Loominum™ 1.0
One architecture on a continuous axis — Marrow ↔ Loom ↔ attention.
Learn More
Research
We work on some of the hardest, most interesting problems in verifiable intelligence
Across substrate, memory, training, agent safety and human–AI collaboration — research that makes powerful systems more capable, understandable and governable.
Breakthroughs
The systems behind the stack
Explore the innovations that underpin Loominum and the ReasonLoom stack.
Atelier
A developmental trainer that grounds models in evidence.
Learn More
Mnemo
Durable, auditable memory that holds at a million entities.
Learn More
Imprint Studio
Forensic media integrity and provenance you can prove.
Learn More
Palestra
A cognitive gym where humans and AI reason by debate.
Learn More
Emmy
Reproducible research pipelines for any scientific field.
Learn More
Obsidian
Neuromorphic silicon where physics meets cognition.
Learn More
Latest news
Breakthroughs and updates from the lab
Structured memory holds perfect multi-hop recall at a million entities
Our memory layer keeps precision-at-5 = 1.00 on multi-hop retrieval as the store grows to one million entities — no degradation from fifty thousand to a million.
Learn more
Rearing models like children beats distillation on sample efficiency
A developmental trainer that grounds, corrects, binds, and consolidates learns concepts from far fewer exposures — and does not forget them.
Learn more
Curvature-native silicon: where the chip’s physics is the model’s math
Our substrate program commits to classical room-temperature silicon that realises computation as curved geodesic flow — co-designed so the hardware’s relaxation is the model’s inference.
Learn more Why we publish the experiments that failed
Every claim on this site has a falsifiable test behind it — and when the test says no, we report the no. That discipline is the product.
Learn morePublications
Areas of inquiry
Beyond transformers
Biology-first compute and substrate-native cognition — including the negatives.
Learn more Structured memory
Analogy, counterfactual, compositional binding — measured at the million-entry mark.
Learn more
Geometry-native compute
Silicon where the physics and the geometry of cognition are designed together.
Learn more Palestra — a cognitive gym for human–AI collaboration
A debate-and-effort gym where humans and models practise reasoning under live evaluation.
Learn more Emmy — a research orchestrator for any scientific field
A workspace orchestrator that runs reproducible research pipelines across scientific domains.
Learn moreMeasuring reasoning under real constraints
New evaluations for planning, factual humility, tool reliability, and collaborative problem solving.
Learn more
Safety methods for agentic systems
Research on oversight, capability boundaries, secure tool use, and robust refusal behavior.
Learn more
Resilient training across distributed clusters
Techniques for keeping frontier training stable when compute, data, and teams span continents.
Learn moreFrom the lab
Uncovering how intelligence works
RL-L1: a continuous-time model that ties clinical SOTA at a fraction of the size
On real ICU data, our continuous-time model matches the standard clinical baseline while running far smaller and faster — the case for treating time as an input, not an afterthought.
Learn more
Mnemo turns research-grade memory into production infrastructure
A durable, multi-tenant memory bridge with compliance and audit endpoints as first-class citizens — fully covered by tests, retrieval quality intact.
Learn more