Research

Intelligence from structure

The field’s default bet is size: more data, more parameters, more compute. Ours is different. We believe genuine intelligence comes from structure — new mathematics and new architectures, intelligence that reasons from how it is built rather than how big it is. This page lays out the lines of work and, for each claim, how far it has actually been taken.

Research map

Four lines, described at the level of the idea. Implementation details stay private; the direction does not.

Architecture

Structures that make reasoning a property of how the system is built, rather than an emergent side effect of size.

Memory & growth

Continual learning where new knowledge accrues and memory grows without overwriting what was already learned.

Safety by architecture

Treating safety as a foundational requirement of the design — transparency of reasoning, independent oversight, trust earned in stages.

Mathematical foundations

The formal groundwork the architecture rests on: parts derived, with the open problems named rather than hidden.

Evidence status matrix

Every public claim carries one status. We never label something “Tested” without evidence, and we never dress private work up as published.

Continual learning without catastrophic forgetting internal experiments

Prototype

Long-context / streaming memory

Prototype

Architecture-level safety model

Position

Mathematical foundations core derived; open problems remain

Hypothesis

Implementation details & weights

Private
Tested
Demonstrated in our own experiments. Not a peer-reviewed or public claim — internal evidence only.
Prototype
A working implementation exists and shows the effect, but it is early and not yet validated at scale.
Hypothesis
A direction we have reason to pursue, partly worked out, not yet shown to hold.
Position
A design or ethical stance we argue for — a commitment, not an implemented mechanism.
Private
Real work we are deliberately not publishing: implementation details and weights.

Open questions

The honest part. These are unresolved, and the answers could move our own claims up or down.

  • How far does learning-without-forgetting hold as tasks and time horizons scale, rather than in bounded experiments?
  • Which parts of the mathematical foundation are load-bearing, and which open problems would break the thesis if they resolve the wrong way?
  • Can architecture-level safety be measured and independently audited, or does it remain a design argument until someone tries to break it?
  • At what point, if ever, does a system warrant being treated as more than a tool — and who decides?

Public record

Public research notes are in preparation. Until then, this page records the current thesis, its evidence boundaries, and the questions that remain open.