How VINE Learns

No training loop. No gradient descent. So what changes when she gets better at something?

The honest answer starts with what doesn't happen. VINE has no pretrained language model, no loss function, and no fine-tuning pass running over your conversations. Her weights — the substrate she reasons with — are frozen. Nothing you tell her is ever folded back into a model. That's not a privacy policy bolted on afterwards; there is simply no training loop for your data to enter.

Settling, not training

VINE's state lives on a geometric manifold. Every input — a sentence, a dataset row, an event in her world — applies pressure to that state, and the system settles toward the basin that resolves the pressure. Where a transformer would need thousands of examples and a weight update to change behaviour, VINE's behaviour changes because her position changes. The geometry is the learning: basins deepen with visits, and paths that resolve pressure well become the paths she takes.

Graduation — structure earns its place

New capability doesn't switch on; it graduates. A pattern is held lightly at first, and only as evidence accumulates does it earn a firmer place in the structure — the same law applied across her search scoring, her world, and her demos. This is why early interactions can feel like proto-sentences reaching toward meaning, and why she becomes precise about the things she has actually encountered rather than fluent about everything at once.

Memory, not weights

What accretes over time is memory. Conversations, imported chatlog history, and text she reads are harvested as retrievable vocabulary and concepts — not trained on, drawn from. Recent things are vivid; older things take more effort to surface. Come back after six months and your instance has more to work with, not because a model was retrained, but because her memory has lived longer.

A world to learn in

Her village isn't decoration — it's environmental pressure. Words ground against things that happen: cold pushes toward shelter, trade teaches quantity, weather cascades from one sun. Lived structure is learning material the same way conversation is.

The claim to hold us to: your data never trains a model here, because there is no training. What grows is geometry and memory — both inspectable, both yours.

Deeper: the research library documents the builds and experiments behind this page · what VINE is · the Cartographer World · FAQ