Portability without degradation

Portability is easy to promise and hard to deliver. Exporting a model’s weights is the trivial part; the harder part is everything wrapped around it — fine-tuning history, retrieval indexes, learned preferences, tool configurations — that quietly stops working the moment you move providers.

Most of what looks like vendor lock-in is actually this: not the model itself, but the accumulated context around it that never had an export path. We’ve been mapping out what a genuinely portable stack needs to treat as first-class, movable state, rather than an implementation detail.

Our working hypothesis is that portability has to be designed in from the data layer up, not bolted on as an export feature later. More on this as we test it against real migrations.