---

name: openclip_recovery_harness description: Run an auditable reduced OpenCLIP scaling-law recovery by invoking generated skills and producing validator-compatible recovery artifacts.

OpenCLIP Reduced Recovery Harness

Use this skill when full LAION-scale OpenCLIP training is infeasible but soft-mode recovery permits a declared proxy that must exercise the paper mechanism.

Inputs

• Attempt directory with module_plan.json and environment/runtime_handoff.json.
• Generated skill root containing the scale protocol, contrastive objective, zero-shot/retrieval evaluation, and power-law scaling skills.

Outputs

• Recovery result JSON with target metadata copied from module_plan.json.
• Mechanism checks for each generated core module.
• Invocation and command evidence.

Workflow

1. Create deterministic synthetic paired embeddings and scale points.
2. Invoke generated skill scripts rather than duplicating their logic silently.
3. Check contrastive loss, zero-shot accuracy, retrieval Recall@1, and power-law R2.
4. Mark the run as proxy and reduced; do not claim full model training.
5. Save source and invocation evidence for the Distiller recovery validator.

Validation

Run python tests/test_recovery_harness.py.

Limitations

This harness validates mechanism faithfulness only. It cannot substitute for full LAION-2B pretraining metrics.

Cycle 4 API stress check

Run python scripts/stress_checks.py --check api_imports --attempt-dir <attempt_dir> --skills-root <generated_skills_root> --output <json> to verify all generated helper APIs import cleanly.