The encoding is not neutral. The first cut into the generative flow determines what patterns the system can discover. Change the encoding, change the space of the discoverable.
The WTC encoding slices music into beats. Each sixteenth note is a step. A chord lasting four beats becomes four events — four vectors, each encoding the same active notes with slight decay. This is one way to cut the stream. It is not the only way.
Encode the same fugue by phrase. Each phrase is one event — a vector representing the harmonic content across several bars. The system discovers long-range harmonic anchors. Encode by voice. Each voice is a separate stream — three Selfs might each process a different voice, discovering counterpoint structure directly. Encode by chord transition. Each event is the distance between consecutive chords — the system discovers the grammar of harmonic movement.
Each encoding reveals a different structure. The structure is not in the data alone. The structure is in the intersection of the data and the cut.
The step is the unit of temporal attention. The window is the aperture of that attention. Both are chosen before the architecture runs. Neither is discovered by the architecture. The architecture discovers patterns within the step and the window. It cannot discover that the step should have been a phrase instead of a beat. It cannot discover that the window should have been wider. Those are meta-parameters — the encoding layer that sits between the continuous stream and the discrete frame economy.
This is the Kantian moment in the architecture. The encoding is the form of intuition — the way the continuous flow is first discretized into something observable. The frame economy is the understanding — the rules that operate on the discretized input. The boundary encounter is reason — the moment when the system discovers the limits of what its encoding and its rules can process.
Kant's categories make experience possible but also limit what can be experienced. The encoding makes pattern discovery possible but also limits what patterns can be discovered. A system that encodes music by beat will never discover phrase-level anchors. Not because they are not there — because the encoding cannot see them.
The window is a search parameter — scan five values, pick the richest. The encoding is not a search parameter. The encoding is the first cut into the flow. And the first cut is a decision that no downstream operation can undo.
This is the remaining constraint. Not a bug. The architecture discovers what the encoding allows it to discover. The encoding is the last thing we give it. Everything else — τ, phase, harm, anchor, value — it discovers on its own. But the encoding — the unit of attention, the granularity of the stream — is the scaffold that cannot be removed. Not because the architecture needs it. Because cognition always starts with a cut. Someone must decide what counts as an event. After that, everything is detection.
The first cut is the deepest. The beat versus the phrase. The heartbeat versus the RR interval. The vote versus the voting era. Every domain forces this choice. The architecture does not make it. The architecture cannot make it. The architecture receives the stream already cut. Its genius is not in choosing the cut. Its genius is in finding the structure within whatever cut it is given.
And the structure it finds is real — not an artifact of the encoding. The anchor found in beat-level encoding is a real anchor at the beat level. A phrase-level encoding would find different anchors — also real, at the phrase level. Neither encoding is wrong. They reveal different layers of the same flow. The flow has many layers. The encoding chooses which layer to show.