Objective
Start the Probe model-training lane only after GEPA produces clean traces and route scorecards.
GEPA is distributed rollout optimization over text artifacts. This issue is the separate later neural/model lane for LoRA, Qwen, and MLX/Apple-silicon adapter work.
Scope
- Select training data from GEPA traces, route scorecards, and failure-family deltas.
- Separate prompt/Blueprint optimization evidence from model fine-tuning evidence.
- Define adapter evaluation against retained and validation Probe benchmark tasks before any wider claim.
- Track Qwen and MLX/Apple FM paths as backend experiments, not benchmark authority.
Acceptance criteria
- Training-data selection uses public-safe GEPA trace refs and failure-family deltas.
- LoRA/Qwen/MLX runs are labeled as model-training experiments, not GEPA rollout optimization.
- Adapter evaluation writes Benchmark Cloud-compatible evidence records.
- Public summaries avoid model-training claims until external gates exist.
Related work
Objective
Start the Probe model-training lane only after GEPA produces clean traces and route scorecards.
GEPA is distributed rollout optimization over text artifacts. This issue is the separate later neural/model lane for LoRA, Qwen, and MLX/Apple-silicon adapter work.
Scope
Acceptance criteria
Related work