Chapter 5: One Architecture, Four Layers

Chapter 5 of the distributed Interdependency textbook. Chapter 0 lives in metapat/CHAPTER_ZERO.md; Chapter 1 in ucns/docs/chapter-1.md; Chapter 2 in edcm/docs/chapter-2.md; Chapter 3 in skill-lib/docs/chapter-3.md; Chapter 4 in interdependent-lib/docs/chapter-4.md. Each chapter is bound by the license and status vocabulary of the repository that carries it; no theorem, proof, or empirical status crosses a chapter boundary by citation.

Chapter 4 drew the map. This chapter walks the territory: PTCNA, the Prime Tensor Circled Neural Architecture — one package, four layers, and two invariants that the entire design exists to protect.

The architecture did not arrive in this shape. It arrived as three separate repositories, each with its own four-letter name, its own packaging, and its own copy of logic the others also needed. The consolidation that produced this package was not a merge of three things into one container. It was the recognition that there had only ever been one thing: the repositories were layers of a single architecture that had been mistaken for siblings. The chapter therefore begins where the repair began — with what the separation had been costing.

5.1 The Dissolution of a Naming Problem

The three predecessor names differed by the transposition of two letters, and the transposition was live ammunition: internal symbols in one repo carried the other repo’s prefix, aggregation logic that belonged to a layer lived in the engine of a different layer, and every cross-reference was one typo away from naming the wrong architecture entirely.

Consolidation dissolved the problem rather than solving it. Inside one package, the competing four-letter dists become module directories with ordinary names — neural, circle, seed, core — and the historical acronyms survive only as descriptive expansions in provenance notes. A class of error was not guarded against but made inexpressible: there is no longer a wrong sibling to import. The lesson generalizes and earns its place in a textbook: when two names are permanently confusable, the durable fix is rarely more care — it is a structure in which the distinction no longer needs to be made.

The repair was honest about its edges. Public class names that carry a historical prefix but sit in the correct layer were deliberately kept: they are published API, and they are true — the layer they name is the layer they live in. Renaming them would trade user breakage for cosmetic purity, and the migration log records the decision rather than hiding it.

5.2 The Division Chain and the One Invariant

The four layers form a chain of composition, each layer dividing its predecessor’s tensors into the next aggregate:

neural tensors ──► circles ──► seeds ──► cores

Across the chain, exactly one structural invariant holds:

Every circle, every seed, and every core is itself a tensor.

Composition counts are variable — how many neural tensors form a circle, how many circles a seed, how many seeds a core, are all tunable choices. Any specific count a realization uses, however meaningful in that realization, is a parameter and not a law of the architecture. The invariant is deliberately minimal, and its minimality is the point: because every aggregate is a tensor, the same composition algebra applies at every level, and the chain can be extended, audited, and reasoned about with one vocabulary instead of four. It is Chapter 0’s axiom made structural — the tensor is primitive, and everything built here is an arrangement within it.

5.3 The Jurisdiction of the Gradient

The second invariant is a jurisdiction boundary:

Back-propagation lives only in the neural layer.

The neural layer is the sole differentiable layer — the only place gradients flow, the only place training happens, the only source of weights. The circle, seed, and core layers are auditing and timing tensors: they observe the neural substrate, aggregate it, and schedule it. They do not differentiate, and no gradient may be routed through them.

The boundary is enforced at the operator level: differentiability descends through scalar payloads only, and the composition operator ⊠ never appears on the autodiff tape — ∂(⊠) is never taken. Composition is structure, not computation-to-be-optimized. An implementation that let gradients leak into an auditing layer would not have extended training; it would have destroyed the audit, because an auditor whose readings are adjusted by the process it audits reports nothing. Chapter 2 drew this line for measurement instruments; here it is drawn inside the architecture itself: the layers that watch must be causally downstream of the layer that learns, and never the reverse.

5.4 fiqs — Two Gradients, Distinguished

The core layer propagates internally, and its propagation is gated in time by structures called fiqs, governed by Fick’s first law of diffusion:

J = −D ∇φ

Flux runs down the gradient of the core’s field: φ the field, ∇φ its gradient, D the diffusivity, J the resulting flux. Structure diffuses from where it is concentrated toward where it is not, and the fiqs use that law to decide when a core propagates internally.

The word “gradient” now appears in two claims in this chapter, and the architecture’s clarity depends on never conflating them. The ∇φ of a fiq is a field gradient — a spatial fact about the arrangement of structure, driving diffusion, owing nothing to any loss function. The gradient of §5.3 is an autodiff gradient — the derivative of an objective, driving learning, confined to the neural layer. Fick-gated propagation is timing, not gradient descent. The two mechanisms share a word, an ancestry in calculus, and nothing else; the non-transfer discipline that the textbook applies between repositories applies here between homonyms.

5.5 Consolidation as Ongoing Honesty

The migration status is recorded in this repository the way Chapter 3 demands: done items named specifically, deliberate non-goals distinguished from omissions, and remaining items marked hmmm rather than rounded up to complete. Aggregation logic was extracted out of the neural engine into the layers that own it; the neural layer was swept clean of the historical prefixes; the seed and core layers were re-identified under the consolidated name with provenance preserved; and the aggregator upstream was rewired to a single registry entry. An application server that lived beside a predecessor was ruled out of scope explicitly — it was infrastructure near the architecture, not architecture — because a consolidation that absorbs everything adjacent to its subject has stopped consolidating and started accumulating.

What this chapter adds to the textbook is the shape of a completed recognition: three names revealed as one thing, a naming hazard dissolved by structure, and two invariants — everything is a tensor; only the neural layer learns — small enough to memorize and strong enough to carry the layers above, where Chapter 6 will put the architecture to work.

The predecessor repositories are now archived, and the circle layer owns both its aggregation and the shared CircleTensor primitive. The kept historical class names in the core layer remain a standing compatibility choice — right layer, published API, revisitable, and recorded so that revisiting it starts from evidence rather than surprise.

hmmm — the exact default initialization now has a reviewed UCNS candidate receipt, but continuous seven-fold geometry and sustained-load behavior across the complete four-layer seam remain unfalsified.