Machine definitions 0.1

Status: complete working draft for prototyping and collaborator review. Version 1 is reserved for the first reviewed public release.

This document defines the human meaning of schemas/machine/0.1.schema.json. A machine is a reusable description of the execution environment for timing or another machine-dependent score. A result may omit a machine only when it makes no machine-dependent claim.

Machine class

class is a coarse hardware partition used for filtering and honest leaderboard grouping:

• cpu: execution is attributed to general-purpose CPU resources;
• gpu: execution is attributed to GPU resources;
• fpga: execution is attributed to an FPGA implementation;
• asic: execution is attributed to an application-specific integrated

circuit; and

• hybrid: the claimed execution materially combines more than one of those

classes.

The class is not a normalized performance unit. Circuit Bench does not convert measurements from different classes into a universal hardware-independent speed ranking.

Execution-environment description

description states the exact device and resources used to the extent needed to interpret and reproduce a machine-dependent measurement. It should include, where applicable, processor or accelerator model, device count, relevant allocation, and material host constraints. Software builds and runtime details specific to a run remain on the result.

The description names the resources actually used, rather than the maximum resources physically installed in a host.

Evidence status

physical means the reported measurement was obtained from the described physical execution environment. simulated means it came from a simulation of that environment. estimated means the value is an analytical or extrapolated estimate rather than an execution measurement. These categories must remain visible when machine-dependent results are compared.

Correction and infrastructure

A published machine record is immutable. A materially corrected description creates another machine record which may identify its predecessor using supersedes_machine_id. Existing results retain their original machine reference.

schema, id, slug, lifecycle state, submitter, and timestamps are identity and infrastructure fields. They carry no scientific performance value by themselves.