Corvair is a governed knowledge layer that sits beneath both AI agents and the people who make decisions in regulated industries. It unifies licensed expertise, your own content, vetted external sources, and each user's personal files into one trust-scored base, keeps it current, and answers in plain language with a citation and a signed record behind every word.
Enterprise and web search return results and leave you to open, judge, and reconcile them. AI copilots go further and write an answer, but they hide the three things that matter most in a regulated decision. When an answer is good, it is hard to say why; when it is wrong, it is hard to say where.
Whether the search surface covered everything material, so nothing load-bearing was missed.
Whether the documents it found were the live versions rather than superseded ones.
Whether the answer rests on those documents rather than being partly invented, in its sources, its content, or both.
For most people the face of the substrate is the Virtual Expert: the domain assistant that answers from one governed base, grounded and warranted. It looks like a chat box, but what sits behind it is the difference.
The prototype is more than a knowledge base with a chat box. The capabilities behave as one system, over the same governed content, with the same attribution and the same proof, reached identically by people and by agents.
A virtual expert answers in plain language, grounded only in the base. Every answer shows its premise chain, its sources and their trust tier, and a Validity Warrant.
One query runs across everything a user is allowed to see, licensed, organisational, and personal. Ranked, attributed, and matched semantically, so the right knowledge is found even when the wording differs.
A multi-pass investigation that clarifies the objective up front, plans and reasons in steps, and returns a structured, cited report with a warrant.
Follow the topic graph to see what the base knows and where it is thin, filtered by geography, framework, time, or sector, and drill down to the sources behind it.
Monitored news on the subjects you care about, each story tied back to its source and its trust tier, with a digest and a one-tap path into an inquiry.
Each user's own files and saved results, governed under the same model and served only to them unless shared, alongside connected drives and crawls.
Raise a gap, dispute, or help request and route it to a steward, so weaknesses in the base are fixed rather than worked around.
The same governed base, reached the same way and governed identically, is the single source your people and your AI agents share.
Trust is not a feeling. Each source carries a tier, every claim names the sources it rests on, and each answer is accompanied by a signed record of how the question became the answer.
A signed record attached to every answer, in five parts.
The prototype lets you switch persona on the web build to see each surface. Each role sees only what it should.
Asks, searches, researches, and receives cited, warranted answers. Contributes within bounds.
Owns what a base knows: sources, tiers, the curation queue, scope, and the recipe.
Runs the platform: workers, jobs, budgets, access. Never sees content.
Verifies after the fact, reads warrants and framework coverage, and exports.
The way DocuSign made a signature provable and binding, Corvair makes a decision provable and defensible. Whether a person or an agent reaches it, the decision draws on the same grounded knowledge and is sealed in a Validity Warrant: a cryptographic envelope that records what was known, who or what decided, and under what authority, then signs and timestamps it. Open it months later and the decision still verifies.
A relationship manager records a decision in the flow of work. The warrant captures who they are, the policy bounds they stayed within, and the exact knowledge behind it.
An autonomous agent acts through the API under its authority matrix. The warrant records the agent token and the same grounded knowledge, governed identically to a person.
A Corvair base is not configured by hand and left to drift. Cooperating agents keep it current under governance, a steward holds the judgement calls, and the whole base is defined as code, so it is reproducible, portable, and auditable like software.
Finding the problem and proposing the fix are deliberately separate steps, so each is auditable on its own. The agents propose; for anything consequential, a steward decides, and the loop runs continuously.
Cooperating agents, each with a narrow job, run under the same trust, veracity, and PII controls as a person.
Discovery finds sources. Ingest and processing normalise and type content. Problem-detection flags gaps and conflicts. Recommendation proposes fixes with candidates pre-staged. Maintenance revalidates, applies decay, and promotes or demotes items over time.
Low-risk changes apply automatically; consequential ones are staged, ready to approve. The role is judgement, not labour.
Intranet access is explicit, least-privilege, and revocable, granted per repository. Circuit breakers cap activity, and every action, by a person or an agent, is recorded for audit. An autonomous agent can act as a steward through the API, under the same controls.
A base is defined by a recipe: a declarative file stating its intent, scope, sources, trust, freshness, access, and curation policy. Hydrate it and you get a running, governed base; hydrate it again and you get an equivalent one, reproducible and auditable like software.
One deployment can run many bases side by side, each with its own recipe, governance, and virtual expert: a wealth base, a risk base, an operations base. There are two levels of asking across them, and a way to consolidate several into one addressable base.
A single query already spans the licensed, organisational, and personal content that base holds, ranked and attributed, with licensed material kept distinct from local policy on every answer.
A query is posed to several bases at once, answered by each base's own expert from its own governed content, then synthesised into one response with each base's contribution attributed. Governance is not merged; the answers are composed.
A composite groups several bases from one environment into a single addressable base, queried together under a chosen strategy (per base, synthesised, or best answer) with a conflict-resolution policy, while each member stays governed in its own right.
A domain expert's knowledge becomes a co-branded domain pack: a prebuilt virtual expert with eight or more named capabilities, licensed reference databases, and how-to guides, produced and kept current by the partner and licensed to institutions on subscription.
Define a recipe and hydrate a governed base. It ships as a prebuilt virtual expert, riding on the same substrate every customer already runs, so there is nothing new to deploy.
Each pack is licensed on a recurring subscription. The partner curates and ships versioned updates that recombine automatically, so every subscriber stays current without a redeployment.
A single query spans every pack and the institution's own content, with licensed material kept distinct on every answer. The platform, the data, and the audit trail stay with the institution.
These are interactive prototypes. The data is illustrative, but the flows, the governance signals, and the proofs are real. Open any one and click around.
The complete surface across Inquiries, Topics & News, Vault, and Cases, with the Steward, Administrator, and Auditor workspaces behind the persona switcher.
The Application-user experience refactored for mobile, at functional parity with the web for everyday use.
The building blocks behind both apps: tokens, controls, trust and status, the warrant timeline, and the mobile patterns.
