Run from a checkout of the v4.0.3 repo (or the Docker image):
./reproducibility_smoke_test.sh
Three sha256[:16] checks against the three fixtures in fixtures/, all roundtrip-verified. Exits 0 on full pass. Measured values:
| Method | Fixture | Compressed (B) | Ratio | sha256[:16] |
|---|---|---|---|---|
bf16_split | TinyLlama_layer3_q_proj.bin (8 MiB) | 5,660,476 | 1.4820× | fc3544c35489eb94 |
qb_mixture_k4 + 277 B profile | Llama32_3B_layer14_gate.q4k.bin (13.5 MiB) | 13,472,852 | 1.0507× | c79972a64d11b717 |
decomp_perstream_zstd19_bg | Qwen2.5-0.5B-Q4_K_M.gguf (379 MiB) | 383,733,236 | 1.0367× | 1c6f077f8c228cf2 |
The 379 MiB GGUF fixture is auto-fetched from this mirror if absent locally (GitHub's 100 MB per-file limit prevents shipping it in-repo).