{"model": "gpt2", "n_tensors_roundtripped": 149, "bytes_roundtripped": 326078464, "max_new_tokens": 32, "seed": 42, "temperature": 0, "token_identity_8_of_8": true, "wallclock_s": {"load": 8.9, "before_gen": 265.6, "roundtrip": 21.9, "after_gen": "~265 s (asserted token-equal)"}, "_note": "End-to-end empirical demonstration of byte-exact-roundtrip -> token-identity. SCAPE bf16_split per-tensor compress+decompress with byte-equal assert on every tensor before re-loading state_dict. All 8 prompts produced identical token sequences across before/after passes. Run on c3-highcpu-88 (openzl-bench-h3 type CPU) on 2026-05-14. Full 11-model sweep at max_new_tokens=256 should run on genom-beast-gpu (T4) per inference_sanity_check.py docstring."}