Machine profiles

Capture real machine identity, then explicitly exercise the refusal path without inventing calibration times:

Start Julia in a project containing PerfChecker (] add PerfChecker if needed).

using PerfChecker
profile = machine_profile(label="local-recipe")
context = Dict("suite_revision"=>"refusal-fixture-v1",
    "environment"=>"local-recipe", "measurement"=>"warm median latency",
    "unit"=>"seconds", "resource_policy"=>"no calibration executed")
target = Dict("machine"=>profile, "context"=>context,
    "calibration"=>Dict{String,Float64}(), "measurements"=>Dict{String,Float64}())
result = estimate_performance(target, Dict{String,Any}[], "heldout")
println((result["status"], result["estimate"], result["ci_gate_eligible"]))

Observed output: ("insufficient_evidence", nothing, false). The receipt records a real perfchecker-machine-profile/1 profile, 64-character specification hash and limits_verified=false; external transfer validation is not_performed. There is no multi-host calibration dataset here. A useful transfer experiment needs overlapping real calibration workloads and independent held-out measurements on distinct machines; see machine calibration.

Calibration limits

Calibration limits

A calibration record maps workload IDs to positive elapsed times in seconds. Shared workloads are the same IDs measured on both the target machine and a reference machine. At least three are required for each reference. Use at least two distinct reference-machine labels; none may be the target label. Freeze workload definitions and units, and exclude the workload being predicted from the target's calibration record.

The algorithm scales each reference machine's measured result by its typical target/reference calibration ratio. It rejects unstable transfers using leave-one-workload-out error. The returned envelope spans the selected references' estimates, widened by that calibration error; it is not a confidence interval or a measurement.

Mismatched contexts, incompatible OS/architecture and self-prediction are refused; duplicate labels are errors. spec_id identifies a specification class, not a physical device. Capture does not verify quotas or affinity, and similar_machines produces only a shortlist. No cross-machine external validation is published; estimates retain ci_gate_eligible=false.