Machine calibration

Use this when CI and developer machines differ and you want a rough estimate before running a full local suite. The output is an estimate with diagnostics, not a measured local result and never a release gate.

If both runs use one machine, use an ordinary comparison instead.

Machine profiles

profile = machine_profile(label = "runner-a")

Records CPU, architecture, OS, core counts, SIMD width, installed memory and Julia threads.

• spec_id identifies a class of specifications, not a physical device.

• Host names and transient CPU load do not enter the hash.

• Record quotas, affinity and container limits explicitly in limits; the capture does not verify them.

similar_machines(target, profiles) ranks same-OS/architecture profiles by numerical differences and CPU identity — a shortlist, not evidence of comparable performance.

Calibration battery

estimate_performance(target, references, workload) is an experimental nearest-neighbour transfer model. Each input record has this shape (the numbers are format examples, not measurements):

Dict(
    "machine" => machine_profile(label = "runner-a"),
    "context" => Dict("suite_revision" => "...", "environment" => "...",
        "measurement" => "warm median latency", "unit" => "seconds",
        "resource_policy" => "one pinned core; declared memory limit"),
    "calibration" => Dict("small_compute" => 0.01, "memory" => 0.02, "dispatch" => 0.03),
    "measurements" => Dict("heldout_workload" => 0.5))

• Use a small battery that represents the workloads: computation, memory, dispatch, native calls and I/O need different coverage.

• Calibration names must refer to the same frozen workload definitions, with positive times in the same unit.

• The held-out workload must not appear in target calibration.

• At least three shared calibration workloads and two separately labelled reference machines are required.

• Incompatible contexts, OS/architecture and self-prediction are rejected. Duplicate machine labels are errors.

The result is a central estimate plus an envelope widened by calibration errors. It is not a confidence interval. Only synthetic controls have been tested; no cross-machine external validation has been performed.

CI policy

• Use estimates to prioritize the next real measurements or choose representative runners.

• They are never eligible for a regression gate: ci_gate_eligible = false.

• A single sample, incompatible contexts or an unstable calibration yield insufficient_evidence. PerfChecker never fabricates an estimate.

The CLI exposes machine --label=runner-a and estimate --source=transfer.json (with target, references and workload fields).