experiment-harness

To validate spell outputs, the experiment harness executes bounded native skill runs and legacy runtime adapters, capturing real runtime outputs and validation evidence. It then generates timestamped reports with optional observability telemetry for analysis.

A profile-aware workflow coordinates prompts, native skill runs, and observed outputs across specific regimes. It produces observable artifacts like validation reports and telemetry to verify runtime behavior.

Yes, jq and codex are required dependencies to operate the experiment harness. They provide the host runtime adapters and processing logic required to execute bounded runs and parse validation artifacts.

Yes, a contract at SKILL.md or README.md is required to initialize the development harness. The harness uses this contract to identify the artifact and structure the initial prompts, regimes, and output directories.

Use the harness's built-in loop-based improvement workflow to iterate bounded native skill runs against your fixtures. This process validates outputs repeatedly to ensure consistent runtime behavior and generates timestamped reports.

Avoid using an experiment harness without a defined contract at SKILL.md or access to host runtime adapters. Without structured prompts, regimes, and fixtures, the harness cannot execute bounded runs or generate validation evidence.