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.