Audio Atlas

Before this, the row was a filename in Finder. You couldn't hear it. You couldn't see what was in it. You couldn't do anything to it without leaving the catalog and opening another app. Each era above earned the row one new thing it could do for itself.

The pattern that emerged: every time the row could do more by itself, a whole UI surface could be deleted. The dedicated Listen page went away once the waveform played. The Library tab went away once Buckets could filter and sort. Workbench went away once cropping and stem-splitting moved into the card's expand panel. Tabs got added back later for the pipelines downstream of the row (Engineer, Sounds), but the row itself stayed the contract underneath.

This is the row that used to be a filename. The audio plays inside it. The metadata edits inside it. The tags you used to manage in different surfaces live in the expand panel. The row stopped being a label on a file and became the place where you work with the file.

The catalog organizes itself in two passes. Pass one is a Browse strip at the top of every Buckets page: All · Jam · Loops · Field · Tracks · Inbox. An auto-classifier reads each row's content type, source, and duration, then writes its bucket. Jams (long-form musical captures) and Field rows (voice memos plus audio extracted from videos) land in their named buckets automatically. Loops and Tracks stay curator-picked. Inbox is the residual that hasn't been auto-classified yet or that I haven't gotten to. The first catalog run auto-bucketed 1,389 field rows and 329 jam rows without me touching anything.

Pass two is the tag chip strip below it. The chips are frequency-ranked tags in the current bucket+category view. Click one and the row count narrows. The counts beside each chip update to reflect what's actually left. The system never asks you to read a documentation page about how filtering works. The chips show what's there and what filtering by it would do.

This is the part of the build I didn't know I needed until it started returning summaries. Eight years of audio I'd never gone back to became eight years of rows I'd actually read. A jam session I forgot existed surfaced because Claude noticed two 16-bar sections worth pulling. A voice memo got flagged because the same conspiratorial chorus theory had shown up six times across the archive.

A hallucination sweep runs over the whole catalog too. Whisper sometimes spirals on long files with music sections (the model keeps writing "thank you thank you thank you" while ambient guitar plays). The sweep flags rows with low unique-word ratios or repeating top-4-grams and relabels them as music-only, so the catalog stops claiming a 17-second clip is a profound speech about gratitude. The first pass relabeled 91 of 100 hallucination-shape rows. The engine isn't just classifying. It's noticing its own mistakes.

Before this existed, the same task meant opening the jam in Ableton, eyeballing the warp markers, dragging a loop region, bouncing the selection, renaming the bounce, dragging the new file into Live's browser. Five minutes of context-switching per loop, every loop, forever. Most loops never got pulled out at all because the cost of pulling one out was so high.

Now it lives inside the catalog row, where the audio already is. The crop tool only matters because the catalog already knew the tempo. Librosa runs on every audio row at ingest. When BPM and key are populated, every loop tool downstream gets to assume them. The cropper draws bar ticks. The promoter writes BPM and key into the filename. The mix engineer's renderer crossfades at section boundaries because the analyzer already found them. One analysis pass at the catalog level lights up every tool downstream.

The Promote button settles all of that. One click writes the loop into Live's own browser, normalized to -14 LUFS, named with its BPM and key, ready to drag onto a clip slot. The instinct when extending a DAW is to build a new surface inside it. The right move was the opposite. Ableton already drags audio onto clip slots from any folder you register as a Place, so the engine just writes into a Place I registered once. The DAW's own gesture does the work. No plugin, no Max device, no custom audio API.

Earlier in the build I tried the opposite approach, a Max-for-Live device that rendered the catalog inside Ableton itself. Two architectural walls killed it. The pivot to writing into the affordance instead of building inside it became the spine of the whole build, not just this feature.

The mix engineer is the second pipeline downstream of the row. Where Promote-to-Live ships a clean loop into the DAW, the Engineer ships six full-mix takes back into the catalog for me to judge. The three axes:

• focal · forward_bright vs recessed_warm
• space · punchy_dry vs spacious_wet
• dynamics · modern_compressed vs vintage_open

Six variants render per source. The Mix Faceoff shows them one A/B pair at a time, one pair per axis. Tap A or B to switch playback. Hold to commit a vote. Free-text comment per pair persists into the system. The next render on the same source reads the full comment history, so a space-axis retry that mentions a buried kick will also address a vocal complaint from an earlier focal pair on the same track. The system learns my taste one pair at a time.

Each pair you commit becomes the per-axis winner for that source. Three axes voted, three winners. The Winners surface upstairs consolidates them into one card per track. Variants stop being identified by their axis-and-pole strings ("focal · recessed_warm") and pick up stable color names instead: Citrine, Maroon, Mist, Slate, Pewter, Sienna. The catalog stores them as <source> · <Color> so you can scan a Winners card and remember which mix is which.

Three winners means it's time to crown a king. Tap ♛ on the mix that beats the other two across axes. The card collapses to just the king plus the original (so you can A/B back to the unprocessed take), with a comment box for the note that becomes part of the planner's memory next time it renders. From the crowned state you have one of three forward paths: Publish to Sounds flips it public on the catalog's own /sounds page and tags it with a collection (EP idea, album draft). Make more mixes kicks a fresh batch through the planner with the full vote history plus your crown comment in context. Uncrown drops back to candidate cross-compare. The king card is the king of the loop.

The reference catalog shipped on 2026-05-22. Bon Iver, Tame Impala, Frank Ocean, 070 Shake, others I love, each row carrying per-track notes on what works in it. The planner pulls the closest references for the source into its prompt as taste anchors before any variant gets rendered. Closest-reference is weighted similarity across spectral centroid, LUFS, band cosine, stereo width, and BPM when the ref has been audio-analyzed, and tag-affinity when it hasn't. The mixes come back already shaped by what I've told the engine matters.

The catalog is a set of rows and a set of verbs you can do to them: play, crop, split, promote, send to engineer, publish. Every interaction in every surface is one of those verbs. Surprise and Library got deleted on 2026-05-17 because Buckets did what they did with one fewer page. Engineer and Sounds got added on 2026-05-22 and 2026-05-27 because the row could now feed pipelines, and pipelines needed somewhere to live. None of those moves changed the row or the verbs. The row stayed the contract. The tabs were just where you went to use it.

That's what made the construction tractable. New surface? Take the verbs you need, glue together a new lobby page on top. Surface no longer earning its place? Hide it from nav, keep the route alive for bookmarks, move on. The row is the only thing that has to stay coherent.

Verbs are the contract. Surfaces are disposable. The catalog will still be itself the day I delete the last tab.

Every one of those is a feature an adjacent product ships. Every one is a place where the machine would pretend to author. A catalog that emails me an unsolicited "draft album concept" stops being a catalog and becomes a parasocial collaborator. That trade was off the table from the start.

Whisper transcribes. Claude classifies. The Engineer renders variants and reasons about them. The Sounds page hosts what I've published. None of them synthesize. The synthesis happens later, in Live, by hand.

Adjacent to all of this is a Diary tab that pairs random voice-memo fragments with random catalog audio, so a melody you hummed into your phone in 2023 can surface next to an instrumental sketch you made last week. Coming back to a project six months later, trying to remember what the idea was and where the pieces went, is the friction the catalog was originally built to remove. The Diary is the surface where that friction has nothing left to hide behind.

The public-facing uxjon.com/sounds page is already live. The publishing path is concrete: Send to Engineer from a track in Buckets, vote three pairs in the Mix Faceoff, crown your king on the Winners card, click Publish to Sounds, label the collection. The same database, the same row, the same verbs. The path from "this loop is interesting" to "this track is out" doesn't cross a single tool boundary.

3,750 rows still sit in the inbox bucket. 215 promotable loops are queued. The machine just made choosing cheaper.