I’m designing an offline loop for a local coding model: collect agent traces, keep only the ones that pass tests, require human or reviewer approval before training, train a periodic small adapter, then use a fixed eval suite as a gate, with automatic rollback…