Skip to main content
You are reading the docs of Galtea 5.15. Read this page in the latest version.
Galtea uses role-based access control. Every member of your organization has one role that decides what they can see and change. This page lists the four organization roles and the actions each one can perform. Actions cover the standard operations (Create, Read, Update, Delete) plus a soft delete (a logical delete that hides a record without removing it), and one Assign action per role a user may be given. Many permissions are conditional, shown as a short rule in the cell.

Roles

  • OWNER: Manages users and resources in their own organization. Can permanently delete only Sessions and User Groups; every other entity is soft delete at most. Only OWNER may create, edit, or delete a User Group, or change its users or products.
  • MEMBER: Day-to-day access in the products their user groups grant: view and contribute, with limited edits. Creates a product only with a user group that grants all products and allows product.create. Can permanently delete only Sessions; every other entity is soft delete at most. Reads and edits every metric and reads every User Group of the organization. Cannot create, edit, or delete a User Group, or change its users or products. May link or unlink a group of the organization to a metric it may edit.
  • HUMAN_EVALUATOR: Read-only on the products their user groups grant, plus may annotate AI metrics and submit human evaluations on them. Can also update their own account settings (name, consents and MFA method).
  • READER: Read-only on the products their user groups grant. Cannot create, update, or delete anything except their own account settings (name, consents and MFA method).
Each cell shows what a role can do for that action. ✅ means allowed and ❌ means not allowed. A short rule (for example, “If product in own org”) means the action is allowed only when that rule is true at the time, checked against your organization and the record’s ownership.

Permissions by area

Identity & Organization

User An Assign action applies only together with Create or Update: a role gives a role only to a user it may invite or edit. A MEMBER invites no one and edits only itself, so its Assign Member role has no effect. Organization An OWNER may update only the credit warning threshold, MFA enforcement, the SAML SSO settings, and the OpenTelemetry settings. An OWNER cannot change the organization name. Subscription User Group

Test Assets

Product The Tutorial and Configuration entry rows only control what the dashboard shows for this product (the sidebar entry, its page and its list URL): the API still returns this data to every role that can read the product. Version The Versions entry row only controls what the dashboard shows for this product (the sidebar entry, its page and its list URL): the API still returns versions to every role that can read them. Dataset The Datasets entry row only controls what the dashboard shows for this product (the sidebar entry, its page and its list URL): the API still returns datasets to every role that can read them. Test Case Specification

Connections

Endpoint Connection Phone Connection WebRTC Connection

Evaluation & Results

Session Evaluation Human evaluation also requires membership in a User Group linked to the metric. A HUMAN_EVALUATION metric with no linked group reaches nobody. Annotating an AI metric has no group check. One gate covers the whole annotation, so anyone who may annotate may also edit or remove an annotation another person wrote. The table shows the role-level gate only. The evaluation metadata row only controls what the dashboard shows: the API still returns the metadata to every role that can read the evaluation, which for a MEMBER, READER or HUMAN_EVALUATOR means only on the products their user groups grant. The inference metrics row is enforced by the API: a role without it gets null judge tokens and credits on evaluations, null tokens, latency, cost and credits on traces, a null aggregated latency and total cost, and null spent and pending monitor credits (the monthly cap stays visible), and null cost per session and total cost in the organization products table, from the dashboard, the SDK, the CLI and the assistant alike. The API also refuses the cost summary of the organization products page and of a product’s Analytics Summary to such a role with 403. Span latency (Spans tab, trace page) and the Analytics version summary (latency, tokens and average cost) stay visible. Trace Span Comment Free-form reviewer text on a trace. Access also requires being able to read the parent trace: it must be in your organization, and for a MEMBER, READER or HUMAN_EVALUATOR in a product your user groups grant. Deleting is a soft delete; you may delete your own comments, and organization owners may delete any. Monitor Run

Reference Data

Metric Model

Analytics & Insights

Analytics The Read row is analytics.read. Without it, the dashboard hides the product’s Analytics section, a route for one product answers 403, GET /analytics/summary answers an empty row for that product, and GET /analytics/products returns the product’s row with empty stats. The summary boxes of the organization products page count only the products with analytics.read, and answer 403 when the request names another. The Coverage row (analytics.readCoverage) only controls what the dashboard shows on the Analytics page: the API still returns this data to every role that can read Analytics. The Inference row follows the inference metrics permission of Evaluation: the API returns the aggregated latency and total cost as null to a role without it. The latency, token and average cost summary of the Analytics payload stays visible.

Dashboard

Dashboard views These rows decide what the dashboard shows or hides for a role. They do not grant access: the platform still checks the permission behind each action.

Permission list ids

A user group’s permission list may hold only these ids. The role still caps each one. inferenceResult is a Trace in the dashboard, and trace is a Span.

Notes

  • Read-only roles: READER and HUMAN_EVALUATOR read what a MEMBER reads, except that metrics are limited to the ones the products granted by their user groups use (all of them for a group linked to all products whose list allows product.read); the dashboard also hides the Coverage section of Analytics, and the Datasets, Versions, Tutorial and Configuration entries of the product sidebar, from them. They cannot create, update, or delete, except their own account settings (name, consents and MFA method). HUMAN_EVALUATOR additionally may annotate AI metrics and submit human evaluations, on those products only; READER cannot. The dashboard also hides from READER the evaluation metadata on an evaluation page, and the judge tokens and judge credits of every evaluation list, the inference metrics of the evaluation, session and version pages, the Inference section of Analytics, the latency and cost columns of the versions list, the cost column and the cost box of the organization products page, the cost box of a product’s Analytics Summary, the monitor credits, and the metrics tab of the Trace card; the API returns those figures to READER as null, refuses the cost summary of the products page with 403, and refuses a READER’s sort on them with 400. Span latency and the Analytics version summary (latency, tokens and average cost) stay visible.
  • Permission lists: An OWNER or ADMIN can give a user group a list of product permissions, each written entity.action, for example analytics.readCoverage. The list applies to the products the group links. A MEMBER, READER or HUMAN_EVALUATOR may do an action on a product when the role allows it and at least one of the user’s groups that reaches the product lists it, so the role caps the list: a list entry above the role does nothing. A list never limits an OWNER or ADMIN. A group with no list grants every permission the role allows, and an empty list grants nothing. A user in no group reaches no product, and lists answer an empty page. When every group of a user leaves out test.read, GET /tests answers 403; the same holds for monitor.read and GET /monitors. Other lists answer an empty page. A list that holds any entry needs product.read, an action on an entity needs that entity’s read, and evaluation.readInference needs inferenceResult.read. Only actions on entities that belong to a product come from a list; the Permission list ids table shows which. Every other action stays with the role.
  • Product access: A MEMBER, READER or HUMAN_EVALUATOR reaches a product, and everything inside it, only when one of their user groups links it. A MEMBER may also create and edit inside those products. A user group with the all-products setting grants every product of the organization, current and future. A user in no group reaches no product. Only a user in a group with the all-products setting whose list allows product.create (a group with no list does) creates a product. When a user who cannot create a product reaches exactly one product, sign-in opens that product instead of the product list. It lands on Analytics, or on the first section the user may open when Analytics is not allowed. Records with no product, such as a global dataset, stay readable. When read by id, a product outside the granted set answers like a record of another organization. On the dataset list, naming an ungranted product answers 403, and naming another organization’s product returns no dataset. In a user group, their product_ids list only the granted products. For a READER or HUMAN_EVALUATOR, metrics follow the products: unless one of their groups links all products and allows product.read, a metric is visible only when a granted product evaluates it, links it to a specification, or monitors it, and every revision of that metric is visible then. Any other metric answers like a metric of another organization. A MEMBER reads and edits every metric of the organization, but cannot delete a metric that has results or specification links in a product where their user groups together do not list both evaluation.softDelete and specification.update: the API answers 403. A metric’s specification_ids list only the specifications of granted products, and a user group’s metric_ids only the visible metrics. The analytics impact of excluding a metric revision, and the data metric optimization learns from, count only the granted products. An OWNER is not limited this way.
  • Organization scope: “In own org” means the record (or its parent, such as a Product) has an organizationId equal to the signed-in user’s organization. This keeps each client organization’s data separate.
  • Parent scope: For nested records (Version, Dataset, Evaluation, and so on) access is derived from the parent’s organization. Example: a Version belongs to a Product, and the Product belongs to an Organization. An Evaluation is reached through session to version to product.
  • Global records: Some records (Metric, Model, and a Dataset with no product) can be global, meaning not tied to one organization. Global records are readable by every role in addition to their own organization’s records, except that a READER or HUMAN_EVALUATOR in no all-products group sees a global metric only when a granted product uses it.
  • Hard delete: Each entity table has two delete rows. Soft delete hides a record without removing it. Delete is a permanent removal, shown as allowed for Session (OWNER within their own organization, MEMBER within the products its user groups grant) and for User Group (OWNER only, within their own organization); for every other entity it is denied to all roles listed here.