Provider and private model routing
Govern Luna, Terra, and Sol model assignments across approved LLMGateway routes and encrypted private endpoints.
Veriom separates model choice from evidence access and action authority. A workspace route selects how a bounded role is executed; it does not expand the evidence that role can retrieve or the actions it can perform.
Three reasoning tiers
| Tier | Default role | Typical workload |
|---|---|---|
| Luna | Fast bounded work | Routing, formatting, classification, and citation verification |
| Terra | Specialist work | Architecture analysis, security review, and remediation proposals |
| Sol | Escalated analysis | Explicitly approved decisions that justify the largest reasoning and cost envelope |
Owners and Admins choose a routing posture, route each tier, set an optional reasoning budget, and decide whether a private-endpoint failure may fall back to the platform gateway.
LLMGateway routes
The model inventory is retrieved server-side with the platform credential and returned without secrets. Inventory is not approval: workspace policy only sees live, tool-capable models that are in Veriom's approved catalogue. A selected model carries an explicit provider namespace, model identifier, capability metadata, and published token pricing where available. Production usage records the actual model, token counts, cache use, provider request identifier, latency, and measured cost.
Qualification gates
The released Luna, Terra, and Sol routes form a frozen baseline. A new gateway model is admitted only after an operator submits a versioned replay export and every deterministic gate passes:
| Gate | Default admission threshold |
|---|---|
| Replay coverage | At least 5 unique cases |
| Answer quality | Mean score of 0.85 or greater |
| Provider retention | No replay retained for more than 30 days |
| Latency | p95 no greater than 30 seconds |
| Tool discipline | Every used tool must be present in that case's allowlist |
| Cost | Mean replay cost no greater than $0.25 |
Veriom stores the replay-set hash, policy hash, aggregate measurements, individual gate decisions, evaluator identity, and expiry. It does not store replay prompts, outputs, or evidence content in the qualification record. Results expire after 30 days by default. A later rejection or expiration removes the model from workspace catalogues even when the provider still advertises it.
The operator dashboard shows the provider inventory and catalogue admission state separately. This makes a newly available provider model visible for evaluation without silently making it selectable in production.
Private endpoints
A private endpoint must be HTTPS and OpenAI Responses API compatible. The workspace supplies a name, base URL, allowlisted model identifiers, and write-only API key. Veriom encrypts the credential using the platform envelope key with workspace and endpoint associated data. The key is decrypted only inside the worker for a selected run and is never returned by the API or exposed to the operator plane.
If fallback is disabled, an unavailable or invalid private route fails closed. If fallback is explicitly enabled, the run may use the configured platform tier and records that route decision.
Policy and audit history
Routing is versioned inside the workspace policy. Private routes must reference an active endpoint and an allowlisted model. Every endpoint creation, credential rotation, activation change, and routing-policy update creates an immutable audit event.
Enterprise identity and onboarding
Add MFA and tenant-scoped OIDC, link every sign-in to one canonical user, and resume setup without losing security context.
Model resilience and safe recovery
Route qualified model work through explicit fallback, circuit recovery, non-publishing shadow evaluation, and immutable checkpoints.