Same bytes on two machines

A result you cannot rerun is a rumour. That is the first thing we ask of every Drylab result: given only the record of how it was made, can we make it again, exactly?

Today we tested that on a real run. It was a TP1 screen of every single-gene deletion in baker's yeast, S288C: all 5,061 genes in the model's list, ranked by the predicted chance that deleting each one is lethal. It ran in production on Cloudflare's Linux machines.^[1] Its record pins the model by checksum, the seed (0), the conditions and the engine versions. We took that record to a different machine, a Mac on an ARM chip, running a different patch release of Python, and ran it again.

Original and rerun produced the same 511,148 bytes, with the same SHA-256: 287ec280…b93f1c.

Why it works

Three choices do most of the work. The model is pinned by checksum, so a run can only ever use the exact artifact it was quoted on. The hashed document contains only science: no run id, no timestamps, nothing that would differ between two honest runs. And it is written canonically, with keys sorted and separators fixed, so the same content always becomes the same bytes.^[2]

def canonical(obj) -> bytes:
    return json.dumps(obj, sort_keys=True, separators=(",", ":"), allow_nan=False).encode()

What this does not show

This is one run, at one tier, on two machines. TP1 is a deterministic model; the stochastic tiers depend on seeds being honoured everywhere, which is a separate promise and a separate test. We report this case because it is real and checkable, not because one case proves the rule.

1. Notes
2. [1]Production run 02fe2195: TP1, S. cerevisiae S288C, all single deletions; model vcl-tp1-go-essentiality v1. Source ↗
3. [2]Drylab docs, How we check: reruns from the record alone, compared byte for byte. Source ↗