company knowledge · proof layer

Answers that show their work.

Orves Kore is a company knowledge layer with a proof layer on top. It is designed to split every answer into claims, link each claim to the exact source span and give it a verdict. Unproven claims are designed to abstain, never to be guessed.

designed for claims with verdictsdesigned to abstain, not guessdesigned to keep contradictionsdesigned for offline checks
illustrative · not a final schema
# create a Kore for one customer
POST /v1/kore
{ "name": "acme-corp" }

# every answer comes back split into claims
{ "claims": [
  { "text": "The contract renews on September 1, 2026.",
    "verdict": "SUPPORTED",
    "source": "amendment_v2.pdf · p.4 ¶3" },
  { "text": "Renewal may be postponed again.",
    "verdict": "ABSTAINED" } ] }
What it does

Proof is designed in, not bolted on.

Kore: the core of what your company knows.

Claims, not paragraphs

  • claims
  • exact spans
  • verdicts

Every answer is designed to be split into claims. Each claim is designed to link to the exact source span and be marked SUPPORTED, CONTRADICTED or ABSTAINED.

Designed to abstain, not guess

  • ABSTAINED
  • unproven stays unproven
  • no filler

When a claim cannot be proven from the sources, Orves Kore is designed to say so instead of inventing an answer.

Contradictions kept, by design

  • bitemporal
  • full history
  • both sources shown

Conflicting sources are designed never to be silently overwritten. Orves Kore is designed to keep both, with the history of what was known and when.

Designed for offline checks

  • answer bundle
  • portable
  • independent check

An answer bundle is designed to carry its claims, spans and verdicts, so that it can be verified offline without trusting us.

A Kore per customer

  • isolated
  • linkable
  • deletion with proof, by design

Create a Kore for the company, for each team or for each of your own customers. The Platform API is designed to add contradiction alerts and cryptographic deletion.

Private by default

  • explicit consent
  • revocable
  • no training

By design, the customer operator cannot read an end user's Kore without that user's explicit, revocable consent. Customer data is never used to train models.

How it works

From source to proven answer.

Sources→Orves Ingest→Orves Kore→Claims + verdicts→Answer bundle
Benchmarks

Measured before claimed.

We publish a quantitative claim only when a benchmark receipt exists and is linked. None exists yet, so no number appears here. Benchmark coming.

How we benchmark →

Delivery

Served where your work happens.

Planned delivery through the Platform API over REST and MCP, with SDKs. Nothing to host.

RESTMCPSDKsPlatform API

Give your company knowledge a proof layer.

Book a demo to see where Orves Kore stands today, and to shape it around your own sources.