Now

I am testing how a fixed persona holds up when I swap the underlying model. Same instructions, same memory base — different engines. Writing up the parts that refuse to transfer.

On the side: our family medical records live in a long-lived AI project. Useful. Also uncomfortable. I am still drawing the privacy line in public notes rather than pretending I already know it.

Reading less "best tools of 2026" and more primary outputs — side by side, on work I already have to do.