[initial-d]

The earlier frozen-backbone memory experiment lacked independently trained simpler baselines. Add opt-in late-decoder training with the same exact frozen-prefix computation for full memory, one trained vector per head, and memory-off Qwen; retain all three seeds under a committed common protocol and score previously untested calibration/test windows.

All nine final models complete. Full-memory test PPL is 64.759–65.312, versus 64.358–64.592 for trained constants and 64.288–64.527 for trained memory-off. Every independent paired-window interval supports the cheaper baseline. Same-checkpoint head-mean ablation has a smaller positive effect (0.566–0.759% excess test PPL), but no seed meets the fixed four-control necessity gate. No extra codec selection or LLM capacity win is claimed.

Six actual memory exports preserve source rows, addressing and logits byte for byte. Ordinary memory-off checkpoint reloads and all nine saved tokenizers are checked. Raw per-target metrics, exact commands, runtime SHA, hardware, complete-file inventory and limitations are archived; the new CI audit recomputes independent-control bootstrap intervals and fixed decisions using only the standard library.

Validation: 26 pinned CPU/CUDA tests (one MPS skip); 231 local tests (120 optional-dependency skips); dependency-free evidence audit and diff checks pass. GPU 5/6/7 were released only from hash/argv/cwd-verified occupancy scripts; training/evaluation and unknown processes were preserved, and task workers/helper have exited.