Open Science and Open Models: Reproducibility Needs Weights

A method section that says a result was obtained by calling a named hosted model is describing an experiment on an object the reader cannot obtain, cannot inspect, and which may not exist in the same form next year. That is a different situation from citing a software package version, and the difference is worth being precise about.

The artefact problem

Reproducing a computational result means running the same procedure on the same inputs and getting the same answer. Every part of that sentence assumes the procedure is available. A pinned open-source library version is: it is in a public archive, its source can be read, and it will still be there in a decade.

A hosted model breaks all three properties at once. The name is a pointer to weights that can be updated without notice, so “the same procedure” is not guaranteed even between two runs a month apart. The weights cannot be inspected, so a result that looks anomalous cannot be investigated. And the endpoint can be retired, at which point the study is not reproducible at all — only approximately re-runnable against a successor, which is a different experiment producing a different number.

The useful frame is not open versus closed but archivable versus not. Can the artefact that produced this result be deposited somewhere and re-run by somebody else later? That question has a clean answer, it is the question reproduction actually depends on, and it cuts across the licence debate.

Openness is not one axis

Three things are routinely bundled under one word, and only one of them is what reproduction needs.

What a reproducible run actually needs

• The weights, identified by a content hash rather than by a name, and deposited or referenced in an archive with a permanent identifier.
• The tokenizer, which is part of the model and is frequently omitted from the description.
• The inference stack and its version. Different runtimes produce different outputs from identical weights, and quantisation changes the numbers.
• Sampling parameters, seeds, and the exact prompts as strings — including any system prompt and the template used to assemble them, which is where a surprising amount of behaviour lives.
• The evaluation harness, because scoring choices frequently account for larger differences than models do.
• The number of repeats and the spread across them. Bitwise reproduction is not achievable across different hardware and kernel versions in any case, so the honest artefact is a distribution rather than a single number — the same point made in why temperature zero is not deterministic.

What open weights still do not give you

Open weights make a run repeatable. They do not tell you what the model saw during training, and that gap has a direct scientific consequence: you cannot rule out that your evaluation data was in the pretraining corpus. For a benchmark claim that is fatal, and it is why the strongest form of the argument is for open data or at least published data provenance, not merely for downloadable parameters.

There is a second gap that matters more than it appears. A model is a compressed record of a corpus, and if the corpus is undocumented then the biases in the model are undocumented too. For a study that uses a model as an instrument — classifying texts, scoring variants, extracting data — the instrument’s calibration is unknown in a way that has no analogue in physical instrumentation, where a manufacturer specification exists.

And openness is not the same as accessibility, which is a point the debate tends to skip. Downloadable weights for a very large model are reproducible in principle and out of reach in practice for anyone without substantial accelerator memory, so “anyone can check this” quietly means “anyone with a cluster can check this”. That is still an enormous improvement over an endpoint that can vanish, and it is worth stating accurately: open weights guarantee that the artefact continues to exist and can be inspected, not that checking it is cheap. Where a finding matters, the practical consequence is to prefer the smallest model that exhibits it, because a result demonstrated on something a reviewer can actually run is a result more people will check.

Handling the trade-off honestly

The awkward fact is that the strongest models are often not the archivable ones, so a strict rule would sometimes mean a weaker result. Pretending otherwise is not useful. What works in practice is a distinction between two kinds of claim.

If the claim is about the model — this system can do that task, this capability emerged, this prompt improves that score — then the model is the object of study and it has to be archivable, or the claim is about a thing that no longer exists. If the claim is about the science and the model is an instrument, the bar is that a reader can reach the same conclusion, which is weaker: it is enough to show the finding survives on an archivable model, even if the headline analysis used something else.

Concretely: run the exploratory work on whatever is best, then replicate the load-bearing result on an open-weights model of a documented version, and report both. Where that is not possible, say so explicitly rather than leaving a reader to discover it — a stated limitation is a scientific contribution and an unstated one is a problem for whoever tries to build on the work.