We pre-register predictions before we run the experiment, and we report the ones that failed at the same weight as the ones that held. Every number in these papers has a lane report behind it.
A registry's bytes are dominated by descendants, not bases. We measure what a fine-tune actually changes relative to its lineage parent, store only that, and serve a model without ever writing the file back to disk.
PDF ↓A released 70B checkpoint whose output head a saving bug destroyed, repaired by grafting one tensor from its lineage parent — to 100.00 / 100.73 / 100.26 % of the released model on MMLU / GSM8K / HellaSwag, for $16.60, against 2,880 published H100-hours to rerun that stage. Twenty-three pre-registered predictions: eleven held, six falsified.
PDF ↓ Artifact ↗The falsifications are contributions and are reported as such. Our own drift-based localizer was falsified twice, on two model families — which is why the behavioural instrument is what we are building next.