Model & Compute — Later capabilities¶
Target
This page describes intended behaviour or structure. No implementation is claimed unless a section is explicitly marked Current implementation or Evidence.
| Capability | Purpose | Trigger | Owner |
|---|---|---|---|
| Compute scheduler | Provision pools on demand through Runtime Cloud and Pulumi under privacy, region, memory and cost constraints (original handoff §14) | self-hosted demand exceeds registered pools | BC-41 with BC-25 |
| Automatic promotion proposals | Propose a candidate that scores better at lower cost for the regression gate | capability suite runs continuously | BC-07 with BC-22 |
| Model failover orchestration per policy | Policy-driven failover beyond per-profile fallback chains | provider incidents in the pilot | BC-07 |
| Real-time token and cost governor | Per-run and per-tenant spend control inside the gateway | FinOps evidence | BC-07 with BC-05 |
| Prompt-cache aware routing | Prefer the deployment holding the run's cached prefix | provider cache metrics available | BC-07 |
| Historical-success weighting | Weight candidates by observed task success per agent and profile | evaluation feed in place | BC-07 |
| Regional capacity reservations | Reserved throughput per tenant for pilot-interactive profiles |
tenant SLO commitments | BC-07, BC-34 |
| GPU sharing with isolation evidence | Share pools across tenants where the boundary allows | cost pressure on tenant-private pools | BC-41 |
| Pool hours as a Factory meter | Bill tenant-private pool time | commercial model extension | BC-41 with BC-05 |