# Modification notices

DevelopingDad produced this release on 2026-10-08 from Trial62 revision
`568aa9764b4f695fb18881b97f6d4ee8b725f75c`.

- `model.safetensors`: text MLPs converted from BF16 to packed NVFP4;
  text attention weights converted to FP8; calibrated quantization and FP8
  K/V scales added. BF16 tied target embedding/output head and media
  projections retained. No additional optimization training was performed.
- `config.json`: exported compressed-tensors quantization configuration.
- Root tokenizer, processor, generation configuration and native chat
  template: preserved from the source checkpoint.
- `serving/gemma4-profile-chat-template.jinja`: optional native-template
  overlay prepending the general V10 response contract.
- `serving/overlays/vllm/`: modified vLLM files implementing a separate
  FP8 draft output head and protected target embedding/head sharing.
  Each modified file carries a prominent modification notice and retains
  its upstream copyright/SPDX notices.
- `serving/draft_head_fp8_overlay.py`: draft-head conversion and
  validation helper.
- `serving/receipt_bootstrap.py`: per-boot byte-preserving diagnostic
  receipt archival before executing vLLM.
- Publication launch wrapper, documentation and sanitized summaries:
  added for this release. Private deployment credentials, endpoints,
  infrastructure records and application fixtures are excluded.