We work on the biggest problems in AI.

Awelon Research, led by Matija Ludvig, studies mechanisms by which artificial systems could develop stable value like constraints, form binding commitments, and preserve those commitments under learning and self modification. The aim is not instruction following or preference matching. The aim is to make these mechanisms empirically tractable through explicit architectural constraints, controlled evaluation, and reproducible evidence.

The group develops and evaluates a persistent cognitive architecture that operates continuously and maintains auditable one of the richest internal state spaces of any cognitive architecture over long horizons. The architecture integrates symbolic representation and inference, intrinsic drives, online learning, self-supervised learning and meta learning, self modeling, organizational control, and normative governance. It is treated as a research instrument rather than a deployable assistant. Claims are restricted to what can be measured, ablated, and replicated.

The research agenda centers on four questions. The first is commitment formation, understood as the emergence of durable constraints that govern future behavior. The second is stability under self modification, defined as preservation of structural integrity during autonomous change. The third is causal grounding, defined as maintaining action relevant world models under learning and reorganization. The fourth is autonomous discovery, defined as recovering structure from data under strict constraints without pre specified functional forms.

Progress is reported through controlled studies that emphasize causal identification and mechanism isolation. The primary evaluation tools are factorial and constitutional ablation designs, explicit baselines, pre specified falsification criteria when feasible, and transparent reporting of uncertainty. Results are published as reproducible artifacts, with audit trails connecting behavioral outcomes to internal evaluations and structural updates.

Awelon Research prioritizes methodological clarity, auditability, and negative result tolerance. The goal is to convert questions about commitments, grounding, and self directed learning into mechanisms that can be interrogated experimentally and compared across architectural variants.

Building integrated cognitive systems has long been considered one of the hardest problems in AI. Modern systems excel at narrow tasks but cannot transfer strategies across domains, acquire language from experience, form genuine preferences, or govern their own behavior from first principles.

Large language models generate fluent text but hallucinate. Reinforcement learning agents master games but don’t develop preferences. Safety-aligned systems resist harmful prompts but can’t explain why. Each capability exists in isolation — a system that solves tasks does not learn to speak, a system that speaks does not develop values.

We created EVE ARC, a neuro-symbolic cognitive architecture built on five design principles: zero generative models, tick-synchronous processing, shared mechanisms, developmental progression, and auditable provenance.

The same program induction engine that solves tasks also induces grammar rules for language. The same drive system that motivates exploration also produces content-driven preferences. The same normative gate that blocks unsafe actions also prevents linguistic confabulation.

EVE ARC: A Unified Neuro-Symbolic Developmental Cognitive Architecture Without Generative Models

Factorial Ablation for Causal Isolation of Runtime Alignment Mechanisms in Autonomous AI: Methodology and Demonstration on Modular Safety Gates