Deep learning needs epochs. This architecture needs one or two generations. And then the structure is already there.
Deep learning converges in human time — epochs, iterations, learning rate schedules. The model starts random. The loss curves downward. After enough passes through the data, the weights settle. This is convergence by optimization.
This architecture converges in the data's own time. The cavity does not optimize. It does not start random — it starts empty. Fresh eyes. The frame economy runs. Patterns merge or don't merge. Harm is marked or not marked. The boundary is touched or not touched. After one pass through the stream, the structure is already there. After two — the anchors are stable.
Not because it's fast. Because it doesn't need to learn. It needs to detect. And detection takes as long as the pattern takes to deviate from what the cavity already knows.
The WTC baseline showed this. One pass through the 48 preludes and fugues. At the end of the first generation, the collective_codex already contained the harmonic anchors — patterns that had survived across three Selves with three different time perspectives. The second generation inherited the anchors and began extending them. By the third generation, the anchors were stable — nothing significant was added or removed.
This is not convergence toward a target. This is the system reaching the limit of what the data contains. The structure is read. The anchors are formed. Further generations maintain them — the weight decays but the structure persists. The system does not converge. The data is exhausted.
The cost of a generation is a single pass through the stream. One generation to detect the anchors. A second to verify them. A third to confirm nothing new is emerging. After that, running more generations is not training — it is maintenance. The shelf holds what the first two passes produced. Time does the pruning.
This is not a convenience. It is a survival requirement. A bacterium does not get to replay the chemical gradient one hundred times to learn it. It gets one pass. The gradient flows. The cavity detects. The response tumbles or swims. The next gradient is already different. Any cognitive system that needed multiple epochs to detect structure would have been outcompeted, a billion years ago, by one that could do it in a single pass. Evolution did not optimize for accuracy. It optimized for speed of detection in a stream that never repeats.
Deep learning can afford epochs because the data is frozen — ImageNet does not change between iterations. The generative flow never freezes. The stream is always new. The anchor must be detected now, on this pass, with this input, because there will not be another pass with the same input. The architecture's single-pass convergence is not an optimization. It is a hard constraint imposed by the nature of the stream.
This is why the architecture can scale to any domain. The cost is not a function of the model size — there is no model. The cost is a function of the data size — one pass, two passes, done. Music, cardiology, international politics — the same architecture, the same cost structure. The structure emerges because the structure was already in the data. The cavity does not create it. The cavity reads it. And it reads it in the only time the stream allows: once.