Research
Dominion Labs is developing persistent intelligence systems: AI that maintains knowledge, learns from experience, reasons, acts, verifies outcomes, and develops capabilities over time. Our research focuses on the systems and mechanisms required for intelligence to persist beyond a single model invocation—including epistemic governance, grounded perception, reasoning, planning, autonomous action, independent verification, cross-domain transfer, persistent learning, and functional metacognition.
We build these capabilities as working systems and evaluate them through reproducible experiments, with the aim of advancing a new class of AI that can accumulate knowledge and capability through continued operation. We welcome researchers, engineers, and institutions interested in rigorously exploring and extending this direction.
Perceive, Induce, Name
A proof of concept for vision in a cognitive substrate with no perceptual model: classical algorithms recover visual structure, the substrate learns to name it from a few labelled examples, and it abstains when it has not. Held-out instances are named at 100% recall with zero false namings and zero language-model calls; an ablation shows the naming ability is carried entirely by what was learned. The study is repeated across eight categories over 149 images, the supervision each name requires is measured, and the operating range of the perceptual stage is mapped.
Reasoned vs. Believed
A cognitive substrate knows in two ways: by deriving over the structure it has learned, and by holding graded beliefs. We study the gap between them, why it appears, and the discipline that keeps a reasoner honest: derive over curated structure, bound derivation, let belief back-stop it, and never assert beyond what is grounded.
Knowing a Fact vs. Being Able to Act
A taught system holds two independent kinds of mastery: the facts it knows and the operations it can perform. Measuring them as orthogonal axes lets it tell a missing fact from a missing skill and abstain honestly on that basis.
Task Completion as an Internal, Grounded Judgment
When is an autonomous agent's task actually done? We make completion a grounded belief that must pass the agent's own doubt, separating what it did from what it independently observed, and report zero false completions on a fixed judgment suite.
Systemic Epistemic Governance
An operational semantics for provenance-gated admission in persistent cognitive systems. We separate the ability to derive a proposition from the authority to persist and act on it, and enforce that separation as a system-wide state-transition invariant with proved structural guarantees.
Structure Before Meaning
Giving a cognitive substrate sight. Seeing divides into perceiving structure (solved by algorithms with no learning) and naming concepts, the one part that must be learned and which every current system buys from a pretrained backbone. We argue the substrate can learn that half itself, and build vision as a single faculty: one entry point for all sight, perceiving structure by algorithm and learning meaning by induction.
Lyric Unified Cognitive Substrate
The canonical architecture of Lyric: a persistent cognitive substrate that organises reasoning, learning, memory, action, belief, competence, and epistemic governance over one shared state, model-optional by design, with the cross-faculty loops, grounded task completion, and governance boundary that connect them.