A timbral axis measured from the audio, and the map it would make worth drawing

[robert-dean]

Written: 2026-08-28, promoted out of comparable-stations, where it
was ranked 7th — the largest item on that list and the only one that unblocks a file currently marked
blocked. It had no file of its own, which for the biggest deferred thing in the directory is the
wrong way round.

It is the answer to a question station-moment gives up on. That file says mood-biased
selection is blocked on data nobody has: EnrichedTrack.moods exists in the plugin SDK
(capabilities/enrichment.ts:150), nothing fills it, and the MusicBrainz plugin explicitly declines
to. The proposed source is tags, which makes a library only as steerable as somebody else's
folksonomy — and on a library of records nobody tagged, that is not steerable at all.

The other implementation gets it from the audio: a joint text-and-audio embedding per record,
scored against a mood vocabulary written in words, plus nearest-neighbour lookup in the same space.
That is a real answer to "nothing populates it", because it needs no tags and no upstream, and it
gives the station a timbral axis it has nowhere else.

The calibration trap, which is measured and is the reason this is written down

Similarity scores from that kind of model are not comparable across different prompts. On an
11k-track library that is not this one, scoring every track against every mood and taking the top
scores per track produced energetic on 61.7% of the library and rainy on 43.5%, against
calm at 0.9% and spiritual at 0.2%. Ranking raw scores ranks the PROMPTS, not the tracks.

The fix is a per-mood baseline over the whole library — mean and standard deviation — and then
ranking on the z axis, with the margin expressed in standard deviations rather than in raw score.
Two second-order details came out of running it, both of which look like polish and are not:

• The minimum sample size is applied per mood, by pruning that mood, never by gating the whole
  pass on the largest one. Gating reproduces the original bug one mood at a time: a thin new mood
  beside a fat old one wins every record.
• Ambiguity returns null, never a middle bucket. Bucketing an unclear result to "medium" replaces
  one guess with another and then calls it evidence. Null leaves the existing value alone, so
  mistuning costs coverage instead of correctness.

Do not start this without the calibration rule. A first pass that ships raw scores produces a
library where two thirds of everything is energetic, and the station will act on it.

What it costs here, counted before anyone starts

• There is no vector column and no pgvector. Nothing in apps/api/data/migrations declares
  either. That is a migration and an extension, and the extension is an operator's problem on every
  install rather than only on this one — which matters more since the repository is meant to ship.
• The sidecar grows a second model, which puts it straight into the analysis licence rule. The licence that matters there
  is the weights', and it is not in the package metadata. That file is not optional reading for
  this.
• ANALYSIS_SCHEMA_VERSION absorbs the output shape for free. track_analysis.data is jsonb
  with a version, and 0005_music.sql says in as many words that this is what lets a deferred layer
  land with no schema change. The EMBEDDING itself is the part that wants a column, because a blob in
  jsonb is not a thing anybody can search.

And one rule that has to be right on the first pass, because it is invisible afterwards: the
tagger's own output must never be part of what gets embedded. The labels are derived from the
vectors, so feeding them back is circular, and the symptom is a library that agrees with itself more
every time it is re-measured.

The map is downstream of this and is not a separate item

The other station draws a full-screen map of every measured record, placed by genre and lit by
energy. Do not build the map first: it has nothing to draw. What makes one worth having here is
the same thing that makes the calibration trap visible — a mood that took 61.7% of a library is a
number in a report and an obvious stain on a map, and recolouring by CONFIDENCE rather than by label
is the same argument the trace surface makes for a budget: it is hard to argue about a threshold
without seeing what it did.

What this unblocks, and what it does not

It unblocks station-moment, which is the only file in this directory whose
status is "blocked on data nobody has".

It does not unblock the beat layer, which is a different measurement with a different model and is in
track-analysis. The two share the sidecar and nothing else, though they share
one operational note: see that file on a long-running Python process's resident memory.

---

Moved from docs/todo/embedding-axis.md. The file as it last stood: https://github.com/robert-dean/deadair/blob/71e431d45ce7b35a189c7d226832ec4a6247e9dc/docs/todo/embedding-axis.md

[robert-dean]

Status check, 2026-09-22 (main at 23b983e)

Still unbuilt: no pgvector or vector column, and no embedding model in analysis/ (the torch dependencies are for beat_this). One premise has changed: moods is no longer empty, because the Last.fm plugin fills it from listener tags (lastfm.tags.ts).