> Canonical page: https://posttrainllm.com/changelog

Product changelog

# What changed

Verified outcomes from the Mac-local specialist factory, newest first. Future work belongs in GitHub Issues; this page records shipped product capability and measured results.

 2026-07-25

## Factory runs became durable and inspectable

- Factory runs now keep authoritative lifecycle state, safe transitions, recoverable current/latest pointers, and explicit failure history without pretending that an interrupted run completed.
- Fine-tune report cards turn recorded before/after evidence into portable public pages while labeling missing measurements and blockers instead of filling them with zeros.

 2026-07-13

## A second specialist package with broader coverage

- The public qwen3-4b-rest-fused package retained the 100% file-operations gate and improved the recorded breadth score from the stock model’s 59.6% to 65%.
- The package is presented as a narrow research specialist, with historical performance and trace gaps disclosed rather than reconstructed.

 2026-07-04

## The full factory loop produced its first honest decision

- A frozen SQL-hygiene candidate completed training, evaluation, packaging, and reporting end to end.
- The candidate regressed, so the factory recorded a retry-training decision instead of promoting a weaker model.

 2026-05-31

## Mac-local training and serving moved into a native runtime

- The Swift and MLX command-line runtime added local training, serving, and agent surfaces that work with the project’s existing evaluation and packaging assets.