The operational 4Fs

Four system qualities, not four slogans

The operational Four Fs describe what an adaptive system should optimize across deployment environments. They are most useful when each term is converted into measurable acceptance criteria.

Fast

Fast means low end-to-end latency under realistic load, not merely a high token-generation rate. Measure routing, model loading, retrieval, network transfer, aggregation, and tail latency.

Useful metrics include p50, p95, and p99 latency; cold-start time; queue time; throughput at a fixed quality threshold; and time to recover after a model swap.

Flexible

Flexible means the system can replace or recombine components without rewriting the entire application. Flexibility depends on contracts, manifests, version negotiation, compatibility tests, and rollback—not on how many plugins exist.

Measure the time required to add a specialist, change a router, migrate a runtime, or roll back a candidate. Track the percentage of changes that remain local to one component.

Frugal

Frugal means the system uses the minimum resources required for the needed outcome. Include memory, accelerator residency, energy, network transfer, storage, operator time, and evaluation cost. A small model that triggers three retries and a cloud fallback may be less frugal than one medium model.

Measure joules or cost per accepted result, peak memory, bytes transferred, model load churn, utilization, and maintenance hours per capability.

Federated

Federated means computation or learning is distributed across devices, sites, or organizations while respecting data boundaries. It does not guarantee privacy by itself. Federated systems still require secure aggregation, authentication, poisoning defenses, update provenance, and explicit trust assumptions.

Measure communication volume, update staleness, client participation, convergence under non-identical data, privacy budget where applicable, and resilience to compromised clients.

A balanced scorecard

Design tensions

The Four Fs conflict. More specialists may increase flexibility but hurt cold-start latency. Federated learning may protect data locality but increase communication and operational complexity. Aggressive quantization may improve frugality but damage calibration. The correct outcome is a Pareto frontier, not a single maximum.

FUNCTION evaluate_four_fs(system, workload)
    measurements <- RUN_WORKLOAD(system, workload)

    RETURN {
        fast: SCORE_LATENCY_AND_THROUGHPUT(measurements),
        flexible: SCORE_CHANGE_ISOLATION(system),
        frugal: SCORE_TOTAL_RESOURCE_COST(measurements),
        federated: SCORE_DATA_LOCALITY_AND_TRUST(system)
    }
END FUNCTION