인터페이스는 현지화되어 있으며, Skill 소개와 설명 문서는 선택한 언어로 AI가 자동 번역합니다. 코드·명령어·경로·전문 용어는 원문을 유지하며 원작자 문서가 우선합니다.
Skills 둘러보기
생산성 및 협업 · SKILL.md

msd_recovery_protocol

msd_recovery_protocol

Assemble executable soft-mode recovery evidence for Minimum Stein Discrepancy Estimator proxy experiments.

Independent GitHub snapshot · Automated classification · Original documentation governs

Skill 문서

GitHub

name: msd_recovery_protocol description: Assemble executable soft-mode recovery evidence for Minimum Stein Discrepancy Estimator proxy experiments.

MSD Recovery Protocol

Use this skill when recovering Minimum Stein Discrepancy Estimators under bounded runtime. It coordinates the generated Stein-kernel, diffusion-design, and optimiser skills to produce validation-ready recovery artifacts without reading an original source repository.

Inputs

  • Attempt directory containing module_plan.json, module docs, generated skills, and environment/runtime_handoff.json.
  • Recovery mode and target from the run manifest.
  • Paths to generated skills that must be invoked or cross-checked.

Outputs

  • recovery/experiment_plan.md describing full and proxy options.
  • recovery/source_manifest.json listing only allowed sources.
  • recovery/logs/generated_skill_invocations.json with evidence for every core module.
  • recovery/logs/generated_data_item.json and recovery/logs/training_trace.json for reduced proxy evidence.
  • recovery/recovery_result.json and recovery/experiment_validation.json.

Workflow

  1. Confirm full reproduction blockers and soft-mode permission for a declared proxy.
  2. Generate a deterministic Student-t location sample matching the module-plan target.
  3. Call or import the generated diffusion-design skill to obtain Student-t diffusion diagnostics.
  4. Call or import the Stein-kernel skill to evaluate DKSD losses.
  5. Call or import the optimiser skill to minimise the empirical loss.
  6. Save all logs and mechanism checks before running the recovery validator.

Validation

Run the bundled tests/test_msd_recovery_protocol.py for source-manifest and mechanism-check helpers. The full recovery harness in the attempt directory should additionally run validate_recovery_experiment.py.

Limitations

This skill accepts a declared reduced/proxy target only in soft mode. It must never label proxy evidence as full reproduction of the paper figures.

Refinement cycle 3 note

A negative source-boundary check rejected an original_repo source; keep this check as a required recovery-protocol regression.

저장한 Skills