A transcript records what a model said. It does not hold everything the model picked up about the task along the way, so a model can lose ability even when the whole conversation survives.
Different kinds of loss take different abilities with them. A model can keep a fact and still lose the habit of using it, and each loss needs its own recovery. Our July experiments found that a capable model reads a commitment perfectly while it is in view and loses it entirely once the commitment leaves view.
Abilities that outlast the model
Reading a model is a different achievement from steering it, and steering only counts when the targeted behaviour moves and every other ability stays intact.
When work moves to a different family of models, plans, finished work and evidence carry over as they are. The abilities the work depends on have to come too, so that a project’s experience outlives the model it started on.