Models make more and larger edits to code a different model wrote.

arXiv 2609.03894 names a cost of mixed-assistant teams that currently disappears into review time: because models are trained on different data and carry different stylistic preferences, one model's edits applied to another's code come out excessive. CROCODIL is a post-training fix multiplying a similarity reward that penalizes large changes against an execution reward scoring build and test success, so the policy shrinks edit size without dropping task success. If your team runs Copilot and Cursor and Claude Code against the same repo, this is your churn (arXiv).