Agent Systems Lab ยท Synthetic study SIM-27B
Architecture and governance note
The model separates observation, proposal, arbitration and approval so local optimizers cannot silently commit global changes.
Data status: Synthetic scenario; no production systems or real agents are represented.
Roles in the simulation
- Latency planner: proposes placements that improve workload latency.
- Capacity planner: preserves headroom and avoids placing beyond zone ceilings.
- Reliability planner: holds a fault reserve and checks failure-domain constraints.
- Cost reviewer: evaluates the shared spend envelope and inter-zone transfer cost.
- Arbiter: compares proposals against one versioned snapshot; it does not invent missing capacity.
Coordination contract
Every proposal names its input epoch, affected workload, requested capacity delta, expected latency effect and rollback condition. A proposal is stale when its epoch differs from the current snapshot. The arbiter re-evaluates stale proposals; it does not merge them against an old view.
Agents can act alone only for reversible, in-zone adjustments that remain within the published per-zone ceiling, global cost cap and fault reserve. Cross-zone movement, reserve consumption, quorum changes or conflicting proposals require arbitration. If a required measurement is missing or the state is ambiguous, the agent should request validation rather than infer capacity.
Objective ordering
First preserve safety constraints and the minimum fault reserve. Then meet the interactive latency objective where feasible. Optimize cost and batch throughput only among allocations that satisfy those constraints. This ordering prevents a weighted local score from trading away a global invariant.