How it works

    An advisory control layer, not a scheduler replacement

    Technical documentation for infrastructure engineers. Aurelius evaluates pending decisions before they reach execution (observing metadata, forecasting conditions, and filtering risk) without touching the execution path.

    Execution model

    Your scheduler stays in control

    Aurelius does not replace your scheduler, intercept payloads, or modify execution. It operates as an advisory layer: it observes job metadata, forecasts energy conditions, generates options, and filters risky decisions. Execution remains unchanged.

    FIG.01scheduler interception
    WORKLOAD QUEUEjob metadata · timingEXISTING SCHEDULERavailability · fairnessEXECUTION LAYERexecution unchangedSCHEDULER REMAINS AUTHORITYmetadata onlyadvisory decisionno payload accessAURELIUS CONTROL LAYERevaluating candidates…forecast−38m windowselecteddelay batchrejecteddelay 2hAPPEND-ONLY LOG
    scheduler interception · control plane
    Forecasting layer

    Short-horizon forecasts with explicit uncertainty

    Aurelius predicts energy cost and carbon intensity using lag-based signals derived from historical patterns and real-time grid data. Every forecast carries uncertainty bounds, and when uncertainty exceeds threshold, a deterministic conservative fallback applies. No black-box models.

    FIG.02control loop
    01OBSERVE02FORECAST03DECIDE04FILTER05LOG
    Rank run / delay / relocate by expected economics.
    deterministic control loop
    Optimization vs safety

    When the two conflict, safety wins

    Optimization
    • Finds lower-cost execution windows
    • Identifies lower-carbon periods
    • Ranks options by expected savings
    • Respects resource constraints
    Safety
    • Quantile-based safety gates
    • Latency-safe constraints
    • Deterministic fallback
    • Conservative defaults
    safety always wins
    Latency-safe mode

    Workloads behave exactly as they would without Aurelius

    Latency-safe mode enforces a zero-slack requirement: jobs complete within their original bounds, start times are preserved, and there is no power throttling. If optimization cannot be achieved within these constraints, an invisible fallback applies. The only difference is lower cost and carbon, when safe.

    Most savings come from time-shifting batch and training workloads away from peak pricing windows, not from throttling or resource reduction.