01 · The idea
What stayed with us
Failed-attempt accounting, slice metrics, trace review, candidate selection, batch-first post-training, policy lag, and LoRA update geometry make autoresearch inspectable instead of hiding the search process behind a final score.
Candidate-selection framing can make a sparse task measurable when unconstrained generation is too difficult or too ambiguous to grade reliably.
Batch rollouts, offline scoring, and compact updates can make a post-training loop easier to inspect, while policy lag can stabilize learning under some reward surfaces.
Effective update geometry, layer placement, and controlled rank analysis can explain why two adapters with similar parameter counts behave differently.
File-native and code-native harnesses may fit knowledge-work agents better than bespoke tool APIs because the environment already exposes durable state and familiar operations.
02 · The local translation
What we did with it
PostTrainLLM added a technique registry, SQL candidate-selection tooling, slice metrics, trace review, a batch plan renderer, and LoRA geometry documentation and tooling.
The SQL lane gained a candidate-selection step before open SQL generation so selection quality can be separated from generation quality.
PostTrainLLM retained a batch post-training plan but did not schedule an OAPL-style run before a clean reward surface exists.
PostTrainLLM added LoRA geometry diagnostics and retained a controlled rank-and-layer sweep as a possible future experiment.
Future coding-agent work should prefer file and code surfaces when they are the natural task boundary.
03 · The boundary
Where the comparison stops
The distributed autoresearch operating pattern was not copied. Only methods that remain useful and auditable on one Mac were retained.
The framing is a method, not evidence that the PostTrainLLM SQL specialist is competitive. No fresh candidate model result is implied.
Policy lag is not assumed to help every target. It remains a hypothesis until a bounded local comparison isolates it.
Geometry diagnostics describe an update; they do not establish task improvement without the same frozen behavioral evaluation used for the baseline.
The coding-agent product lane is not active, and this study does not authorize a new agent runtime or evaluation campaign.