Diren Kumaratilleke
IChapter

The Four Primitives

Four paradigms, not four projects. BTUT proposes a new substrate for complex multi-agent systems — the class of systems modern civilization cannot run without — and ships one primitive across six live applied domains. TCD-JEPA (Crystara) proposes a new AI training paradigm past transformers: grow the predictor at runtime via topological crystallization instead of scaling a fixed attention architecture. Regenerationism proposes a new economic school — capital with compounding margins measured against cumulative friction — and NIV is its first operational instrument, beating the Fed yield curve on several recession benchmarks. The Participatory Data Estate proposes a new framework for digital data governance — continuous ingestion, public moderation ledger, federal-hardening controls — with SGUNCCH as its first live deployment. Every paradigm below has a table, a tear sheet, or a repo directly underneath it.

Why four, and why these four.

Each primitive names a field and proposes a new shape for it. Reduction — for the mathematics of complex multi-agent systems (BTUT). Crystallization — for the training paradigm of AI (TCD-JEPA / Crystara). Signal — for the philosophical foundation of macroeconomics (Regenerationism, operationalized by NIV). Ingestion — for the architecture of digital data governance (the Participatory Data Estate). Four fields, four paradigm proposals, each with working code and measured results.

BTUT — a new substrate for complex multi-agent systems. Multi-agent coordination is the mathematical problem underneath traffic networks, power grids, drone fleets, logistics, civic coordination, and autonomous vehicles — the systems the 21st century depends on. The dominant default is a PDE on agent density with O(N³) numerics, which crashes before 10,000 agents. BTUT replaces the default with a phase transition on a scale-free network under Fermi-rule updates, mean-field universality, constant 12-iteration convergence from 500 to 10,000 agents, and six live applied domains. Along the way it is also a new approach to DARPA Mathematical Challenge 13.

TCD-JEPA (Crystara) — a new AI training paradigm past transformers. Transformer scaling fixes the architecture at self-attention and pours more parameters and compute through it. TCD-JEPA refuses the fixed-architecture assumption: a recursive three-system loop explores the energy landscape with Fisher-information Langevin dynamics, runs Vietoris–Rips persistent homology on the trajectories, and crystallizes the stable features into typed H₀ / H₁ / H₂ predictor modules at runtime. The architecture is not designed; it is discovered. The first runtime-discovered predictor architecture for the JEPA family, and the first concrete instance of a post-transformer paradigm on working benchmarks: beats vanilla JEPA on three real heterogeneous graphs (+20 to +36.6 AUC pts) and beats supervised GAT (DeepMind), GCN (Google Brain), and GraphSAGE on a Georgetown CSET semiconductor supply chain.

Regenerationism — a new economic philosophy. The claim: the leading indicator of macroeconomic regime health is the velocity of capital formation with compounding margins, measured against cumulative friction — not bond-market sentiment, not linear averages of coincident series, not equilibrium-return dynamics. NIV is Regenerationism’s first operational instrument. It beats the Fed yield curve on several recession benchmarks (ROC-AUC 0.8538 @ 18 mo, 41.71% orthogonal variance), and the ensemble’s own Gini importance picks the regenerative-capital term unprompted at 0.9328. The repository is named regenerationism for a reason — the school precedes the signal.

Participatory Data Estate — a new framework for digital data governance. Every governance environment whose knowledge corpus is amended through human submissions (municipal agencies, NGOs, regulators, standards bodies, policy platforms, student governments) needs a data architecture that is continuous rather than batched, publicly auditable rather than privately moderated, and hardened to federal-agency patterns. The Participatory Data Estate proposes exactly that: Submit → Moderate → Thin → Crystallize, over pgvector + GIN FTS hybrid retrieval, with a publicly-readable approval ledger and federal-hardening controls. SGUNCCH is its first live deployment — a full UNC student-government stack running the framework end-to-end. A security posture student government has never needed, and has never had.

The four beats, stated plainly.

  1. BTUT — a new substrate for complex multi-agent systems. The class that civilization runs on: traffic, grids, drones, logistics, civic coordination, markets. BTUT replaces the O(N³) PDE default with a phase transition on a scale-free network (mean-field universality, β ≈ 0.5, constant 12-iteration convergence from 500 to 10,000 agents). One primitive, six live applied domains: traffic (Eclipse SUMO, 800-vehicle peak stress, zero gridlock), robotics (ROS / Turtlebot3), drone swarms (50 – 200, 100% cooperation), civic data (franklinstreetdata.com), game modeling (bigdunc.com), and four production cloud surfaces. Also: a new approach to DARPA Mathematical Challenge 13.
  2. TCD-JEPA (Crystara) — a new AI training paradigm past transformers. Transformer scaling pours compute through a fixed attention architecture. TCD-JEPA refuses the fixed-architecture assumption — a recursive three-system loop explores the energy landscape with Fisher-information Langevin dynamics, runs Vietoris–Rips persistent homology on the trajectories, and crystallizes stable features into typed H₀ / H₁ / H₂ predictor modules at runtime. +36.6 AUC pts over baseline JEPA on Georgetown CSET’s 519-entity semiconductor supply chain (82.7% vs 46.1%); also beats supervised GAT (DeepMind), GCN (Google Brain), GraphSAGE. +22.1 on GDELT; +20.0 at SEC EDGAR scale where GAT runs out of memory. 16 interpretable modules crystallized from persistent homology, 1-to-1 with real industry clusters — no labels, no prompting. The first runtime-discovered predictor architecture for the JEPA family.
  3. Regenerationism — a new economic philosophy, with NIV as its first operational instrument. The school’s claim: the leading indicator of macroeconomic regime health is the velocity of capital formation with compounding margins, measured against cumulative friction — not bond-market sentiment, not linear averages of coincident series, not equilibrium-return dynamics. NIV writes that school as a scalar. Ensemble ROC-AUC 0.8538 at 18 months across 504 months (1970 – 2024) and six OOS tests. 98.5% false-alarm suppression — 7 critical alerts in 42 years. 41.71% orthogonal variance to the Fed 10Y – 3M spread. Under Gini importance, the regenerative-capital term scores 0.9328 and the yield spread scores 0.0298 — the model picked the school unprompted.
  4. Participatory Data Estate — a new framework for digital data governance. The dominant data- governance stacks treat ingestion as batched, moderation as private, and audit trails as compliance tax. The Participatory Data Estate inverts all three: continuous ingestion (Submit → Moderate → Thin → Crystallize), moderation as a public transition (publicly-readable approval ledger via RLS), hybrid pgvector + GIN FTS retrieval with graceful fallback, federal-hardening controls (time-constant auth, RLS on every table, rate limiting on four action classes, CSP/HSTS, XSS detection). SGUNCCH is the first live deployment — a full UNC student-government stack running the framework end-to-end. A security posture student government has never needed, and has never had.

Every “paradigm” cited above has a working implementation, a measured benchmark against a named incumbent, and an open repository. The deep-dive chapters below give the mathematics, the tables, and the tear sheets in full.

01 · Multi-agent systems · Coordination
Read →

BTUT

A new substrate for complex multi-agent systems — the class civilization cannot run without (traffic, grids, drones, logistics, markets). Phase transition on a scale-free network, hub-weighted Fermi updates, β ≈ 0.5 (mean-field). Constant 12-iteration convergence from 500 to 10,000 agents.

BeatSix live applied domains: traffic (SUMO, 800 vehicles, zero gridlock), robotics (ROS), drones (100% cooperation at 200 agents), civic data (franklinstreetdata.com), game modeling (bigdunc.com), cloud surfaces. Also: a new approach to DARPA Mathematical Challenge 13.
02 · Post-transformer AI paradigm · TCD-JEPA
Read →

Crystara

A new AI training paradigm past transformers. Instead of scaling a fixed attention architecture, grow the architecture at runtime: Fisher-information Langevin exploration, Vietoris–Rips persistent homology on the trajectories, typed H₀/H₁/H₂ predictor modules crystallized from stable features.

BeatBeats vanilla JEPA everywhere tested: +36.6 AUC pts 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 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 macro 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.

BeatEnsemble ROC-AUC 0.8538 @ 18 mo across 504 months (1970 – 2024) and six OOS tests. 98.5% false-alarm filter; 41.71% orthogonal variance beyond the Fed 10Y – 3M spread. Under Gini importance the regenerative-capital term scores 0.9328; the yield spread 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, federal-hardening controls. Generalizes to any governance corpus — municipal, NGO, regulatory, scholarly.

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 on 4 action classes, CSP/HSTS, XSS detection, public approval ledger. A security posture student government has never needed, and has never had.

Reading order.

The chapters are ordered by the shape of the pipeline, not the chronology of the work. Start with BTUT to see the reduction; read Crystara to see what you do with compute once coordination is free; read NIV for a worked example of signal compression; read PDE for how the substrate stays alive. Then Convergence fits them together.