Hadal Research
LAB NOTE
27 October 2026
4 min read

← Notes

What a CPU-only host can and cannot tell you

Hadal Research

method · reproducibility · compute

Environment recordPaper (PDF)

A lab note on the machine behind Changing One Word, because the hardware explains most of the design choices in it.

The host

itemvalue
CPU2 vCPU · Neoverse-N1 (aarch64)
RAM11.9 GB
GPUnone
precisionfloat32 on CPU throughout
modelQwen2.5-1.5B-Instruct + 0.5B control
analysis depsnone — standard library only
From docs/ENVIRONMENT.md in the study repository.

What the host can carry

At 4 bytes per parameter, a 1.5B model is about 6 GB of weights before any activations or KV cache. On 11.9 GB with no GPU that is the practical ceiling — not a preference. A 7B model does not fit, and a quantised 7B would change the thing being measured. So the study is a 1.5B study, with a 0.5B control, and it says so rather than implying a family-wide result.

One trap is worth recording: the 1.5B checkpoint ships in bfloat16, and loading it with dtype=float32 doubles peak allocation and sends the machine into swap. The checkpoint has to be converted to float32 on disk first. Without that step the run either fails or takes long enough to be abandoned mid-grid — which produces a partial grid, not a result.

Why the analysis path has no dependencies

The interval routine is the part of a small study most likely to be wrong, and the least likely to be read. Keeping the reported path to the standard library means a reader can run it on any machine with Python 3.10+ and get the same numbers in under a second. statsmodels was installed separately to cross-check the interval implementation; it is not required to reproduce a single reported figure. That separation is the point: a dependency-free path is one you can actually check.

It also made the v6 correction possible. When the Newcombe routine was found to be missing its n/2 covariance term and failing on zero marginals, the correction changed verdicts in the paper. The correction log records what changed, and the regenerated tables come out of the same raw records — nothing was re-run at the model to make the numbers agree.

What it cannot tell you

Twenty-four items per cell is enough to establish a sign and not enough to bound a size. Nine cells with uncorrected comparisons are probes, not a confirmatory design. One model family, one position, one wording family. Every one of those limits is stated in the paper rather than left for the reader to infer, because a small result that is honest about its scope is more useful than a large claim that cannot be rebuilt.

Sources
Status

Lab note. Describes the conditions a result was produced under, not a new claim.