The shared, secure trust and governance substrate for AI agent and human decisions in regulated industries.

The Corvair Knowledge Substrate

Knowledge you can see. Answers you can prove.

Trusted knowledge, governed reasoning, defensible decisions.

A governed knowledge layer that unifies licensed expert content, your organisation’s content, vetted external sources, and each user’s personal files into one trust-scored knowledge base, keeps itself current, and serves grounded, source-cited answers, with a signed record of each, to both people and AI agents.

Four capabilities in one governed layer

It is more than a knowledge base with a chat box. Four capabilities sit in a single governed layer.

One query across licensed, organisational, and personal knowledge, ranked and attributed.

Q&A

Plain-language answers from a local or cloud model, always with citations and a warrant.

Knowledge management

Curated, typed, trust-scored, and source-attributed. Licensed expertise blended with your own knowledge.

Agentic memory

A bi-directional record that pre-loads context and captures what was said, done, and decided.

Agentic memory works in both directions

Before work begins it pre-loads relevant context into the session. After, it captures what was said, done, and decided. Those memories are promoted into the knowledge base as facts and learnings, and distilled into reusable skills, so the system improves with use.

How you use it

A virtual expert answers in plain language, grounded only in the content, with multi-turn follow up. It works in four modes, mixed freely in one conversation.

Search Q&A Browse Deep research

Grounded answers

Cited to the document, page, and site, with each source and site carrying why it is relevant.

The premise chain

The specific facts an answer rests on and the trust tier of each, so the basis is visible.

Browse and filter

Follow the topic graph and ontology, filtered by tag and dimension, to survey what the base knows.

Deep research

A long-running series of queries assembled into a comprehensive, cited report.

Decision support and export

Reach and record a decision, and export results to a NotebookLM-compatible notebook.

Proof for any part

A Validity Warrant can seal a single answer, a span, or the whole conversation.

Proof on every answer

Any answer or decision can be sealed with a Validity Warrant: a cryptographically signed record of the question asked, what was known, how it was reached, who asked, and whether the request stayed within policy. It can cover a single answer or a whole conversation, and is reconstructable later for audit and legal use.

01 Search or Q&A

The question asked

The exact search prompt or question put to the substrate, including any follow-ups.

02 Knowledge

What was known

The facts used, each with its trust score, pinned to a snapshot of the knowledge at that moment.

03 Reasoning path

How it was reached

The trace from question to answer: the search, the sources selected with why each is relevant, and the premise chain.

04 Authority

Who asked, and bounds

The person or agent that asked, and whether the request stayed within policy.

05 Signature

Signed and replayable

Cryptographically signed and timestamped, exportable, and reconstructable on any future date.

Decisions cannot be guaranteed as correct, but the basis for the decision can be proved.

What goes in, and how it arrives

Optionally, licensed content from a reputable domain expert serves as the authoritative foundation. On that foundation sit your organisation’s knowledge, curated external sources, and each user’s personal files.

Licensed · the foundation

Domain expert content

The authoritative backbone, produced and kept current by a named partner. Optional.

Crawled & ingested

Your organisation’s knowledge

Internal sites, SharePoint, product shelf, policies, and procedures.

Curated · external

Vetted external sources

Reliable information and news, watched and kept current.

Self serve

Each user’s personal files

Their own documents, transcripts, notes, and sources.

›

Search · Q&A · knowledge management · memory

One governed substrate

A governed blend across all four inputs, ranked, attributed, and warranted on every answer.

Ingests PDF · DOCX · XLSX · CSV · web pages · call transcripts · meeting notes · YouTube transcripts

Crawler renders JavaScript and reads scanned documents through OCR. Content-hash deduplication keeps the same fact from entering twice.

How content is governed

Every item is typed, not stored as undifferentiated text. Each fact carries a trust score set by its source tier and loses confidence over time unless it is refreshed. New material must clear a veracity gate before it is admitted, never silently.

Trust scoring and decay

Tier	Source	Trust
T1	Institutional, certified	Highest
T2	Credentialed, peer reviewed	High
T3	Self declared, general	Moderate
T4	Speculative	Low, flagged

Each fact takes its trust from its source tier, then loses weight over time unless it is refreshed, so stale knowledge quietly falls in confidence.

A veracity gate, eight dimensions

Before anything enters the base it must clear a minimum on eight quality dimensions. Items below the threshold are queued for a steward, never admitted silently.

Item score Minimum to enter

Typed knowledge, organised into families

Every item is typed, and types are grouped into families. The set is current and will grow. The Reference and Normative families are contributed primarily by the licensed foundation.

Declarative

what is asserted to be true

Verifiable, attributable facts, the premises that build on them, and the hypotheses they lead to.

FactsPremisesHypotheses

Experiential

what has been observed

What the system has seen: memories of events and decisions, declared beliefs that shape how trust is assigned, and vetted cases and precedents.

MemoriesBeliefsCases & precedents

Procedural

how work is done

How the work is done, from skills, SOPs, and instructions to best practices, guidelines, learnings, and reusable templates and clause libraries.

SkillsSOPsInstructionsBest practicesGuidelinesLearningsTemplates & clause libraries

Reference

the authoritative backbone foundation

The authoritative assets the expert reasons over: structured reference data, canonical definitions and taxonomy, framework crosswalks, and models and formulas.

Reference dataDefinitions & taxonomyFrameworks & crosswalksModels & formulas

Normative

what must hold foundation

What an answer must respect: the rules and obligations that bound it, and the risk indicators that flag trouble.

Rules & obligationsRisk indicators

Signal

external, time-sensitive awareness

External, time-sensitive awareness, from news (press releases through analyst coverage) to social-media signal.

NewsSocial media

Collateral

marketing and sales material

Promotional and explanatory material held for reference and comparison, such as product collateral and competitor marketing.

Product marketing materials

How it stays current

Human experts curate the base, supported by AI agents that operate behind the scenes. The base tends itself rather than drifting.

Crawlers and monitors

Watch designated sources for new and updated material as it is released, and bring it in automatically.

Recommendations

Suggest what to add to close gaps, where to expand into adjacent areas, and what to prune when material is duplicative.

Conflict detection

Surface conflicts and inconsistencies within the material so they are reconciled rather than silently propagated.

Knowledge subscriptions

Optional. Subscribe to a reputable domain expert’s content, kept current by the partner and blended into the base as it updates.

For people and agents alike, reached anywhere

The same grounded interface serves AI agents, not only people, so their actions rest on trusted, current content under the same source attribution and warranting.

Reached through

Chatbot Search Secure API MCP server Claude skills Embedded in your applications

Distinct roles, each scoped per knowledge base

Business user

Reads and contributes within bounds: cannot change the shared base directly, but can suggest sources, trigger revalidation, and curate their own files.

Virtual Expert (chatbot)

The plain-language assistant that answers from the governed base, grounded and carrying a warrant.

Steward or curator

Shapes what the base knows: sources, trust, curation, and scope.

Domain partner author

Builds, packages, versions, and updates a licensed domain pack, the foundation that anchors a base.

Compliance and audit

Opens and verifies the Validity Warrant behind any decision.

Administrator

Runs the platform: schedules, retention, and limits. No access to the content itself.

Autonomous agent

Acts through the API under an explicit authority matrix, in bounds.

How it is built and packaged

Knowledge as code

Each knowledge base is defined by a recipe: a versioned, auditable, portable file that declares what the base covers, what it trusts, how it is kept fresh, and how it is reached. Hydrate the same recipe and you get an equivalent, governed base, so a configuration can be rebuilt identically or lifted to a new institution in days.

recipe.yaml

substrate: wealth-management
scope:
  dimensions: [ geography, time_window, evidence_type ]
sources:
  crawl:   [ sharepoint:/policies, intranet/house-view ]
  trust:   { T1: [ regulator, board ], T3: [ general ] }
content:
  licensed: [ partner/wealth-advisory@2026.1 ]
  grounded: ./knowledge/**
freshness:  decay(half-life = 180d)
access:     [ chatbot, search, api, mcp ]
warrant:    every-answer
curation:   { auto_cull_below: 0.2, escalate_T1: true }

From recipe to running base

Author

Declare intent, scope, sources, trust tiers, freshness, access, warrant, and curation policy.

Hydrate

The recipe builds a running, governed knowledge base. A manifest records exactly what was admitted.

Curate

Agents and stewards keep it current. The recipe governs what auto-applies and what escalates.

Promote

Branch and review in Git, then promote a tested configuration and reconstitute it anywhere.

Git repository›Test›Production

Two ways to fill it

Your own content

Run the platform empty and fill it with your organisation’s knowledge and each user’s libraries. No domain partner and no royalty. Best when your own knowledge is the asset.

Subscription content

Add one or more co-branded domain packs, each a prebuilt virtual expert with eight or more named capabilities, licensed reference databases, and how-to guides, kept current by the partner. Banking is one example family (Wealth, Retail, Commercial, Markets and Treasury, Risk and Compliance, Operations); the same pattern serves other regulated industries. One query spans every pack plus your own content.

How it is deployed

Model and cloud agnostic. It runs in your cloud of choice or on your own infrastructure, always inside your firewalls, and the platform, the data, and the audit trail stay yours. The initial build targets Google Cloud.

Your cloud of choice

Deploy on the public cloud you already run on, under your own accounts and controls, and procure through its marketplace.

Google CloudAWSMicrosoft AzureAlibaba CloudTencent CloudOthers on request

On-premises or private cloud

Run entirely on your own infrastructure, inside your firewalls, when data residency or isolation requires it.

On-premisesPrivate cloud

Loosely coupled by design

The governed core and the research-and-monitor side are loosely coupled, and can be co-located on one cloud or deployed separately, for example the core on-premises with research in the cloud. Only a recipe goes out and only staged candidate files come back, so your knowledge never leaves your environment.

Trust boundary

The substrate is the constant beneath whatever models and agents you run. Nothing is taken into custody: your knowledge, your data, and your audit trail never leave your environment. Licensed packs are co-branded and royalty-bearing, but never take custody of your knowledge.

What it delivers

Seconds to minutes

to a sourced answer, instead of hunting across drives and inboxes

Weeks not quarters

for a new joiner to reach competence

Every answer

source attributed and carrying a Validity Warrant

One source

shared by your people and the agents that assist them

Also in the platform

The same governed layer includes further capabilities.

PII detection, redaction, and blocking Agentic discovery, ingest, processing, and curation Internet and intranet discovery Many bases, federated query with synthesis Knowledge as code, prescriptive or discovered

Built on the patent-pending Corvair Knowledge Substrate. Runs on your cloud or on-premises, inside your firewalls, and every answer can emit a signed Validity Warrant.

corvair.ai
Singapore · UAE · United States