The kernel contract for agentic computing.
Foundational research in agentic systems — conducted with the rigor of a scientific institution.
Stratum
The kernel contract for agentic computing. Safe, auditable, composable by design.
Read the specification →Alloy
Unified cross-modal representation learning. A single representation for arbitrary-dimensional signals.
View the research →Seam
Multimodal data acquisition pipeline. Ingestion, storage tiering, and preprocessing at scale.
View the pipeline →Noted. We will be in touch only when there is something worth saying.
We publish research notes, architecture decisions, and project updates. No cadence, no noise — only when there's something worth saying.
The First Approval
A proof of mechanism is not a proof of scale. It demonstrates that the core claim of the architecture is true — that the thing we said would happen does, in fact, happen.
The plugin architecture of Stratum
The action surface is a plugin host. It has no built-in knowledge of any domain — no filesystem operations, no process management, no networking. Domain packages register plugins at startup.
The Shape of Work
Choosing a queue says these things happen in order. Choosing a graph says these things have relationships — and the relationships are what matter, not the sequence.
The reversibility engine: classifying actions before they execute
Not all actions are equal. Reading a file is not the same as sending an email. The architecture needs to encode this distinction formally — at runtime, not statically.
The audit store: append-only, tamper-evident, always reconstructable
State is a projection of an event log. You can always reconstruct how you got here. The audit trail that captures what the agent did is tractable. The one that captures what the agent understood is not.
Trust expansion through observed compliance: on the behavioral record
Permissions expand as behavior warrants. The initial trust boundary is narrow. It widens through observed compliance, not through configuration — and that distinction is harder to operationalize than it sounds.
Workspace surface: isolating agent context from the global environment
The workspace is not a sandbox in the security sense — though containment is a property we want. It is a principled definition of what an agent can see and touch at any moment of execution.
Why we are building stratum: the substrate problem in agentic computing
There is no agreed substrate for agentic systems. Every team builds their own orchestration layer, their own permission model, their own audit trail — or skips these things entirely and hopes for the best.