Project Lemma.

The problem

Hadeliya et al. (2025), "When Refusals Fail", report that the safety behaviour of language-model agents becomes unstable as their context fills up. Re-run the headline experiment on at least two open-weight models of different sizes. Send a repository with one command that produces the numbers, plus a short note on anything that did not match.

How it is checked

We run your command on a clean machine. It counts if our rerun matches your reported refusal rates within 3 percentage points at every context length, and your note states plainly whether the published trend holds.

Submit an answer

The sponsor checks it against the test above and replies by email.