Model sources for scenarios and CTFs¶
Current support: ADR-068 parks the broker and its controls. Keyless-direct is the supported guest model path. The broker-dependent user flows below describe retained implementation, not an available supported service. The replacement evaluation records the selected LiteLLM OSS replacement with deployment-owned rotation; it does not enable these controls.
The retained sharing API reports that the feature is unavailable before reading membership or publishing changes. An enabled runtime setting does not override the parked status. Independently supported source administration and the keyless-direct model path retain their existing behavior.
A model source connects an approved model to an account that pays for its use. It is independent of where the range or BigRAE runs. A Vertex source may use BigRAE's own GCP project, another GCP project, or a model account elsewhere. The broker handles provider authentication; participants receive only scoped, revocable access for their range.
These controls require an enabled model broker and an approved scenario model profile. A source marked Configured has been stored successfully; that label is not a live provider test or a guarantee of quota. Deployment qualification is tracked separately in the operations guide.
Connect a source¶
Organization administrators open Administer → Model sources, choose an organization, and select Add model source. Choose the provider, model and account identity, then enter the context limit, application token ceiling, prices and price expiry. Cloud sources also require their explicit region and invocation identity. Workload identity uses the broker's approved cloud grants; otherwise supply a provider credential in the form. Secrets are write-only and are cleared after every attempted submission. An edit without new credentials retains the current credential.
Source registration does not give anyone permission to spend. Explicitly allow organization members or select individual users. A source is selectable only while the user also belongs to that tenant. Administering sources and being allowed to use their funds are separate permissions.
| Source | Configuration and behavior |
|---|---|
| Vertex | Approved project, service account, model, invocation region and counting region. The platform project is a valid choice. |
| Bedrock | Approved AWS invocation role, model and region. The hosting cloud need not be AWS. |
| Direct Anthropic | API key and pinned model at the provider's fixed API origin. |
| OpenAI | API key and pinned model; the broker translates the supported text and local-tool protocol to Responses. |
| OpenRouter | API key, model and one explicit upstream provider; automatic provider fallback is disabled. |
Direct API sources use provider-managed geography; they do not promise a cloud region. Both deployment policy and the scenario must allow that geography. OpenRouter does not expose token counting through this integration. It requires a scenario input allowance covering the source's full context window; scenarios requiring token counting cannot use it. Unsupported choices are omitted from the scenario picker. Hosted provider tools, arbitrary API URLs and platform administration tools are not granted by selecting a model source.
The token ceiling is an application limit, not purchased provider quota. Sources using the same cloud account and region share capacity accounting; renaming them does not create capacity. For direct providers, use the same quota identity for sources that share a provider quota. Prices use micro-units of the selected currency per million tokens. A zero price explicitly means no monetary charge is reserved for that component; it does not disable token or rate limits. An expired price prevents new admission.
Choose sources at launch¶
Select a workspace and scenario in the range launch form. For each model alias, keep the deployment default or select an authorized source. Where the scenario and deployment allow weighted allocation, select several sources and their relative weights. For example, weights 2 and 1 express a preferred two-to-one distribution across many allocations, not an exact participant count.
Allocation checks capacity and policy before launch. Source choices do not replace shared event/user budgets, per-range limits, allowed regions or provider pool restrictions. A retry preserves the original launch intent. The broker never retries a potentially billed request on another source automatically.
Configure a CTF¶
The event create/edit form provides the same model-source picker. An organizer with event configuration permission and source-use authority selects sources for the event's aliases. Participants use this event sponsorship without receiving the organizer's credentials or source administration permission. If no explicit model-demand declaration is supplied, admission derives it from the scenario's verified needs and the event's participant ceiling.
With independent weighted routing, the event reserves capacity across the selected accounts in proportion to their weights, rounded to whole participants. Every account's share must fit before the first launch. Accounts sharing the same provider quota still count against that single limit. A policy that keeps the whole group on one source requires that source to support the whole group.
Changing event defaults applies to subsequent preparation and fences affected authority. It does not silently rewrite allocations already running. Use the range administration controls for explicit changes to existing ranges. The event's workspace remains fixed after creation.
The event detail page shows a model-access capacity summary: the assessed outcome, whether it blocks admission, and bounded reason codes. This is planning information only—raw quotas, usage figures and account identifiers stay operator-only, and an unavailable assessment is shown as such rather than as a positive decision.
Share model access across ranges¶
Operators and organizers open Administer → Model-access sharing to share a model profile, provider identity, capacity or budget across more than one range. Choose separately which ranges a binding covers—explicit ranges, an event/cohort/team, a user's ranges, a group, workspace or organization, a named collection, or (operators only) every range—and which facets to share. Each unshared facet stays per-range. Membership can be a fixed snapshot or dynamic.
Preview reports the matched ranges before anything is published; the effective policy preview shows overlapping bindings, their revisions and any priority conflict for a range. Publishing compares an expected definition revision: if the binding changed since it was loaded, the save is rejected and the form asks for a reload. Publication re-resolves membership and authority on the server, records the real publisher, and issues a distinct revocable grant per range—sharing a provider account never hands every participant the same token. Draining a binding stops new use and advances its fence; it never refunds spend, resets an account or cancels a billable request.
These controls express configurations such as:
| Intent | Result |
|---|---|
| Every range uses one provider account, with individual spend limits | Shared provider identity; per-range spend accounting. |
| A set of ranges shares everything | One profile, shared assignment, provider identity, capacity and spend/rate/concurrency; distinct revocable grants. |
| One user's ranges share a budget across two events | One persistent user budget applies to every qualifying range; each event's own limits also apply. |
| A group shares its main model but keeps separate small-model budgets | Shared assignment for the main alias; per-range assignment and budget for the small alias. |
| A cohort shares capacity and spend, with individual concurrency caps | One cohort capacity/spend pool plus separate concurrency ceilings. |
| Two events share a model pool but keep their own budgets | Both bindings reference the same routing/provider pool; each event's budget stays distinct. |
Publishing a binding requires authority over the whole selection; it never grants new IAM, event or workspace authority, and it cannot fund a group the publisher does not already control.
What participants see¶
A participant's range page shows only its own model access: an availability state (active, refreshing, or not available) and the logical model aliases the range may use. It never shows provider or account identifiers, regions, source coordinates, rosters, other members' usage, or shared-pool balances.
Change an existing range¶
Under Model sources → Range model sources, choose the range and edit its assignments. Save compares the revision displayed in the form; if another administrator changed it, reload before trying again.
The change revokes old grants before attempting replacement admission. Existing spend and unresolved provider charges remain. A successful admission reports refresh pending until the guest exchanges its refresh credential. It keeps the original range operation and hard expiry. If admission fails, the range remains blocked and the page displays the reason; retry after correcting the source or capacity. Failure does not reactivate the old grant.
Revoke model access fences the range's current grants and dispatch leases without changing its source selection. It is the immediate control when access must stop. Revocation records the acting administrator, and outstanding usage and unresolved charges stay accounted for under their original prices; it does not refund spend or prove the provider stopped a request already in flight. Re-enrolment is the ordinary source change (renew) above, which delivers a fresh grant to the guest over the range's existing operation path—a credential is never returned to the browser.
Disabling or rotating a source fences dependent grants too. Retire unused credentials deletes obsolete stored versions only after the grace period and when no active allocation or unresolved usage needs them. Current credentials remain, including for a disabled source that may be re-enabled. Failed cleanup can be retried from the same tenant page.
For identity, accounting and provider extension details, see tenant source management.