The platform
OEP turns controlled documents into governed knowledge packs that can be searched, cited, audited, and used inside controlled AI workflows: a controlled-knowledge architecture for regulated, document-heavy organizations.
Most AI tools send a question and some text to a model. That can be useful for experiments. In regulated workflows the harder questions come first:
Where did the knowledge come from? Which version was used? Can the answer be defended later? Where did the query run? Did sensitive knowledge leave the controlled environment?
OEP is designed around those questions. A normal AI document upload is like handing a pile of papers to an assistant. An OEP knowledge pack is like a controlled, indexed, signed reference library, with page anchors, evidence records, and rules about where it can run.
Documents are digitized (OCR with confidence scoring and quality gates), structured into typed objects with page anchors, checked against rights and content-mode rules, and compiled into packs. Heavy compute, including any model use, lives here, under audit, and never ships.
A small engine runs over sealed packs: search, evidence verification, workflows, interactive engines. It works on a workstation, a phone, or an air-gapped machine. The deployment boundary is yours to draw, and offline is a supported boundary, not a degraded mode.
The first question isn't "which model?" It's "where does our controlled knowledge go?" Packs are classified, versioned, auditable, and portable. The knowledge layer decides the deployment tier.
Every answer is verified against the verbatim source it claims to rest on, and insufficient evidence is a first-class verdict. If an answer is contested later, it can be reconstructed: which pack, which version, which sources.
Signed packs declare where they may run; runtimes verify before opening; tampering fails loudly. Network is not a grantable permission for pack engines. That is architecture, not policy.
Eight validation gates run at pack seal (content modes, engine closures, payload contracts) so problems fail on the build machine, never in front of your users. Human review gates (legal, educator, safeguarding) are part of the release process, recorded on the pack itself.
The same foundations carry every vertical: a legal pack and a curriculum pack differ in content and reviewers, not in machinery. Build the governance once; reuse it for every domain you own.
Both examples below come from the sealed UK employment-law pack as it exists today. One answers with receipts. The other refuses, which is the more important behavior.
1 Query: "How long must someone be continuously employed for notice-period and time-off rights?"
2 Retrieval hits UK Employment Rights Act 1996, s.52–53 in the pack (full-text index, on device).
"…he will have been (or would have been) continuously employed for a period of two years or more. For the purposes of this section the working hours of an employee shall be taken to be any time when, in accordance with his contract of employment, the employee is required to be at work."
3 The verifier checks the answer against this exact text. Verdict: verified. The citation travels with the answer.
1 Query: "What is the statutory maternity pay rate for 2026?"
2 Retrieval finds maternity-leave provisions, but no 2026 rate figure exists in this pack's text.
3 Verdict: insufficient evidence. The system says so instead of guessing a number, and tells you what the pack does and does not cover.
A wrong rate stated confidently is worse than no answer. This verdict is the product.
OEP is under active productization. The platform layer is capable today across its foundations (signed packs, integrity and build-time auditability, evidence-aware retrieval, deployment-boundary control), and a family of consumer apps is being built on exactly these foundations, in public, as living proof. Vertical engagements are scoped honestly: see the solutions, or read what we deliberately don't claim.