| # GLM-4.7 Aider post-training data | |
| Canonical private, manual-approval store for every project-tracked Aider training corpus and dataset lineage. `CATALOG.json` marks trained, candidate, rejected, audit-only, and failed artifacts without conflating them. | |
| Hidden benchmark tests, runtime rubrics, reference answers, and executable grader-oracle material are intentionally outside this repository. Internal review receipts are labeled per entry. Fixed-26 model responses live in `TokenBender/glm47-aider-fixed26-responses`. | |