RoomEQ predicted its EQ would cut the error to 1.8 dB, then measured 6.8 dB. Four sweeps without moving the phone found the real cause, and changed how auto-tune compares before and after.
Why generating HTML email markup is a genuinely different problem from generating a web page, and how a four-stage pipeline with chained AI calls handles it.
Why a phone terminal needs a real PTY and a real ANSI emulator on both ends of the wire, not just a text box that appends strings.
Real numbers from skull's memory system show why a single cosine-similarity cutoff can't safely tell 'the user changed their mind' from 'the user mentioned something related' — or from the wrong fact entirely.
A terminal AI agent that writes its own tools at runtime — versioned, sandboxed by static analysis, and composable — and the failure modes that shaped each piece.
A path-based signature scheme for catching an LLM agent stuck calling the same tool with the same arguments, without blocking legitimate repeats in other branches.
How an unlock request from a phone becomes an actual relay click on an ESP32, and what stops a replayed MQTT message from opening the door twice.