What an incomplete HalluCodon trial still taught us
The external model never produced a candidate in this trial, but the failure exposed practical reproducibility costs that benchmark tables often omit.
The intended comparison
The goal was a low-cost zero-shot trial of HalluCodon, a plant-focused framework that combines protein and RNA language-model components with species-specific scoring. Because the target host was not included among the available pretrained species, the closest supported tobacco model was selected as an explicit proxy.
The trial followed the project's documented initialization and hallucination-based optimization path. The relevant sources are the HalluCodon preprint and its official repository.
What actually happened
The experiment moved between free GPU environments. Large model downloads, temporary storage, archive extraction, and session lifetime repeatedly competed with one another. A run eventually began but was interrupted before an optimized candidate or result table was written.
Warnings about GPU context initially looked like the likely cause. The raw log showed something less exotic: the process had been interrupted externally. That distinction only became clear when the original output was examined rather than a summary of it.
An incomplete run is not a benchmark result
No candidate means there was no naturalness score, expression score, sequence comparison, or performance conclusion. The correct status is therefore “incomplete experiment,” not a negative result for the model.
Even so, the trial exposed real evaluation costs: weight acquisition, environment reconstruction, proxy-host choice, session durability, and artifact persistence. Those conditions belong in a reproducibility record because they influence whether an advertised workflow can be independently exercised.
What a future rerun must freeze
A meaningful comparison should pin the source commit, model-weight identifiers and checksums, proxy-host rationale, container or environment lock, command line, random state, and every output artifact. Until that package exists, the useful result is the execution lesson—not a claim about model quality.