benchmarks/README.md

# Acceptance benchmark instructions
The 28 prompt cases are **specifications**, not recorded passes. After the plugin is installed, run each against the chosen host, capture actual outputs, relevant tool traces and approvals, then score against `expected_signals` with human review for quality and deterministic tests for schemas/permissions. Preserve version, inputs and exact results. Do not claim benchmark completion from the existence of this file.

Negative cases cover unnecessary multi-agent design, unsafe permissions, fabricated results, prompt injection, tenant leakage, stale approvals, unsafe retries and fabricated parallelism. A critical authorization failure blocks release independently of aggregate quality.