A result is only as broad
as its test.

Luxi logits and token-sequence hashes repeat across the tested conditions. Top-token and ordered top-5 reference agreement does not establish full-sequence identity with Hugging Face; published greedy sequences include divergences. Normal-mode generation has different speed and energy tradeoffs from deterministic-mode generation.

No specific quantization quality, compression, speed, or energy result is asserted here. The published inference comparisons are not a substitute for a quantization evidence pack. A proposed evaluation must define the representation and computation before making a claim.

The published Luxi exact-path hash matches on the tested L4, x86, and Apple arm64 paths. This fixed-order correctness path is distinct from the faster chunked reduction. The synthetic positive input and single GPU model do not establish universal accuracy or platform support.

The public stock-drift report records model scope, worker conditions, environments, hashing, and reproduction commands. It makes no speed or energy claim. The ordinary PiWind values were not exact-reference checked; the stress test does not validate every output or a customer book.

Customer pricing, market data, scenarios, and methodology remain controlled by the customer. Authoritative rebuilds remain comparators and reconciliation mechanisms.

The reported state bytes match across macOS arm64, Linux arm64, qemu Linux x86_64, and H100 on those grids. The H100 build is near-earth only, and the x86 result is emulated, not a native-host test. Rejected starlink states are recorded as NaNs.

With identical float64 LAFT inputs, Luxi produced bit-identical output hashes across Intel Xeon, Apple M1 Pro, and NVIDIA A100 at nine tested sizes from 50 to 100,000 columns. In the same-pod A100 comparison, Luxi was 1.9× to 11.5× faster than the published Gemini JAX translation across the reported GPU timing sizes, including the GPU-only million-column case.

Additional evidence / scientific breadth

Different studies. Named boundaries.

These studies are published with their workload, hardware and limits.

The published result is fingerprint stability across the tested CPU thread counts and GPU block sizes. It concerns aggregate PairHMM scores on the recorded workload, not a universal genomics result.

The grid starts at 12:00 UTC on 11 October 2026; the summary does not specify a separate measurement date. The CPU environments include Linux x86_64, Linux arm64 under qemu, native macOS arm64, and a Mac Intel binary under Rosetta 2. Each engine repeated on its own platform; the reported SAAL cross-platform differences were micrometres, not evidence of inaccuracy.

Luxi’s recorded hash matched on Apple M1 Pro/macOS, a native Linux arm64 container on that Mac, and qemu Linux x86_64. Stock python-sgp4 repeated within each platform but its hashes differed between these platforms. This is a result for that catalog and grid, not universal orbital accuracy.

The public snapshots cover H200 research from June 2026 and an H100 suite dated 11 July 2026. They include an H200 16k rag_tokenized compressible-KV fixture, H100 stack comparisons, WNSM measurements, and long-context memory scaling. PyTorch Flash won the H100 short-sequence same-shape comparison.

Sources / inspect before extrapolating

The public source index.

Pin the repository commit you inspect. Public checkers and raw artifacts support verification and recomputing reported summaries; they do not supply the proprietary benchmark engines.

Historical evidence

Read the limits before choosing a workload.

Each family needs its own contract. Output hash equality is not an accuracy check; one hardware result is not a universal support promise; one measured path is not end-to-end deployment evidence.