capture-evidence

To establish an offline evaluation baseline, generate repeatable harness, metric, splits, and baseline artifacts from your workload. This creates a traceable local foundation to compare current versus future model performance before making changes.

Local-first benchmarking enables private, offline evaluation without hosted services or accounts. You do not need existing traces or baselines; the workflow generates initial baseline artifacts directly from your repository.

Generate auditable artifacts by running a capture-evidence workflow against your repository. It produces timestamped files like harness.json, environment.json, and baseline.json with source refs to enforce provenance for baseline-driven optimization.

Yes, you can run private workload profiling without hosted services. The local-first workflow operates entirely offline, generating environment.json and workload-profile.md files to support private benchmarking.

The best way to compare model performance locally is by establishing a baseline using generated splits.json and baseline.json artifacts. These provide a traceable path to evaluate future changes against your original workload.