# REMAINING_WORK.md

**No remaining work to ship paper draft v4.0.6.** The empirical results, the regenerated `bf16.zl` (617 B) profile, the smoke-test fixtures and checksums, the eight inference-sanity prompts, the five figures, and the manifest are all vendored and verified end-to-end. The archival snapshot is GitHub release `v4.0.3` and its Software Heritage SWHID `swh:1:rel:55d910f5af170c22719cc9346f4d8a5029f09164`; v4.0.6 is the current head and is byte-identical to v4.0.3 except for archival-identifier housekeeping. No Zenodo DOI was minted; see `DEVIATIONS.md §H`.

For the (small) catalog of current drift between paper text and live artifacts, see `DEVIATIONS.md`.

## What is *deliberately* not in this version

- External-method strict-accounting rows (DFloat11, Cloudflare Unweight, ZipNN). Reasons: DFloat11 + Unweight kernels are Hopper-only (H100/H200), incompatible with the CPU-only reference; ZipNN's official CPU release fails byte-exact roundtrip in default config. Paper §11.2 scopes this out explicitly.
- fp16 OpenZL trained profile. The fp16 row in Section 6.2 / Section 10.1 is deferred until the 778 B fp16 profile is regenerated (paper §11.1).
- 30B / 70B entropy spot-check. Not run; 7B Qwen2.5-7B-Instruct is the largest validation point (paper §5.8).
- Full 11-model inference sanity sweep. Only gpt2 verified end-to-end (149 bf16 tensors, 311 MiB, 8/8 token identity); the byte-exact roundtrip assertion (§4.2) mathematically implies token identity for the other 10 test models at temperature 0, but a CUDA / T4 sweep is the obvious belt-and-suspenders extension.
- Hosted Docker image. The Dockerfile ships in-repo; no ghcr.io publication (no `write:packages` on the build account).
- Zenodo DOI. Dropped (Zenodo↔GitHub webhook installation did not complete on this repository); see `DEVIATIONS.md §H`.