Which model is the whole ballgame

The Recovery Log — 005

Saturday 7 February 2026

If the bleed is encoded in latent space, then the whole project lives or dies on whose latent space. This is the least romantic post we will write and it is also the most important, so settle in.

A recovery is only as good as the cultural manifold you are navigating. You need a model whose pretraining actually compressed enough of the world — enough music, enough image, enough of the texture of how people lived — that the neighbouring regions of a song are dense rather than sparse. Feed it too little and the space around a probe is empty; there is nowhere to walk to. Feed it the wrong things and every trajectory collapses back toward the average of everything, which is to say toward mush, which is to say toward a cover.

So we spent the week doing the unglamorous triage: which base models have the representational depth, which have the coverage across modalities, which we can actually run without a data centre we do not have. Every candidate failed at least one of those. The good ones were too big to touch. The ones we could touch were too shallow to hold anything.

The conclusion we did not want and arrived at anyway: we are not going to find the right model. We are going to have to build a harness on top of one — an embedding layer we control, a traversal method we understand, and a lot of tuning — so that the thing we probe is a space we have shaped, not one we rented. That is months of work we had hoped to skip. We are not going to get to skip it.