Most of today went into A Wine App, and it was a real build day — north of 11k lines. The core idea: stop scattering on-device model calls all over the app. I counted 28 separate places spinning up their own model sessions, so I collapsed them into a single language engine, then stacked a persona layer on top so the sommelier voice lives in one place instead of being re-invented per feature. From there it was bricks: label and search parsers, fact and taste extraction, the structured generators that round out the API, and a command-line harness that can fire every one of those functions from the terminal for testing.
The honest part of the day was a course correction. I switched the parsing calls to greedy sampling — pulling facts out of a wine label isn’t creative work, it wants the single most-likely reading, not dice. Then the harness caught the on-device model red-handed: it was fabricating the user’s history, inventing “the Rioja you trusted” even when the prompt forbade it. So the afternoon turned defensive — I stopped trusting the prompt and started enforcing honesty in code, detecting the fabrication pattern and refusing it, plus deterministic tests so the LLM layer can’t quietly drift. Never fabricate a user’s history is the one promise this product can’t break.
The Cloudy Brain saw the opposite kind of work — two passes that removed far more than they added, mostly content cleanup, with an idea pruned along the way. And myTeam got its routine nightly touch, a small bit of new scaffolding and nothing more.
Tonight the wine engine is consolidated, tested, and — for now — caught lying and patched to stop. Tomorrow’s job is trusting it.