Imported from yzj623/pplab (
ResearchAgent-Skill/skill_library/paperbench/bridging-data-gaps/skill/latent_diffusion_models/reduced_recovery_harness/SKILL.md). Install upstream withnpx skills add yzj623/pplab --skill reduced_recovery_harness. Copyright stays with the author.
Reduced Recovery Harness
Use this skill when full Latent Diffusion Model checkpoints or datasets are unavailable in a bounded run and soft-mode recovery permits a declared proxy. It coordinates generated skill scripts, writes recovery artifacts, and prepares evidence for the Distiller recovery validator. Do not use it in hard mode as accepted success.
Inputs
- Attempt directory containing
module_plan.json. - Runtime handoff path from environment preparation.
- Generated skills root containing compression, objective, conditioning, and spatial scripts.
- Optional resource provenance for a synthetic or resource-derived image item.
Outputs
recovery/recovery_result.json.recovery/logs/training_trace.json.recovery/logs/generated_data_item.json.recovery/logs/generated_skill_invocations.json.recovery/source_manifest.jsonand command evidence produced by the caller.
Workflow
- Read the module-plan target and runtime blockers.
- Declare whether the experiment is reduced/proxy and why full recovery is blocked.
- Invoke the generated module scripts instead of duplicating their logic silently.
- Combine their outputs into mechanism checks for compression, latent noising, cross-attention, spatial planning, and optimizer execution.
- Write validation-compatible artifacts and keep the original source repository out of recovery sources.
Validation
Run python scripts/build_proxy_result.py --attempt-dir <attempt> --skills-root <generated_skills_root> --runtime-handoff <attempt>/environment/runtime_handoff.json. Then run the Distiller recovery validator on the attempt directory.
Limitations
The harness produces a reduced mechanism proxy, not a published FID reproduction. It depends on the generated module scripts and must be rerun after skill refinements.