From information to verifiable knowledge.
Structured evidence becomes a governed knowledge base — canonicalized, deduplicated, versioned and evidence-linked, with lineage and citations down to the source. Served to people, agents and any model via API and MCP.
# a knowledge object — not text GET /v1/knowledge/objects/kn_842 # 200 OK { "claim": "Contract renewal date moved to 2026-09-01", "state": "SUPPORTED", "version": 4, "evidence": 3, "contradictions": 1, "relationships": 21, "lineage": "amendment_v2.pdf · page 4 · bbox", "abstains_when": "insufficient_evidence" }
AI stacks store conclusions and discard the proof.
Ad-hoc chunks, lossy markdown, orphan embeddings, answers with invisible foundations. Six months later nobody can say where a fact came from, whether it is still true, or what changed when it did. Knowledge becomes disposable — and unprovable.
A knowledge base that carries its own proof.
Knowledge objects
Claims with evidence, contradictions, relationships, versions and resolution states — not text. Insufficient evidence means ABSTAINED, never a guess.
Lineage and citations
Every fact traces to its exact source position — page, bounding box, timestamp — through a deterministic, reconstructible canonical substrate. When a defect is found, you can replay and prove exactly what changed.
Served everywhere
APIs and MCP for agents, LLMs, applications and people — plus a native experience, the Orves Console (pilot): ask, explore the graph, compare versions, and drill from any answer to its evidence. Swap the model, keep the knowledge.
Three moves. Nothing hidden.
Connect
Point it at certified parses — or at your sources and let the pipeline run end to end.
Adjudicate
Conflicting evidence is compared, deduplicated, versioned and resolved. Every change creates a new version; nothing is ever overwritten.
Serve
People, agents and any model consume knowledge objects with grounding, citations, abstentions and a certificate.
From evidence to knowledge, stage by stage.
Each stage is measured, versioned and deterministic — change enters through the flow, never by edit.
Where this engine sits in the platform → architecture
Anatomy of a knowledge object.
Not text — a governed object. Every part below is addressable through the API.
Evidence · 3 supporting · 1 contradicting
Relationships · 21
History
Certificate
Watch a claim live.
New evidence arrives; the claim is re-adjudicated; a new version is written. Nothing is ever overwritten — every state keeps its evidence.
Renewal scheduled for Jun 30
contract_2024.pdf · p.12Renewal postponed
renewal_notice.emlNew date proposed: Aug 15
draft_amendment.pdf · p.2Renewal moved to Sep 1
amendment_v2.pdf · p.4The object above is v4 of this claim — same asset, full history attached.
Retrieval finds text. The layer preserves why it is true.
Traditional RAG
- ▾
- chunks
- ▾
- embeddings
- ▾
- top-k
- ▾
- LLM
- ▾
- answer — origin unknown
When the answer is wrong, there is nothing to inspect.
On Verifiable Knowledge
- ▾
- structured evidence
- ▾
- knowledge objects
- ▾
- versions + grounding
- ▾
- LLM
- ▾
- answer — traceable to the page
When the answer is challenged, you show why.
One knowledge asset. Every model.
Models are interfaces — they change every quarter. Knowledge is the asset that compounds. Build it once, serve it to all of them, swap any of them tomorrow.
Swap the model. Keep the knowledge.
Measured before claimed.
Public benchmarks
Pre-registered evaluations against named competitors — published with the pilot, every number with its n.
Internal evaluation
Running today on proprietary gold sets. A number appears below only after it survives its gate — until then, the honest state is shown.
| Dimension | Unit | Status | Notes |
|---|---|---|---|
| Determinism | identical re-runs | measured · internal | byte-level round-trip checks |
| Reconstruction fidelity | knowledge → source | measured · internal | publication after pre-registered eval |
| Claim extraction | precision / recall vs gold | measured · internal | publication after pre-registered eval |
| Abstention calibration | abstains when evidence is insufficient | measured · internal | UNKNOWN counted, never hidden |
| Grounding fidelity | answers traceable to claims | measured · internal | measured on adversarial sets |
| Chunk boundary quality | structure alignment | publication pending | gold set in construction |
A knowledge base that cannot show its grounding is measured as wrong — fluency does not count. Continuous benchmark →
Credits per unit of work.
Certificates included with every artifact. How pricing works →
Every artifact is verifiable.
Certification runs through everything Orves produces. Anyone can verify a certificate — free, forever, no account.
Integrate in minutes.
Everything is delivered through the API. Nothing to install, nothing to host.
FAQ
What exactly am I buying?
A knowledge base where every fact carries origin, context, history, versions and evidence — queryable by people, agents and any model. Not a vector store, not a wiki, not a chatbot: an asset you can audit and prove.
Does it compete with GPT or Claude?
No — it feeds them. Models are interfaces; this is where your knowledge persists. Swap the model, keep the knowledge. We don't build AI — we build knowledge AI can trust.
Where did Canonical and Knowledge go?
They became capabilities of this product. Canonicalization is the mechanism that makes knowledge deterministic and reconstructible; the persistent memory layer is how it is served and accumulated. You buy the outcome — knowledge you can prove — not the internal stages.
Why aren't chunks and embeddings enough?
Chunks discard structure and context; embeddings discard provenance and history. Both keep what the text says and lose why it can be trusted — source, version, relationships, adjudication. Retrieval finds text; the layer preserves why it is true. RAG built on chunks cannot answer 'how do you know?' — knowledge objects can.
Why does determinism matter?
Because proof requires replay. When a defect is found, the canonical substrate lets you re-derive every downstream artifact and show exactly what changed — and what did not.
Does my knowledge stay current?
Yes, when you need it to. Continuous knowledge is a capability of the platform — approved sources are monitored, changes are detected, history is preserved and new versions are created. Knowledge is never overwritten, and provenance is maintained through every update. It runs as a recurring service — see the Living Knowledge solution.
Is my knowledge shared?
No. Each instance is isolated by construction and reasons only over your governed knowledge. Nothing you ingest trains shared models.
Is there an interface, or is it API-only?
Delivery is API-first, and the product includes a native experience — the Orves Console (pilot): question answering with evidence, graph exploration, version comparison and the Show-me-why drill-down from any answer to its source span. It is how you experience the product, not a fourth product.