benchmark/

Package cacheevidence projects prompt-cache request policy without retaining prompt text, cache keys, or other request content.

Package evidenceproxy implements a prompt-blind reverse proxy used only by the benchmark harness.

Package harness defines the reproducible Agentic Coding benchmark protocol.

Package pierbackend binds the generic benchmark contract to a frozen Pier runtime and DeepSWE-style Harbor tasks.

Package pilot provides a development-only, fail-closed five-task runner.

Package report turns content-addressed agentic benchmark artifacts into a deterministic, self-contained HTML report.