How Deforget is built - the honest version. Occasional, technical, and free of announcements.
August 12, 2026
CloudKit lost my diary's words three different ways - a launch-import race, last-writer-wins, and a silently wedged sync engine. The shadow copies, paragraph unions, and three-sentence repair doctrine that survived all three.
August 3, 2026
I added a prompt-injection fence to the extraction pipeline and the eval harness billed it at seven points of recall. The model got scared, not safer - and the fix was six words and a stack of deterministic guards.
July 28, 2026
One corpus, real hardware, every iOS 27 beta, the release candidate and the public build. Raw model scores, repaired pipeline scores, model states that trade places under one OS build, two different models on two devices, and the thirty-odd points in between that stay mine.
July 16, 2026
Running Apple's on-device model against a private diary: what a small model gets wrong, why the model proposes and deterministic code disposes, and how I measure the seam between them.