Agents do not just consume enterprise knowledge. They create a new knowledge layer. Proofpress governs it.
The defining question is not only what an agent can retrieve. It is: What may the next agent or human rely on, why, and under whose authority?
Existing knowledge infrastructure organizes what agents reason from. Proofpress governs what their reasoning produces.
Enterprise knowledge is what the organization already knows: documents, databases, policies, domain ontology, and memory. Agents reason from it.
Agent-produced knowledge is what agent work newly creates: claims, verdicts, findings, analyses, and decisions. It is derived through research and reasoning, then increasingly reused by other agents and humans.
flowchart LR
EK["Enterprise Knowledge<br/>documents · data · policies<br/>ontology · memory"]
A["Agents<br/>research · reason · work"]
APK["Agent-Produced Knowledge<br/>claims · verdicts<br/>analyses · decisions"]
PP["Proofpress<br/>evidence · verification<br/>authority · lifecycle"]
GK["Governed Agent Knowledge<br/>current · scoped · authorized"]
D["Downstream<br/>agents · humans"]
EK --> A --> APK --> PP --> GK --> D
Proofpress governs the second layer. It is not another enterprise knowledge graph, ontology, memory system, workspace, or agent orchestrator.
The adjacent layers are complementary, but their primary objects and questions are different. From granular execution events upward:
//: # (ob:9ad7a7d9)
| Layer | Primary object | Core question |
|---|---|---|
| Observability | Runs, traces, tool calls, execution events | What happened while the agent worked? |
| Memory | Retrieved context and prior interactions | What should the agent remember or read next? |
| Knowledge graph / ontology | Enterprise entities, concepts, relationships | What does the organization know about its world? |
| Proofpress | Agent-produced claims, evidence, authority, lifecycle | Which claims may future agents or humans rely on—and why? |
Observability can supply evidence. Memory can retrieve governed claims. Enterprise knowledge graphs can supply the knowledge agents reason from. Proofpress governs the reusable knowledge that agent work produces.
As agent adoption and autonomy increase, enterprise knowledge may continue to grow relatively steadily while accumulated agent-produced claims and work grow much faster. More agents, more runs, branching research, and agent-to-agent reuse all create derived knowledge without requiring the original enterprise corpus to grow at the same rate.
This is a directional product model, not a claim of a universal mathematical growth law. The important shift is operational: beyond a threshold, teams can no longer review every claim informally. Verification, authority, and lifecycle become infrastructure.
An agent reads enterprise knowledge, researches and reasons, and produces a claim. A later agent then receives that claim as context. Without a governance layer, a derived claim can lose the source, scope, assumptions, review state, or authority that made it usable in the first place.
The failure compounds across handoffs. Retrieval can surface an old claim without showing that its dependency was revoked. A trace can show how an agent worked without deciding whether the result is approved for reuse. A knowledge graph can represent the claim without establishing who may rely on it.
Proofpress makes this transition explicit: agent output becomes future input only through governed admission and scoped projection.
The product object is a graph of claims and the claims they depend on. It binds each reusable claim to evidence and provenance, verification and review, authority and scope, dependencies, and later contradiction or supersession.
flowchart LR
C["Claim"] -->|depends_on| K["Claim"]
K -->|supported_by| E["Evidence"]
C -->|scoped_by| A["Authority"]
C -->|supersedes| P["Previous Claim"]
This is not a generic graph of enterprise entities. It is a graph for answering whether a concrete claim is currently eligible to enter a downstream human or agent's context.
Agent work becomes governed context through extraction → evidence binding → verification → admission or review → governed claim graph.
- Deterministic Checks: enforce fixed rules and required evidence, then return pass or fail.
- LM Judge: evaluates meaning against organizational policy, then recommends or escalates. It cannot authorize reuse.
- Human Approval: an authorized human admits or rejects the claim for a defined scope. Only this gate enables downstream reuse.
Rejected, unresolved, expired, superseded, unauthorized, or dependency-blocked claims remain auditable but stay out of default context. Admission is not a declaration of universal truth. It is a scoped, inspectable decision about reliance.
Proofpress currently provides a local ledger and CLI, local review and context
UI, supported agent adapters, artifact provenance, and portable Markdown and
static-HTML carriers. The context projection returns admitted, current
claims that match the requested scope and actor.
A supported public API/SDK, MCP server, hosted service, and production connectors are planned, not shipped. The local UI's endpoints are implementation details rather than a public integration contract.
A frozen product study tested 7 models across 3 Harvey LAB-derived legal task families and 126 valid paired runs. In that bounded panel, governed handoffs raised rubric completion from 89.3% to 93.4% (+4.1 percentage points) and reduced observed unsafe propagation from 8 to 0 across 63 controlled stress pairs.
This is evidence for one composed mechanism under frozen conditions—not the definition of Proofpress, an official Harvey leaderboard score, a population-level causal claim, a statistical-significance result, or evidence of improved legal intelligence. The retained receipts and boundaries are in the public results.
As agents create and reuse more of an organization's claims, governance must attach to the knowledge produced by agent work—not only to the data the agents started from.
That is the layer Proofpress is building.