This is the merged, directly loadable checkpoint from the 310-trajectory regret-weighted CoMem manager experiment.

The model starts from YWZBrandon/summary-sft-qwen3-4b. Summary generation is trained with GRPO using DeepSWE next-action preservation and bounded context length advantages. The KEEP/SUM selector is trained with regret-weighted soft targets and action-balanced replay. Above the 8,192-token budget, training uses eight summaries; at or below budget it uses one virtual KEEP reference and seven summaries.

The current one-token, unrestricted text-generation gate evaluation can emit a third token instead of the KEEP or SUM verbalizers. For a gate measurement aligned with training, compare the next-token logits of KEEP and SUM or use constrained decoding. The summary-generation checkpoint itself is unaffected by this evaluation-interface issue.

This checkpoint is experimental and has not yet been validated by a complete SWE-bench Verified harness run.