Migrated from open-experiments/Telco-AIX#19 (original ask: RLHF for network decision making in AutoNet), reformulated per the blueprint: the model stays a stateless service, so learning happens offline in the training pipeline, not in the agent loop.
Goal: Close the human-feedback loop — every approve / override / reject at 301's approval gate is a labeled preference judgment about a proposed remediation. Use them.
Blueprint-aligned design:
- Log gate decisions as structured preference records (plan features, scorer score, judge verdict, human decision, operator note) — traceable per convention 5, stored outside the loop per convention 4.
- Offline training job (DS Pipelines, as in 202's train→register→promote) periodically retrains/calibrates the risk scorer on accumulated preferences; MLflow-registered, promoted through the existing KServe serving path.
- Evaluate 302-style: measure scorer-vs-human agreement before/after, and re-measure the LLM-as-a-judge against the updated scorer.
- Optional stretch: DPO-style preference tuning of a small planner model, still served via vLLM as a stateless endpoint.
Non-goals: online/in-loop weight updates; per-session personalization; anything that makes the agent pod stateful.
Related: #18 (its approval-gate decision log is the data source for this issue).
Migrated from open-experiments/Telco-AIX#19 (original ask: RLHF for network decision making in AutoNet), reformulated per the blueprint: the model stays a stateless service, so learning happens offline in the training pipeline, not in the agent loop.
Goal: Close the human-feedback loop — every approve / override / reject at 301's approval gate is a labeled preference judgment about a proposed remediation. Use them.
Blueprint-aligned design:
Non-goals: online/in-loop weight updates; per-session personalization; anything that makes the agent pod stateful.
Related: #18 (its approval-gate decision log is the data source for this issue).