Company / The system underneath
One structure runs underneath most of this site: a fractally recursive, holographic address space held in superimposed states. It is bounded in extent and unbounded in depth — it grows inward, never outward. It was not designed and then applied; it was extracted from the earliest work and kept surviving contact with the next thing.
First, a correction
You will see 7³ on the lattice page, 3 × 3 × 7 on the decomposition page and in the key system, seven axes in the explanatory basis, seven entity kinds in the generated worlds. It is tempting to conclude that the specific integers are the intellectual property. They are not.
Those are the widths chosen for the current instantiations, because they suit the domains and because sharing a width lets the systems compose without a translation layer. Any numbers work — provided the underlying principles below are satisfied. The Arc Lab already demonstrates this: it accepts bases from two to twelve axes and the mathematics is identical, only the basis the input is projected onto changes.
What must hold is the structure. Get the principles right and the dimensions are a configuration decision.
N × N × N templates resolve to N × N × N shapes, stored in Nn states.
A template declares the structure — the axes, the widths, the kinds. It resolves into a shape of the same form: a concrete, addressable instance occupying that space. The shape is then held across Nn states — n being the depth of descent, since every cell opens another full level beneath it. Choose N to suit the domain. The mathematics does not care, and neither does the implementation.
Because the current instantiation is 1 × 3 × 7, and each factor is a deliberate ceiling rather than a preference.
One engine. A single evaluator, not a committee of specialised models. If cognition is split across several engines you can no longer attribute a result to the structure, and the audit trail stops being one trail.
Three axes. The minimum basis for a decision that has a cost, a payoff and other parties in it — and, not incidentally, the number of spatial dimensions a person can actually hold a structure in.
Seven realms of complexity, maximum. Seven is a ceiling on how many distinguishable things can be reasoned about at one level before they stop being distinguishable. Past that, you are not adding resolution — you are adding entanglement, which the Arc Lab will measure and report back as a bad basis.
Not from theory. It came from a chess variant — Cubex³,
built years before any of this was infrastructure. The board was 8 × 8 × 8 and the
evaluator needed a true centre: a single cell equidistant from every
face, that a piece could occupy. An even-edged cube has no such cell; its middle is a
corner shared by eight. Going odd produced one, and the code still carries the fix —
ORIGIN_COORD = [3,3,3], the centre of a 7-wide cube.
Seven rather than nine is the ceiling argument above: past about seven, items at one level stop being distinguishable. That seven is also prime is a real consequence rather than the reason — an edge length with no non-trivial factors cannot be folded into a coarser tiling, so no group of cells can be aliased into one. Useful, and honestly labelled: it is a property we get, not a property we chose for, and it has not been measured.
The principles
The name is a mouthful, so here is each term operationally — what it means for the running system, not as metaphor.
A cell of the structure contains a complete instance of the structure. Not a subtree — a whole lattice, with its own addresses, its own faces, its own edges.
Consequence: one implementation serves every depth. Retrieval, decay, scoring and validation are written once and work at any level, because there is only ever one kind of thing.
A cell is not a fragment awaiting context. It carries the full explanatory basis itself — a centroid and every face — so it is interpretable standing alone, at any scale, without reassembling its neighbourhood.
Consequence: a single recalled cell is a complete, auditable object. You can hand one row of the trace to a reviewer and it means something. That is what makes the audit export readable rather than a dump.
A position is not one value. It holds the set of states it could resolve to, weighted, simultaneously. Addressing it resolves one — and the alternatives are retained rather than discarded.
Where this stands: the retrieved-but-not-injected distinction is real and audited in the running engine. The fuller form — every declined alternative retained and queryable at every level — is architectural intent, not something the current code demonstrates end to end. It is on this page because it is the design, and it is flagged because it is not yet the implementation.
Capacity is not added by growing the structure outward with more nodes. It is added by descending into a cell, which opens a full level beneath it. The outer extent never changes.
Consequence: the address space is finite and enumerable at every level while being unbounded in total capacity. Locality is preserved permanently — neighbours stay neighbours no matter how much is stored, because nothing is ever inserted between them.
The one that matters most
Not every store has this problem — a content-addressed store like git never rebalances, and a hash written years ago still resolves. But the stores you would actually reach for here do: balanced trees split and move keys, vector indexes shift neighbourhoods as points are added, LSM stores rewrite on compaction. Insertion rearranges, so what is near a given item keeps changing and the physical structure that defined an address keeps being rebuilt underneath it.
Descending inward avoids that. A saturated region does not push its neighbours apart — it opens beneath itself. The parent address remains exactly as valid as it was; it simply now has an interior. An address written a year ago still resolves, and still means the same place.
This is not a new idea and we are not going to pretend it is. Octrees, quadtrees and tries all add capacity by subdividing a cell rather than extending the outer bound. What differs here is that they subdivide adaptively — cells split where the data happens to be dense, so the shape of the structure depends on its contents. This does not. Every level is the same 7³ with the same reserved portal, which is why one implementation of retrieval, decay, scoring and validation runs at any depth with no special-casing, and why any level can be enumerated exactly. The invariant geometry is the contribution, not the inward growth.
The honest trade: capacity is not free, it is relocated. Breadth is fixed and depth is open, so a read at depth d walks d levels, and locality is preserved within a level but not across one. Descent is implemented and unit-proved; it has not been stressed under real write pressure, so we quote no figure for what deep descent costs until we have measured one.
What follows from it
| Memory Cognitive Lattice | The address space used as memory. Cells hold recalled knowledge; descent handles growth; the flattening is the audit export. |
|---|---|
| F.R.A.C.C. | The address space used as decomposition. Cells hold candidate goal-elements; descent handles goals too large to act on; the flattening is the reasoning path. |
| Arc Lab | The address space used as explanation. The basis defines the faces; superposition holds competing readings; the flattening is the cosine matrix and the influence ranking. |
| Key System | The address space used as identity. Keys are addresses, kinds are faces, determinism is what makes an address reproduce its contents. |
| Cubex³ | Not an instantiation — the origin. The tensor board, the evaluator and the coordinate scheme came first, and the move from an 8-wide board to a 7-wide one to obtain a true centre is where the width of everything above comes from. Its own multi-commander doctrine layer is designed and not built. |
| SeCURED | The applied domain, and the oldest work here alongside Cubex³. A 228-page manual on smart-city resilience and emergency deployment, plus a simulator — terrain, road network, EMS allocation, utilities and a doctrine clause system — that runs and typechecks clean. The simulator is days old and has no test suite; the manual is unpublished. Neither is a deployed municipal system and we present them as neither. |
It would be tidier to say this is one architecture instantiated six times. It is not, and the true version is more interesting: the constants came out of the earliest work and kept surviving. A 7-wide cube was chosen to give a chess evaluator a centre. The same width later turned out to be the ceiling on distinguishable items in an explanatory basis, the shape of a decomposition, and the address space of a world generator — which now holds 63 of 63 object keys, complete, on disk.
That is convergence, not deduction, and it is worth less than a proof and more than a coincidence. It is also checkable: the constants are literals in code, asserted by tests, and countable in a directory tree. If the structure is wrong they fail together — and we would rather find that out on a game board than in a hospital.
“It is not spaced outward into branches. It is spaced infinitely inward.” William Wallace Morehead · Applied Intelligence Architect