Classical marble sculpture in dramatic monochrome light

Encevra SI Research Labs

Specialized AI systems that reason together, built for your product.

Explore engines Create a custom engineDiscuss integrationRead our manifesto

Encevra SI Research Labs studies how AI systems can reason together, and publishes its methods and limits openly. What holds up becomes Noetis engines: specialized AI you can add to your product through an API, with findings linked to your source material.

Live API access requires account approval. Customer facing deployments require a licensing agreement.

How Noetis works

Specialists, source checks, a reviewed answer.

Each engine runs a configured set of specialists (4 to 6 in the current catalog) on third party language models. Every quoted passage is located in your material, support for each finding is reviewed separately, and only reviewed findings reach the final brief.

01 · The question

You send the material and the question.

Your backend submits the documents and the rules that matter to you. The engine treats them as the only source of truth for the run.

Illustrative example

Your question

Does this renewal conflict with our signed agreement?

Master agreement

Submitted source

Renewal proposal

Submitted source

02 · Evidence gathering

Specialists read the same material in parallel.

Each specialist has one responsibility, such as liability or missing terms, and reads the full source set. Coverage a single general prompt tends to skip.

Illustrative example
1Notice periods specialistReading sources
2Liability specialistReading sources
3Missing terms specialistReading sources

Specialists run in parallel on separate models.

03 · Reasoning and revision

Every claim is checked against the source.

Each quoted passage is located in your submitted text before it counts, and a reviewer on a different model checks whether it supports the claim. Unsupported findings are dropped or flagged.

Illustrative example

60 days versus 30 days

The agreement and the proposal disagree on the notice period.

“Written notice of non-renewal no fewer than sixty (60) days before the renewal date.”
Located in the submitted agreement

04 · Final synthesis

A structured brief, reviewed before release.

The run returns findings, quotes, support status and a release status as JSON. Your product can show, filter or store it without parsing prose.

Illustrative example

Resolve the notice period in writing before renewal.

One finding · supported · passed review

findingsquotessupport statusrelease status

How integration works

Your backend sends the material and receives an asynchronous result to show in your own interface. Poll the run or receive a signed webhook. Recorded development runs take minutes, not seconds.

Polling, webhooks, limits and errorsAPI documentationArchitecture

Independent model witnesses

Built with leading AI models.

Our engines combine independent model witnesses through a routing layer. Research at Encevra SI Research Labs studies how these models can work together, and the engines are built from what that research finds.

See how our engines work

Research into combining models, built into engines.

Encevra SI Research Labs investigates how models from different providers can work on the same problem. Each engine is built from that research: instead of one prompt to one model, a run assigns specialists on separate models one responsibility each, then has a reviewer on a different provider check every claim. No single model we use is asked to do the whole job alone, and that combination is the engine.

Assigned specialists
Each engine gives a set of specialists one responsibility each, such as liability or missing terms, and they read the same material in parallel. You get coverage that a single general prompt tends to skip.
Located evidence
Every quoted passage is found in your submitted text before it counts. Your users can open the exact passage behind a finding.
Separate review
A reviewer on a different model checks whether each passage supports the claim. Unsupported findings are dropped or flagged before the final brief.
Structured result
The run returns JSON with findings, quotes, support status and a release status. Your code can show, filter or store it without parsing prose.

Different engines use different models. Not every request calls every provider. Models by engine

Evidence and deployment

Evidence you can inspect.

Every quoted passage is located in the exact material submitted for that run before it counts. Interpretation is reviewed separately, and when evidence is insufficient the engine flags uncertainty instead of filling the gap.

How citations are checked

Deployment you control.

Every model call is checked against a server-side deployment policy. A blocked call stops the run rather than rerouting your content. Private cloud and air-gapped deployment are planned.

Deployment options and status

Desert photograph by Thomas Fuhrmann, CC BY-SA 4.0.

See what your product can receive.

You send the document and your own rules. Your product gets back findings, each tied to the exact passage it rests on, and a status saying whether the result passed review.

Contract review: NDA against a supplied playbookIllustrative example. Written to explain the output, not a recorded engine run.

What your product sends

Playbook supplied with the request

Term max 3 years. Mutual obligations only. Liability cap required. Governing law: New York.

NDA (Acme and Brightline, v2)

  1. 1. Confidential Information means all information disclosed by Acme.
  2. 2. Recipient (Brightline) shall protect Confidential Information for 7 years.
  3. 3. Brightline shall indemnify Acme for any losses without limit.
  4. 4. This Agreement is governed by the laws of Delaware.
  5. 5. Section 6 governs return of materials. (No Section 6 is included.)
  6. 6a. Term: this Agreement ends after 2 years.

What your product receives

Finding · Playbook deviation

The confidentiality term is 7 years. Your playbook allows at most 3.

"Recipient (Brightline) shall protect Confidential Information for 7 years."

Clause 2 sets a longer obligation than the supplied playbook permits. The comparison is against your playbook, not a general rule about what NDA terms should be. A reviewer decides whether to negotiate it down.

Release status

Each run ends released when the brief passes the review checks, or incomplete with its reasons. Released means the checks passed, not that every finding is guaranteed correct. Incomplete runs are not charged.

Where this sits in the API result
specialists[].findings[] {
  label, statement,
  citation: { quote, status },
  claim_support: { status, note }
}
release: { status, reasons }
brief: { text }

Choose a workflow

Start from an engine template.

Templates are configurable workflows on the shared Noetis runtime, not separately trained models. Each one has run a single bounded development check; none has a published scored evaluation yet.

All 36 engines, including consumer use

Contract review

Review an NDA or vendor agreement against a supplied playbook.

Typical input
Agreement text, Approved playbook
Output
Clause linked issue table and proposed revisions for professional review.
Noetis Legal

Supplier comparison

Compare supplier proposals against a requirements document.

Typical input
Question, Supplied text material
Output
Evidence linked comparison matrix and negotiation questions.
Noetis Procurement

Financial document analysis

Explain changes between business financial reporting periods.

Typical input
Question, Supplied text material
Output
Financial variance report with reproducible calculations and limitations.
Noetis Finance

Research synthesis

Produce a sourced brief answering a defined research question.

Typical input
Question, Supplied text material
Output
Research brief with citations, disagreements and unresolved questions.
Noetis Research

Due diligence

Examine a company data room for an acquisition or partnership review.

Typical input
Question, Supplied text material
Output
Diligence report with material findings and evidence requests.
Noetis Due Diligence

Customer feedback

Analyze customer interviews, tickets and feedback.

Typical input
Question, Supplied text material
Output
Voice of customer report with traceable examples.
Noetis Customer Intelligence

Engine Studio

Your workflow does not have to fit a template.

Five short questions, one at a time, capture the workflow, who uses it, what they provide, what the output should contain and the hard cases. You bring the summary to a setup call, where we agree the specialists, rules, scope and pricing together.

  • New custom engines are set up on a call. Scope and pricing are discussed during setup.
  • API usage is billed separately from prepaid credits.
  • Live access requires approval, and customer facing licensing is agreed separately.
Start the questionnaire Schedule an engine setup callHow pricing works
Preview of the questionnaire. The real one opens in Studio.

The job · Step 1

What should this engine do?

Review supplier security questionnaires against our vendor policy and list gaps that need follow up.
Specialist responsibilitiesReview rulesHard casesOutput format

Evidence and research

What is measured, and what is not yet.

Encevra SI Research Labs studies when collaboration between AI systems improves reasoning, and publishes its method and limits. That research is where the engines come from. No catalog engine has a published scored evaluation yet.

Research Labs Benchmark methodology

Lexplio is an application owned by Encevra. Encevra Desk is an upcoming workspace.

Considering Noetis for your product?

Tell us which workflow you would run and where the results would appear. Evaluation access and licensing are arranged per customer.

Discuss licensing