Safety

    Aurelius begins as an analysis tool, not a control system.

    Every evaluation begins as a historical replay against recorded scheduler traces. Aurelius evaluates scheduling decisions entirely offline before any shadow mode or optional production integration.

    Unlike systems that immediately participate in production scheduling, Aurelius is designed to begin entirely offline, on your own recorded workloads, before any interaction with a live control plane.

    Initial evaluation

    Metadata flows in. Nothing flows back.

    During an evaluation, Aurelius reads recorded scheduler metadata, replays it offline, and produces a savings report. It does not write to your scheduler, move workloads, or read the contents of your jobs.

    Your infrastructure
    Scheduler
    queue state · capacity · job history
    read-only metadata
    Aurelius
    Offline evaluation
    Offline replayBaseline comparisonSavings report
    No production changesNo workload movementNo payload access
    What Aurelius can access

    Read-only means scheduler metadata, not the work itself.

    Aurelius reads scheduler metadata such as queue state, capacity, and job history. During evaluation it does not access prompts, model outputs, training data, or customer payloads, and that telemetry is not used to train a foundation model.

    Reads · scheduler metadata
    • Queue state and depth
    • GPU and resource availability
    • Historical job events
    • Timestamps and durations
    • Placement history
    • Resource utilization
    Does not access during evaluation
    • Prompts
    • Model outputs
    • Training datasets
    • Customer payloads
    • Secrets
    • Credentials
    How evaluation works

    Constraints are checked before a recommendation exists.

    A recommendation is the output of a sequence that runs mostly offline. Recommendations that violate configured hard constraints, such as latency, SLA, placement, or capacity limits, are filtered before presentation.

    1. Historical replay

      Your own recorded traces are replayed offline. No live system is involved.

    2. Constraint validation

      Candidates that violate your configured hard constraints, such as SLA, capacity, or placement limits, are filtered before they become a recommendation.

    3. Baseline comparison

      Each remaining candidate is measured against your current scheduler, so a recommendation is grounded in your own recorded results rather than a generic claim.

    4. Shadow recommendation

      Aurelius observes live metadata and emits recommendations only. It does not act on the live scheduler.

    5. Operator review

      A person evaluates and approves before anything changes.

    6. Optional production integration

      Enabled only if you choose to. It is never the default, and it remains under your control.

    Operator control

    By default, Aurelius evaluates historical scheduler behavior and produces recommendations. It does not modify production infrastructure unless an operator explicitly enables production integration.

    Historical replay precedes shadow recommendations. Shadow recommendations precede any optional production integration. Operators remain in control throughout every stage.

    ← Back to home