The optimal scheduling decision depends on constraints that haven’t emerged yet.

    Working with a small batch of infrastructure operators.

    +724% avg SLA-safe goodput / $·−84% GPU-hours·~1.5M replayed requests · vs production scheduler
    01Hypothesis

    Most infrastructure waste is created because schedulers optimize before future constraints are visible.

    Power prices change. Queue pressure changes. Capacity changes. Deadlines tighten. The decision that appears optimal now can become economically suboptimal later. Aurelius tests whether forecasting those future constraints before execution produces measurably better economic decisions.

    02Architecture

    Forecast. Simulate. Rank by Economics.

    Conventional schedulers optimize the cluster they can observe. Aurelius uses a predictive world model to forecast the constraints they cannot yet observe, simulate candidate decisions, and rank those decisions by economic outcome before execution.

    Scheduling starts too late when it only reacts to the present cluster state. A decision that appears optimal now can become expensive later when power prices rise, queue pressure increases, capacity disappears, or deadlines tighten.

    FIG.01Forecast · Simulate · Decide
    Decided
    Current state
    Cluster telemetryWorkload queueCapacity stateSLA targets
    World model
    4,723,920possible decisions
    Selected planHighest projected goodput / $SLA constraints satisfied
    +724%

    average SLA-safe goodput per dollar vs a production scheduler

    Pareto-safe at ~84% fewer GPU-hours. Mean of +698% / +718% / +755% across PJM, ERCOT, and CAISO. Uncapped replay of public production traces (~1.5M replayed requests). Simulated replay, not a production deployment.

    03Evidence

    Backtested on public production traces.

    Validated through offline replay and read-only shadow mode before any rollout, measured against the operator’s own scheduler.

    • Offline replay
    • Read-only shadow mode
    • Uncapped high-load replay

    Replay evidence. Results depend on workload mix, constraints, and scheduler baseline.

    04Evaluation

    Read-only until you decide otherwise.

    Start with scheduler metadata. Aurelius replays historical decisions, simulates counterfactual outcomes, and produces an audited savings report before any production rollout.

    01Upload telemetryWeek 1
    02Offline replay
    03Savings estimate
    04Shadow deployment
    05Controlled rollout
    05Get access

    Run a read-only savings audit.

    Upload historical scheduler metadata. Aurelius replays decisions offline, simulates counterfactual outcomes, and produces an audited savings report. Metadata only. No payload access. No production changes.