benchmark

name	benchmark
description	Defines how an @samchon/lint-plugin-evidence benchmark campaign is set up, launched under frozen inputs, run, supervised, recovered, and reported, from issue creation through pull-request completion. Use whenever operating, supervising, or reporting a benchmark run.

Benchmark Operation

A campaign measures one coding engine building the same application twice: once with @samchon/lint-plugin-evidence and its guidance, once with neither. One subject and arm is a cell, one execution of a cell is a run, retained under benchmark/output/<subject>/codex/<arm>/runs/<run-id>/.

Two perspectives, and they are never interchangeable.

• Measurement — the benchmark's own view: how a cell is run, whether what it did still counts, and what may be published about it.
• Intervention — your view as this repository's agent: what you may change, how you warn a cell, and how you recover one.

A rule written for the operator never binds a cell, and a cell's permitted edit is never yours to make.

Standing Rules

• Change nothing but the arm.
• Report only what the record retained.
• Never repair a measured workspace.
• Warn before resume, resume before derive, derive before restart.

Commands

Under pnpm --filter @samchon/evidence-benchmark:

• start — launch, resume, or derive a run. running.md, recovery.md
• dashboard, audit-suspensions — the live pull-request record. dashboard.md
• report — the tracked benchmark/aggregate artifacts. aggregate.md
• supervise — apply a hand-written Plain review verdict. plain-review.md
• warn — deliver an operator warning to a running cell. warning.md