Skip to content

SELF_LEARNING_DOE.md

Latest commit

 

History

History
69 lines (52 loc) · 2.72 KB

SELF_LEARNING_DOE.md

File metadata and controls

69 lines (52 loc) · 2.72 KB
Raw

Self-Learning DoE Runbook

Self-learning in the public scaffold means disciplined memory, not autonomous physical-execution control. The planner records what happened, what was excluded, what should not be repeated, and what needs review before the next experiment round.

Artifacts

ferm-doe plan-wave2 writes three learning artifacts:

  • learning_ledger.csv: structured events an agent can carry forward.
  • hiccup_review.md: human-readable review notes.
  • negative_result_memory.json: arm-scoped regions that should not be repeated by default.

Learning Ledger Fields

Field Purpose
learning_id Stable row id.
scope campaign, arm, results, responses, or wave2.
event_type monitor, hiccup, assay_power, decision, or a user-defined type.
trigger Why the event exists.
observation What happened.
action_taken What the agent or scientist did next.
status open, monitoring, closed, or user-defined.
owner_role Human or agent role responsible for closeout.
follow_up_ref Artifact or issue reference.

Hiccup Policy

Record a hiccup when:

  • First-batch rows fail QC or are excluded.
  • Trust scores are low.
  • assay-power policy is missing or weak.
  • a requested scale or downscale move lacks bridge eligibility.
  • a recommendation is pause or stop.
  • the agent changes factor space because the best row is on a boundary.

Do not convert a hiccup into a hard failure automatically. The public lane should remain versatile: a scientist may decide to confirm, narrow, expand, pause, stop, or branch depending on goals and practical constraints.

Negative Memory

Negative memory is arm-scoped by default. A bad plate condition should not automatically block reactor planning unless a bridge policy explicitly says the signal transfers.

Recommended policy:

{
  "adaptive_wave2": {
    "self_learning": {
      "enabled": true,
      "negative_memory_scope": "arm",
      "learning_ledger_path": "wave2/learning_ledger.csv",
      "hiccup_review_path": "wave2/hiccup_review.md",
      "closeout_rule": "Every open hiccup must either be closed with evidence or carried into the later campaign manifest."
    }
  }
}

Agent Loop

  1. Run ferm-doe validate.
  2. Run ferm-doe plan-wave2 after first-batch results exist.
  3. Read result_ingestion_report.json and assay_power_results.json.
  4. Review wave2_recommendation.md.
  5. Update or accept wave2_manifest.patch.json.
  6. Carry learning_ledger.csv and negative_result_memory.json into the next planning wave.

The learning setup is designed for handoff. A later agent should be able to tell which decisions were evidence-backed, which were caveated, and which need human review.