A retry decision was enough to close the first factory loop

The first complete PostTrainLLM factory run did not ship a model. It produced a retry decision.

That still mattered. The run moved through one reproducible sequence: target, data, post-training, evaluation, packaging, and report. Its folder kept the config, dataset identity, frozen baseline, candidate evaluation, decision, and human-readable report together.

The candidate was the frozen qwen06 SQL-hygiene target. The run completed on July 4, 2026, and passed the repository's factory-run validation. The result exposed a training failure instead of smoothing it into a launch story.

The important change was procedural. A completed training command was no longer the finish line. Every experiment had to end with evidence strong enough to support ship, reject, or retry. If the evidence supported retry, that was the recorded outcome.

That first loop made later routed specialists and report cards possible because the project had a durable place for negative results.

The source devlog is at https://posttrainllm.com/devlog.