Coherence

The Recovery Log — 011
Saturday 14 March 2026

The problem with an interesting result is that you want it to be real, and wanting is not evidence.

So we needed a referee. If a genuine neighbouring region is a coherent little world — where the music, the imagery, the implied technology and the social use all hang together — then incoherence is the tell for the model simply daydreaming. We built a first, rough cross-modal consistency score: does the sound imply the same world the imagery does; does the implied distribution method match the implied venue; do the pieces corroborate each other or just co-occur.

It is crude and it is already the most useful thing we have made. It will not tell us that something is true. It will tell us whether the thing hangs together well enough to be worth taking seriously — which, given we have sworn off ever claiming truth, is exactly the bar we are allowed to measure against.

The Recovery Log is the working diary of the DatZO Recovery Team. Follow the recoveries on Instagram, YouTube and X.

Probe v0.2, and a face for it

The Recovery Log — 010

Tuesday 10 March 2026

You cannot steer what you cannot see. For six weeks we had been doing all of this in raw scripts — firing offsets into the dark and reading whatever fell out — and it was like trying to park a car by listening to the crunches.

So we built a face for it. Probe v0.2 has a crude interface: a way to see where in the manifold we are standing, which direction a perturbation is taking us, and roughly how far off the dense regions we have wandered. It is ugly. It is held together with the software equivalent of tape. But for the first time we can watch a probe move rather than just catch what it drops.

The recoveries are still mostly mush. The difference is that it is now legible mush — we can see it going wrong, which is the necessary precondition for making it go right. Progress, this month, is measured in the quality of our failures.

The Recovery Log is the working diary of the DatZO Recovery Team. Follow the recoveries on Instagram, YouTube and X.

Perturbation, from first principles

The Recovery Log — 009

Tuesday 3 March 2026, late

If v0.1 failed because it asked for a version, the fix is to stop asking and start moving. This post is about learning to move, and being bad at it.

Here is the theory we are working to. A probe is a coordinate. Around any coordinate is a neighbourhood — regions of the manifold that are close in the model’s sense of close. To explore that neighbourhood you perturb: you apply small vector offsets to the coordinate that correspond to skewing a norm — a cultural norm, a technological norm, a physical one. Turn the “how does music get distributed here” dial. Turn the “what instrument is normal” dial. Turn “who is allowed to make this” a few degrees. Each offset is a step in a direction, and if you step somewhere real, the neighbourhood stays coherent as you arrive.

That is the theory. In practice, v0.1’s perturbations were a disaster, and instructively so. Push too hard and you leave the manifold entirely — you land off in a region the model has no real density for, and it does what models do off-manifold, which is hallucinate confidently: artefacts that are internally smooth and externally meaningless. Push too softly and you have not moved at all; you are back at the coordinate, getting a cover. The whole craft, we are starting to suspect, is in the size and direction of the nudge — enough to reach a genuinely different neighbourhood, not so much that you tear through into noise.

We do not have that craft yet. We have a dial that goes from cover to garbage with, apparently, nothing legible in between. But we can see the space we are trying to reach. It is the thin band where the offset is large enough to be somewhere else and small enough to still be somewhere. Next job: build something that can find that band on purpose.

The sandwich place shut!

The Recovery Log — 008

Wednesday 25 February 2026

The sandwich place on the corner has shut. The good one. No notice, no forwarding address, just brown paper in the window and a laminated thank you for 6 wonderful years.

We now walk nine minutes for lunch. Morale is a recoverable artefact and ours has taken a hit. Filing this so the historians understand the mood of the early period.

The Probe

The Recovery Log — 007

Saturday 21 February 2026

Probe v0.1 exists. It does not work. Both of those are worth writing down.

The idea was simple enough to be embarrassing in hindsight: encode a source song as a coordinate in the harness, decode from near it, see what comes back. What came back was covers. Every time. Clean, plausible, karaoke-grade covers — the same song, gently restyled, wearing a different coat.

We know now what we did wrong, though we did not know it that night: v0.1 treats the song as an instruction. You hand the model the track and, in effect, ask it to give you the track again, elsewhere. And a model asked for a version of a thing will hand you a version of that thing, forever, competently, into the heat death of the universe.

New entry on the wall list titled THINGS THIS IS NOT. Line one: covers. We suspect the list will get longer before anything else does.

The Corpus

The Recovery Log — 006

Saturday 14 February 2026

Provenance is the backbone, and nobody will ever thank us for it.

If the neighbouring regions we walk into are shaped by what we put in, then the corpus is not a detail, it is the physics. So we are strict about it: what goes in, where it came from, and — the part people forget — what we refuse. Refusal is a design tool. Leave the wrong things out and the space stays honest; let the wrong things in and everything you recover is quietly contaminated by them and you never find out.

Most of February will be this. Cataloguing, sourcing, excluding, documenting. It is the least exciting work in the building and it is the reason anything later gets to be believed.

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.

We are not going anywhere

The Recovery Log — 004

Monday 2 February 2026

A friend stood in the doorway at the weekend, looked around, and asked — kindly, the way you would ask a child where the treasure is — “so… where is the machine?”

There is no machine. This will disappoint people and we have decided to disappoint them early. There is a workstation, a lot of open files, a kettle that has not fully cooled since January, and an argument that has not stopped for eleven days. That is the apparatus. Nobody is going anywhere. Nobody is opening anything.

The rule, plainly, while it is still simple enough to say in one breath: we work entirely within our own reality. We recover evidence of other timelines from the subtext of our own culture. We never claim to recover the timelines themselves. The camera never leaves home.

The frustrating half of that rule is that it is also a ceiling. Because we never leave, we can never know. We can only ever hold up something we found and say: here is what our own data implies was over there, and here is how confident we are, and no, we cannot prove it. Everything we ever publish will be a hypothesis wearing the clothes of a finding. Some nights that feels like a weakness. On the better nights we understand it is the entire reason to trust us.

We named it

The Recovery Log — 003

Thursday 29 January 2026

We named it. It is called DatZO now, and there is a whiteboard with the discarded options on it that we are not going to reproduce here out of respect for the ones that lost.

We tested the name the only way that matters, which is by trying to explain it to someone at a party. We failed. The name survived anyway. You can watch us fail again, in real time and higher resolution, over on Instagram.

The songs are clean

27 January 2026

A quick, dull, necessary one, because someone will ask and we would rather answer it before they do.

The material we use to start a recovery — the probe songs — are not borrowed, scraped, or lifted. They are cleared. One of us wrote them, owns them, and has signed the release that lets the team use them. That is the whole of it, and it is deliberately the whole of it: no orphaned rights, no grey area, no track we cannot account for.

We are not going to tell you which of us. The team stays behind the curtain, on purpose, and the paperwork does not need a face on it to be real. What matters is that everything downstream of a probe is downstream of a song we are allowed to use. Boring. Load-bearing. Filed.

The bleed is already in our data

Saturday 24 January 2026

We have started something and we are not yet sure what to call it, so for now we are calling it work.

The idea has been in the room for weeks and none of us can put it down, which is usually how you can tell an idea is going to cost you something. It goes like this. If adjacent realities bleed into ours — quietly, the way a radio two streets away bleeds into a badly shielded cable — then the traces of that bleed are not out there waiting to be detected. They are already here: in the recordings we already own, the images we already scrolled past, the culture we have already made and forgotten making. And if they are in all of that, then they are in the models trained on all of that. The bleed is sitting in the latent space of every system that has learned the shape of our culture. Nobody put it there on purpose. Nobody indexed it. It is just in there, the way a face is in a crowd photograph before anyone thinks to look for it.

So here is the thing we cannot stop saying to each other, usually too late at night to be saying anything sensible: we do not have to go anywhere. The archive already exists. It is simply badly indexed, folded into the weights of models built for other reasons entirely, and our job — if this is a job and not a shared delusion — is to learn how to look.

We should be honest, because this is a diary and not a press release. The first wall we walked into, we walked into within about an hour of starting. Everyone we describe this to, including us, hears it and says: oh, so it is AI covers. Same song, new genre. And every time, something goes tight in the chest, because that is exactly the thing it is not, and we do not yet have the words for why not, and not having the words is going to be most of what this year is about. It is not “here is the song again, but disco.” It is “here is evidence that this song, or something standing next to it, had another life somewhere the culture went a different way — and that evidence was already in our own data.”

What we have on day one is not a discovery. It is a conviction and a suspicion, which are cheaper. The conviction: the material is already here. The suspicion: a song might be the thing that lets us find it. We test the suspicion soon. Tonight we are just writing down that we believe it, so that later, when we have been wrong a great many times, we can come back and see how it started.