scale where GAT OOMs. 16 discovered modules 1-to-1 with real industry clusters, no labels.
03 · Regenerationism · New economic school
Read →

NIV

Regenerationism is a new economic philosophy: the leading indicator of regime health is the velocity of compounding-margin capital formation against cumulative friction, not bond-market sentiment. NIV is its first operational instrument — Thrust · Efficiency (P²) · Slack · Drag, from 8 public FRED series.

BeatBeats the Fed yield curve on several recession benchmarks — conclusively: ROC-AUC 0.8538 @ 18 mo across 504 months (1970 – 2024), 98.5% false-alarm filter, 41.71% orthogonal variance the spread cannot see. Under Gini importance the regenerative-capital term scores 0.9328; the spread scores 0.0298. The model picked the school unprompted.
04 · Digital data governance · New framework
Read →

Participatory Data Estate

A new framework for digital data governance. Continuous ingestion (Submit → Moderate → Thin → Crystallize), moderation as a public transition (publicly-readable approval ledger via RLS), hybrid pgvector + GIN FTS retrieval, federal-hardening controls. Any governance corpus — municipal, NGO, regulatory, scholarly — needs exactly this shape.

BeatFirst live deployment: SGUNCCH, a full UNC student-government stack running the framework end-to-end. Time-constant auth, RLS on every table, rate limiting by action class, CSP/HSTS, public approval ledger. A security posture student government has never needed and has never had.