Cadence

Cadence is a fitness-coaching app that adjusts its guidance to how the user actually is that day — sleep, prior training load, soreness, energy — rather than running a fixed program regardless of state. A deterministic rules engine computes a readiness score from that data; an AI layer turns the score into a written coaching response, but never overrides it.

The AI layer is told, explicitly, that if the rules engine says the user needs recovery, it must recommend recovery. That instruction isn't trusted on its own: a second, deterministic check reads the AI's actual generated text for phrases like "push hard" or "max effort," and if it finds one that contradicts a recovery day, it throws the AI's answer away and substitutes a fixed, pre-written response instead. It's a narrow, almost crude check — a keyword list, not a second model call — but it reflects a real judgment: don't let a generative system have the last word on something that could affect someone's health, even after you've prompted it correctly.

The same instinct shows up in what's missing. Error tracking hit a launch-crashing bug twice in earlier versions, both times traced to the crash-reporting library's own native init code running before the app could recover from it. Rather than fight it a third time, the reporting calls were replaced with a stub that logs to the console and does nothing else, with the failure history written directly into the code as a comment so the mistake doesn't get repeated blind.