Organisation, projects and access
Organisation, project, project domain, environment, API keys and scopes, roles, and sign-in.
Organisation
What it is. Your company's account. It holds the members, every project, usage and billing, and the policy audit log.
Why it matters. People, plan and audit history are managed once, not per project.
How it relates. An organisation contains projects. Each member has an organisation role, which sets their default role in every project.
Where you see it. Administration › Organization (Usage & billing, Projects, Audit log) and Administration › Members & roles. See Organization.
Example. Northwind Insurance has three projects: claims-assistant, benefits-chat and finance-sql-analyst.
Project
What it is. One AI application. It holds that application's answers, its evaluation policy, datasets and runs, experiments, reviews, alerts and API keys.
Why it matters. A score only means something against one set of rules. One application per project keeps its policy, history and keys together.
How it relates. Creating a project creates its evaluation policy (v1). An API key belongs to exactly one project.
Where you see it. The project switcher, Getting started › Create a project, and Administration › Project configuration.
Example. benefits-chat answers members' questions from Summary of Benefits documents; its answers never mix with finance-sql-analyst's.
Project domain
What it is. The subject area of the documents your AI answers from: Automatic (by each answer’s topic), Finance, Healthcare, Insurance, Legal or General.
Why it matters. The domain picks the grounding models used to check claims against your sources. Automatic chooses per answer, by its topic.
How it relates. Applies to new answers; it does not change the policy or past verdicts. Policy admins and owners can change it.
Where you see it. Step 1 of Create a project, and Project configuration › General.
Example. A benefits chatbot answering from plan documents: Insurance. A text-to-SQL agent over the general ledger: Finance.
Environment
What it is. A label an answer is logged with, such as prod, staging or ci. It is a label, not a separate policy.
Why it matters. You can filter answers by it and use it in scoring rules, collection rules, queue rules, alerts and the Max-Accuracy scope.
How it relates. One policy version is in force for the whole project; an environment never chooses a different policy. observe(), log(), the Google ADK and LangGraph integrations (through variably.context()) and MLflow's log_traces() set it; OpenTelemetry answers arrive without one.
Where you see it. The All environments filter in Live events, the Details tab of an answer, and rule conditions under Evaluation policy › Scoring rules. Environments in Project configuration are for feature gates.
Example. client.log(q, a, environment="prod"); a rule "Environment is prod → Max-Accuracy 100%" then double-checks production only.
API keys and scopes
What it is. A secret your app or CI job sends with every call. It belongs to one project and carries scopes: what it may do.
Why it matters. Answers land in the key's project, and a key can do only what its scopes allow, so a logging key can't change your policy.
How it relates. A new key gets Log answers and Read experiments and results. Policy admins and owners create and revoke keys. Revoking a key, or deleting its project, stops it at once. The key is shown once.
Where you see it. Project configuration › Integration & keys, and the Connect step of a new project. See Connect your app.
Example.
| Scope | On screen | Lets the key |
|---|---|---|
events.write | Log answers | Send answers to be scored |
experiments.read | Read experiments and results | Read scores, dataset runs and gate results (CI) |
experiments.write | Change experiments | Create, start and stop experiments |
policy.write | Change the evaluation policy | Push and publish the policy; the SDK settings calls |
flags.read, flags.write, gates.read, gates.write | Read / Change feature flags and gates | Feature management from code |
analytics.read | Read analytics | Read analytics and usage figures |
provisioning.write | Create API keys | Mint further keys (trusted provisioning jobs only) |
Roles
What it is. What a person may do. Each member has an organisation role (Owner, Admin, Member, Viewer) that gives a default project role (Owner, Policy admin, Editor, Viewer). A Project role override changes it for one project.
Why it matters. Changing how every answer is scored, spending, and deleting data are kept apart from day-to-day work.
How it relates. A button your role can't use is disabled with the reason, for example "Policy admins and owners can publish. Your role in this project: Editor."
Where you see it. Administration › Members & roles. See Members & roles.
Example.
| Role | Can |
|---|---|
| Viewer (from org Viewer) | Read everything in the project |
| Editor (from org Member) | Edit the policy draft (rules, weights, bars, data definitions, YAML) and discard it; datasets, collection rules, alerts, human review verdicts and queue rules; experiments |
| Policy admin (from org Admin) | Also publish the policy, re-score, create and revoke API keys, change the project name and domain, set project role overrides. Org Admins also invite and remove members and create projects |
| Owner | Everything, including deleting a project, the organisation's single sign-on and plan changes. Can't be overridden per project |
Sign-in and staying signed in
What it is. Sign-in is by your organisation's single sign-on or a listed provider (Google, GitHub, Microsoft, LinkedIn). There are no passwords. People join by invitation only.
Why it matters. Access follows your identity provider. Removing a member takes effect immediately, including reading.
How it relates. Your session renews itself in the background while you work, so you stay signed in. A "your role can't do this" refusal shows on the page and does not sign you out.
Where you see it. The sign-in page; Members & roles for invitations.
Example. An adjuster invited as Viewer signs in with the company's single sign-on and sees the claims project read-only.