Skip to content

Graph-Native Optimization State (CONCEPT:ORCH-1.31)

Overview

Persists the GEPA Pareto frontier (candidates + ancestry) into the durable epistemic-graph so optimization is resumable and accumulates across sessions — a synergy unique to our graph-native architecture (neither predict-rlm nor the GEPA reference impl does this). Extends ORCH-1.13 (+KG-2.7 ingestion/persistence).

How it works

  • Snapshot round-trip (ParetoCandidatePool.to_snapshot / load_snapshot) — pure serialization of the frontier to/from plain dicts, preserving candidate prompts, scores, and parent_ids ancestry.
  • Persist / resume (GEPAOptimizer.persist_frontier(run_id) / resume_frontier(run_id)) — writes a GEPAFrontier graph node (extending the existing OptimizationTrajectoryNode/EvaluatorFeedbackNode writes via create_or_merge_node) and reads it back to seed the pool. A killed run resumes from the persisted frontier with no loss of the best candidate; prior frontiers are reusable optimization state.
  • Best-effort — absent a backend, persist/resume degrade quietly (return falsy), never raising.

Key files / API

Piece Location
Snapshot + persistence rlm/gepa.py (ParetoCandidatePool.to_snapshot/load_snapshot, GEPAOptimizer.persist_frontier/resume_frontier)

Wiring (≤3 hops)

graph_orchestrate(action="rlm_optimize") → optimizer → persist_frontier/resume_frontier → graph (≤3 hops).

Research provenance

Synthesis of GEPA (frontier/ancestry) with agent-utilities' durable epistemic-graph persistence — verified against rlm/gepa.py (OptimizationTrajectoryNode/EvaluatorFeedbackNode).