Eval

Same task — which executor wins, and what did it cost?

Eval stays micro. One task, N executors, deterministic criteria. Not a benchmark platform.

Overview

@ruah-dev/eval is a micro-benchmark harness. One task (prompt + seed files), N executors, acceptance criteria as the judge, fresh sandbox per run.

ruah eval init
ruah eval run .ruah/evals/sample-eval.json --out results.json
ruah eval run spec.json --runs 5
ruah eval compare a.json b.json
ruah eval report scorecard.json --format md
ruah eval report scorecard.json --format html

Requires Node.js >= 18. Criteria are verify-shaped (command / file_absent / file_matches / unverifiable) but judged in-process — eval does not import @ruah-dev/verify.

Acceptance

• Sandboxes are always temp dirs. Writing outside fails the run.
• unverifiable never becomes pass.
• Mock executors exist so the harness can test itself without API keys.