The everyday user reads and asks. Behind them, a set of roles makes those answers trustworthy and the base maintainable: a steward shapes what the base knows, an administrator runs the platform without seeing its content, an auditor verifies after the fact, and a domain partner packages expertise others subscribe to. Access is relationship-based and scoped per base, and each role sees only the surface built for how it works.
The same governance that makes an answer provable also splits the work, so authority is bounded, every action is recorded, and the people who run the platform are not the people who can read what it knows.
Searches, asks, and runs deep research, browses the knowledge bases they are allowed to see, and uploads their own private material to a personal vault. Cannot change the shared base.
Directs the scanning of internal and external sites, manages the organisation's content, approves recommendations, and directs the agents and reviews their output.
Runs the platform and its agents, sets budgets and access, and promotes recipes across environments. Never reads the knowledge content.
Verifies decisions after the fact and surfaces disclosure gaps. Reads warrants and coverage, exports for a regulator, and curates nothing.
Authors and packages licensed domain expertise, trust-scores and versions it, and ships updates that recombine into every subscribing base.
Acts through the API within the authority its role is granted, escalating consequential actions, with every action recorded like a person's.
The owner-operator of a base. The steward directs the scanning of internal and external sites, manages the organisation's content, approves the recommendations the agents bring, and directs those agents and reviews their output. They own the base's sources, trust tiers, declared scope (what the base is meant to cover, against which gaps are measured), beliefs (its settled positions on contested points), and the recipe. The role is judgement, not labour: the agents stage the work and the steward decides.
Declared, observed, and available coverage on each dimension, so thin spots are visible before anyone hits them.
Claims sized by epistemic gravity (how much rests on a claim, so being wrong about it would cascade through the answers built on it) and coloured by trust, so the load-bearing sources stand out.
Approve, hold, or re-source what is waiting. Gaps and conflicts arrive with candidate sources already pre-staged.
Which sources have moved, gone dark, or degraded, flagged by validity alerts before coverage rots unnoticed.
Set the declared scope and manage beliefs. Untrust a premise or revise a belief and see the cascade before it commits.
The recipe as code on one side and as a navigable object on the other, the single declaration of what the base is.
The platform owner. The administrator runs the machinery, workers, jobs, agents, budgets, and access, and promotes recipes across environments, but has no path to the knowledge content itself. The console shows what is running and what it costs, with the controls to keep the agents in bounds, never what any base actually knows.
The running machinery at a glance. Heavy stages such as discovery and ingestion can run as multiple instances, scaled to load.
Per-agent controls, autonomy levels, and circuit breakers (automatic cut-outs that cap activity the moment it runs hot).
Resource consumption and spend, with budgets set per base and per agent so cost stays predictable.
When jobs run, how often, and how long data and logs are kept.
Who can reach which base and in what role. Access is relationship-based and scoped per base.
Move a recipe along governed promotion paths, from sandbox to UAT to production, with the right approvals.
Second and third line. The auditor verifies decisions after the fact and surfaces disclosure gaps, and does not curate. The work is to read, verify, and export: confirm coverage against a framework, inspect the signed trail behind any answer, and hand a regulator a warrant they can check independently.
Coverage against a reporting framework, requirement by requirement, surfacing the disclosure gaps that need attention.
Every answer's signed, timestamped warrant as an inspectable, exportable record, ready to hand to a regulator.
Where admitted sources disagree, mapped so a reviewer can see and reconcile the tension.
How validity holds across the base and over time, replayable as of any date.
Export a warrant or a coverage report for inspection outside the system.
Verify, do not curate. The auditor has no write actions on the base, so the line stays clean.
The partner-side author who builds, packages, and maintains a licensed domain pack, the authoritative foundation a base can be built on. They produce the pack's named capabilities, reference databases, and how-to guides, trust-score and version it, and ship the periodic updates that keep it current. They work on the pack itself, never on any one institution's live base, and their releases recombine automatically into each subscribing base without disturbing local content.
Shape the pack's sources, structure, and trust in a steward-style control room scoped to the pack, not to any subscriber.
Ship a prebuilt virtual expert with eight or more named capabilities a subscriber can call from day one.
Bundle the licensed reference data and how-to guides that make the expert authoritative in its domain.
Tier the content and cut a dated release, so every subscriber knows exactly which version they are on.
Push periodic updates that recombine automatically into each subscribing base, without a redeployment.
The partner works on the pack, never on a subscriber's content, so the institution's data and audit trail stay theirs.
A base is not configured by hand and left to drift. It is declared by a recipe: a versioned file stating its intent, scope, sources, trust, freshness, access, and curation policy. That one decision is what makes knowledge portable, packageable, and auditable like software.
Promote the same recipe from sandbox to UAT to production. Because each environment hydrates from one declaration, what you tested is what you ship, and the promotion is governed and recorded rather than a manual rebuild.
A partner authors a recipe and ships it as a domain pack. Many institutions hydrate the same pack against their own entitlements, so one body of expertise reaches all of them without anyone hand-rebuilding it.
Hydrating a recipe (building the live base by collecting and admitting what the recipe points to) produces a running, governed base and records a manifest of every source collected, every item admitted, and every exclusion with its reason. Hydrate the same recipe in a different environment and you get a compatible base, current as of what that environment can reach.
A recipe can be derived from a base that already exists, reading its scope, sources, and structure back into a declaration that would reproduce it. This brings an ungoverned base under governance, or forks one whose recipe was never written down.
A user asks within bounds, a steward shapes under preview-then-commit, an auditor verifies but never curates, an administrator runs the platform but cannot read it, a partner packages expertise without touching a live base, and an agent acts only within the authority it is granted. Every action, by a person or an agent, is recorded.
