genesis-evals

Automating maintainer-side evaluation requires running the self-eval suite to validate PR changes, applying deterministic scoring and convergence checks across parallel sub-agents to guarantee stable, repeatable results before shipping.

Cold-spawn sub-agent orchestration runs parallel scenarios independently, generating pre-spawn records and applying explicit model prompts to ensure isolated, deterministic evaluation without user-facing context leakage.

Validating PR updates involves running parallel sub-agents that apply deterministic scoring, reporting convergence within a maximum of three iteration loops to ensure stable results and produce an auditable run summary.

The loaded-skills context remains explicitly outside the user-facing bundle to prevent context leakage, ensuring that evaluation prompts and pre-spawn records stay isolated during the scoring process.

Yes, the orchestration scripts depend on pyyaml and jsonschema to parse configuration and validate schemas during the automated evaluation and deterministic scoring workflow.

Limiting convergence to three iteration loops ensures repeatable, stable scoring results while bounding the computational overhead of parallel sub-agent execution during the evaluation cycle.