The producing process forgets the map
We have hours of audio rendered from real observations, and for much of it the link between second and pixel was thrown away at render time. It cannot be reconstructed by wanting it.
There is a gap at the centre of this work, and naming it precisely took longer than it should have.
We have a large archive of audio rendered from real astronomical data — JWST cubes, DESI spectra, months of studio passes. We also have a pipeline that can prove, for a given window of audio, exactly which measurements it corresponds to.
These two things do not currently meet, for a reason that is entirely our own fault:
The producing process forgets the map; the atlas refuses to invent it.
When those WAVs were rendered, the renderer knew the correspondence. It had to — it was stepping through the data and emitting frames. But it did not write that correspondence down. It wrote the audio. The traversal-to-frame map existed for the duration of the render and then went away.
The obvious repair is to infer it: the file is 24 minutes, the sweep had 1,400 steps, so each step is about a second. That inference is usually approximately right, which is exactly what makes it dangerous. It is also, in the atlas’s vocabulary, a quarantine rather than a success — recorded as assumed_from_duration, which never qualifies a map, and which cannot even be constructed without writing down the assumption in prose.
So the archive splits into tracks. Qualified renders carry a map that is ground truth by construction, or a recovered map that passes golden frame-boundary tests. Listening renders are the production audio without a verified map — perfectly good for radio, for human audition, for mining motifs as audio, and not admissible in anything that claims a correspondence. Linked renders are listening-grade audio pointing at a qualified observation: same source file, different map grade, no pretend precision.
The rule that governs all of it is one line: a qualified claim may not be built on a listening-grade sync.
What we want to underline is that this is not a data-loss story with a technical fix pending. Nothing is corrupted. The audio is fine and the FITS files are fine. What is missing is a relation between them, and a relation that was never recorded cannot be recovered by wanting it badly enough. It can only be re-derived by reimplementing the producing process exactly and proving the reimplementation agrees — which is a real project, or re-rendered, which is cheaper and less interesting.
The lesson generalises past astronomy: when a process transforms A into B, the transformation is usually the most valuable artefact and the one nobody thinks to save.