Model & Compute — Workflows¶

Target

This page describes intended behaviour or structure. No implementation is claimed unless a section is explicitly marked Current implementation or Evidence.

Profile activation¶

A model profile is a Published asset (BC-04, D32): Draft → InReview → Published with evaluation evidence and Platform-admin approval. ModelRouterService consumes AssetReleasePublished and activates the profile version; deprecation and revocation retire it. Runs pin a profile version at acceptance and keep their route plan even if a newer version activates.

Deployment lifecycle (ModelOps)¶

sequenceDiagram
    participant PA as Platform admin
    participant R as ModelRouterService
    participant Q as QualityScoreService (BC-22)
    PA->>R: RegisterDeployment (provider, model, region, license metadata)
    R->>R: Candidate
    PA->>R: EvaluateDeployment
    R-)Q: capability suite run requested
    Q-)R: EvaluationCompleted (capability-suite record)
    R->>R: CapabilityEvidenceRecorded, Evaluated
    PA->>R: ApproveDeployment (publication-class action, approval required)
    R->>R: Approved, DeploymentStateChanged
    loop every 2 minutes
        R->>R: health probe, Degraded or Approved
    end

Hold "Alt" / "Option" to enable pan & zoom

Thresholds are the quality gates' (tool selection ≥ 97 %, valid arguments ≥ 99 %, structured output ≥ 99.5 %, multi-turn ≥ 95 %, instruction compliance ≥ 95 %, patch compile ≥ 90 %) and a deployment may not lower task success or raise tool-error rate against the one it replaces.

Route-plan resolution and invocation¶

Step	Who	What
1	AgentRuntimeService	`ResolveRoutePlan(run, profile)` once per run
2	ModelRouterService	policy decision (privacy class, region, tools), candidates filtered and scored, plan persisted
3	AgentRuntimeService	`Invoke(request hash, idempotency key, budget)` per model call
4	ModelRouterService	credential brokered, first candidate called with streaming; equivalent fallback only if the profile enables it
5	ModelRouterService	receipt with actual deployment, outcome, usage status; `RecordUsage` to Usage Accounting
6	ReceiptReconciler	`Unknown` and `Estimated` usage confirmed from provider statements

Embeddings and reranking follow the same path through Embed and Rerank with embedding profiles whose compatibility pins a knowledge index to the profile version.

Degradation (XL-16)¶

A degraded deployment is skipped by the router without changing the plan's constraints; when no healthy candidate remains the invocation fails with class transient and the attempt retries within its budget. Nothing switches privacy class to keep a run going.

Compute placement and warm state¶

sequenceDiagram
    participant R as ModelRouterService
    participant C as ComputePoolService
    participant P as Pool agent
    R->>C: Place (deployment, memory, privacy, region, price)
    C->>C: hard filters, score warm, price, utilization, latency
    C-->>R: PlacementDecided with pool endpoint reference
    P->>C: ReportUtilization (stream)
    C->>C: WarmKeeper loads or evicts models per policy