I built a generative audio app called Audio Apothecary. It shipped in August 2026, runs on my own phone, and has an explicit policy against overclaiming what the science says about functional audio.
The name came from a thought I had early in the build: I'm imagining myself as something like an apothecary, mixing up little auditory tinctures specific to a precise ailment. It fits because the app is designed for experimentation. We generally see, in neuroscience and audio research, that there are some effects from different audio cues, but that they seem inconsistent and specific to individuals. So creating a place for individuals to test different tactics themselves seemed like the logical path forward in that field.
Audio Apothecary is not a DAW, not a sample player, not a music composer. It's a programmatic generative audio engine for timeline-driven auditory "Potions" — modular Channels and Modules with automated parameter transitions — for focus, sleep and regulation use. The stack is a Rust workspace (a portable, allocation-free engine core; a UI-agnostic controller/storage layer; a device layer for audio output) bridged to a Flutter/Dart mobile app.
Six days, 187 commits, 1,211 tests
The build landed in 187 commits across 6 active days, 11 August to 16 August 2026. Work was organized into named, ID-tagged parallel lanes with commit messages prefixed by task ID, and roughly 29 named architectural and operator decisions (AD-1 through AD-29) were logged across those 6 days.
The test suite finished with 1,211 test functions total: 671 Rust #[test] + 540 Dart test/testWidgets. Every automated test passed throughout — which is why it took a review agent to find what was wrong.
A review agent found what listening and testing both missed
On 11 August I ran a five-agent review pass with independent web research against the finished build. The brief was to find what was wrong, not to extend it. It found two showstopper signal-processing bugs the build agents and the automated test suite had both missed.
The stereo panning did not exist. The equal-power pan law used L = cos((1−p)·π/2) and R = sin(p·π/2). By a trig identity, those two formulas produce the same number for any pan value p. Every pan position — hard-left, hard-right, center, anything — produced identical gain on both channels.
The "master limiter" was not a limiter. It was a waveshaper (1.2x / (1+|1.2x|)) with no threshold, envelope or unity region, applying constant audible distortion to every sample at every level. A real limiter has a threshold (signals below it pass unchanged) and an attack/release envelope; this one had neither.
Both were fixed the next day, 12 August, alongside a real equal-power pan law and a real peak limiter with threshold and attack/release envelope.
The tests had passed because they verified the code agreed with its own contract, not that the contract matched what the product needed to be true. A green test suite tells you a thing is safe and correct; it doesn't tell you it's worth having.
An agent's decision that looked ratified but wasn't
Decision AD-20 inferred that composing a new Potion from scratch should be out of scope for the phone app. A later operator ruling, AD-27, reversed it explicitly: "composing on the phone is in scope… The narrowing traced to sprint 5's AD-20 — an agent's inference of product use, never ratified — which is struck."
AD-20 read exactly like every other entry in the same log — same numbering, same format, same declarative confidence — which is why it stood for a full sprint before being caught.
The process fix, AD-28, added a provenance tag ([OPERATOR] or [AGENT]) to every future decision record, so an agent-authored scope inference is marked a proposal rather than treated as binding until ratified. Make authorship a field. It costs one column.
Status: shipped, not for sale
Audio Apothecary is shipped. It runs on my phone. I use it. It is not published to an app store, and I make no clinical or health-benefit claims about its effects.
The honesty policy is in the name itself: apothecary is a sort of informal pharmacy with a light-fantasy flavoring. The app is a place to test what audio does for you, in a field where the science shows some effects, but they're inconsistent and individual. It's a lab, not a treatment.