L’interface est localisée ; les présentations et documents des Skills sont traduits automatiquement par IA dans votre langue. Le code, les commandes, chemins et termes techniques restent en version originale, et le document de l’auteur fait foi.
Explorer les Skills
Données et recherche · SKILL.md

offline_recovery_harness

offline_recovery_harness

Execute a bounded D4RL-style offline recovery proxy with fixed-dataset training and normalized-score evidence.

Independent GitHub snapshot · Automated classification · Original documentation governs

Documentation de la Skill

GitHub

name: offline_recovery_harness description: Execute a bounded D4RL-style offline recovery proxy with fixed-dataset training and normalized-score evidence.

Offline Recovery Harness

Use this skill when full D4RL packages or datasets are unavailable but soft-mode recovery permits a declared mechanism-faithful proxy. The harness should still perform executable work: validate a fixed transition dataset, run a parameter update from that dataset, compute normalized scores, and save mechanism evidence.

Inputs

  • module_plan.json containing the fast recovery target.
  • environment/runtime_handoff.json describing package blockers and allowed sources.
  • Generated skill root containing dataset validation, score normalization, and taxonomy skills.
  • An output recovery directory.

Outputs

  • recovery_result.json with numeric metrics and target metadata.
  • logs/generated_data_item.json with transition content and provenance.
  • logs/training_trace.json with loss_before, loss_after, params_before, and params_after.
  • Invocation logs proving generated skills were used.

Workflow

  1. Read the module plan target and runtime handoff.
  2. Build a tiny fixed offline chain dataset with explicit synthetic-proxy provenance.
  3. Validate the dataset through offline_dataset_protocol.
  4. Classify proxy metadata through benchmark_property_taxonomy.
  5. Fit a one-parameter linear Q approximation with a deterministic gradient update.
  6. Compute raw and normalized return through d4rl_score_normalization.
  7. Save mechanism checks: no online collection, fixed dataset consumed, reduced training executed, optimizer step changed parameters, full D4RL packages unavailable, and normalized scoring used.
  8. Keep full-runtime flags false when the runtime handoff reports missing packages.

Validation

Run the harness script used by the recovery attempt, then run Distiller's validate_recovery_experiment.py on the attempt directory.

Limitations

This harness is a reduced proxy, not a full D4RL benchmark run. It should be accepted only in soft mode after full recovery is blocked and the experiment gate passes.

Skills enregistrées