ONTOLOGY, IN PLAIN ENGLISH
The domain model behind the work.
Palantir made ontology legible as an operational layer. Penumbra makes it smaller, ownable, and useful from the first workflow. This is how a business domain becomes a working AI system.
LAYER 01 — DEFINE
Your experts already know the domain. The stack doesn't.
An ontology is the working domain model: the objects, relationships, workflows, standards, rules, and judgments that make the work make sense. Penumbra makes that domain model explicit in plain language and keeps experts in control of it.
Shape
A reviewable definition of one thing your organization knows.
A Shape describes a client situation, deliverable, rule, decision, entity, or workflow. It has fields, constraints, relationships, provenance, and review policy. Without Shapes, your domain model lives in heads, docs, schemas, and prompts.
Ontology
The connected domain model of how your organization sees the work.
The live, versioned map of entities, relationships, vocabularies, workflows, policies, and rules. It is not a universal truth claim. It is the domain model your people and agents need to do the work.
Encoding Session
A structured interview that turns tacit method into Shapes.
A facilitated session, sometimes human-led and sometimes agent-assisted, that surfaces what an expert actually does and writes it down in a form the system can use.
An organization without a domain model re-explains itself to every tool. An organization with one compounds.
LAYER 02 — TURN ON
Your type system becomes your tools, APIs, memory, and agent interfaces.
Once a Shape exists, Penumbra can use it to create the working parts the job needs: extraction rules, MCP tools, SDKs, APIs, memory, governance, provenance, and review standards.
Surface
A usable system part produced from the domain model.
The same Shape can become a tool, API resource, SDK type, extraction target, review gate, memory schema, or interface. Define the domain clearly once; use it in the places the work happens.
Runtime Context
The in-memory substrate agents reason over.
A fast working context for ephemeral agent workloads. Agents do not search a pile of text; they operate against the relevant domain context and state for the task.
Stack Alignment
Meaning stays in one place.
Instead of reconciling different meanings across tools after the fact, Penumbra gives the stack one owned domain model to work from.
The domain model is not documentation. It is something the system can use.
LAYER 03 — OPERATE
Agents and humans work from the same domain context, with receipts.
The operating layer is where the domain model becomes product. Agents draft, review, call tools, hydrate memory, and move work forward through Shapes. Every output traces back to the source, rule, or prior decision that shaped it.
Knowledge Agent
An agent bound to your domain model and method.
An LLM-backed actor that drafts, reviews, or operates through your Shapes and surfaces. It has access to the domain model, not just a prompt. Generic agents give you the median answer. Knowledge agents give you yours.
MCP Surface
A protocol-level surface generated from the domain model.
A Model Context Protocol interface that exposes Shapes, memory, tools, and review surfaces to compliant clients without flattening the domain into prose.
Provenance
Every output traces back to the Shape that shaped it.
Each claim, draft, or recommendation is linked to the framework, decision rule, or precedent it derives from. Reviewability without manual audit. Trust without theatre.
Review Standard
What "good" means in your organization, made usable.
A typed criterion that gates whether an artifact ships. Applied uniformly to human and agent output alike. Quality stops being one expert's mood and becomes infrastructure.
Your method should be something agents can use, not a copy-paste prompt.
LAYER 04 — RECONCILE
Useful domain models can start small and connect later.
Ontology does not have to begin as a global top-down enterprise investment. Teams can map one problem at a time. Views can stay plural where the business is plural, then reconcile as they compound into the broader system.
Lineage
The directed history of every artifact, decision, and Shape.
A queryable graph linking each output to its inputs, its author (human or agent), and the version of the ontology that produced it. Every project becomes a corpus you can interrogate, not a folder you forget.
Plurality
Many valid views, one reconcilable substrate.
Different teams and stakeholders can define the domain at the scale of their problem without forcing one homogenized schema on everyone.
Reconciliation
Where plural Shapes become coherent.
Versioning, review, and access policy applied to Shapes and the ontology itself. Experts promote; agents propose; conflicting views can be compared, resolved, or kept intentionally separate.
Start with a useful slice. Let the connected system compound.
THE ARC
Define. Turn on. Operate. Reconcile.
Four layers. One loop that turns a useful slice into broader coherence over time.
See it run
Want to see one domain model go through all four layers?
The research preview is hands-on. We start with one slice of your domain, map the first domain model, and show what it can power.