Open Territory
The map shows what's unbuilt.
An empty cell with occupied neighbors is not a gap. It is a prediction. The Stack does for synthesizers what Mendeleev's table did for elements — it tells you something exists, and what shape it has, before anyone has built it. This is the field's open territory. Someone should build these.
Part 01 · The Mendeleev move
Predict the cell from the gap.
In 1869 Mendeleev left holes in the periodic table and described the elements that would fill them — their weight, their density, their chemistry — years before anyone isolated gallium or germanium. The prediction worked because the table had structure: a missing element is constrained on every side by its neighbors.
The Stack's tensor does the same thing. A cell that is empty but whose neighbors are occupied is not a gap — it is a synthesizer that should exist, whose coordinate fixes its cost, its verifiability, its interface, and its encoding regime before anyone has assembled it. The teal cells on the table are predictions. The structure tells you their shape.
This page reads four of those predictions out loud: the cell, the coordinate move that reaches it, the occupied neighbors that fix its shape, and what it would take to be its first inhabitant. None of these require a research breakthrough. Each one combines two disciplines that already exist, that nobody has bolted together because they live in different rooms.
Part 02 · The transformation rule
A move, productized, becomes a cell's first inhabitant.
Every coordinate move — the moves named on the thesis page — opens a region of the Stack. Where that region contains a predicted-but-empty cell, the move, done systematically and sold as a product, is that cell's first inhabitant.
A move is what you do to one synthesizer. A cell is what you build when you do that move systematically — and sell the systematization.
RAG wasn't a new idea — retrieval is ancient. RAG was retrieval aimed at the navigation-encoding cell and productized. Caching isn't new either; the versioned-substrate cell is caching-with-invalidation productized. The pattern is the deflation that "AI is software" makes, applied to AI progress: every advance the field celebrates is a known move, aimed at an empty cell, industrialized. The Stack is what makes that legible.
Not every move points at empty territory. Agentic chaining stays in the already-settled reactive-substance cell — a product, not a land-grab. The land-grabs are the moves that point at empty cells. The Stack is what tells you which is which. That's its job.
Part 03 · The catalog
Four cells. Each buildable now.
Auditable physical AI ⟨ Z₂, active, L₁, {control, verify}, bodies, citation, world model ⟩
Today's physical AI is Z₁↔₂ active with delayed verifiability — deterministic control envelopes (Lyapunov controllers, lane-keeping) where you learn it was wrong when it crashes. The empty cell is physical AI whose reasoning is citation-anchored before it acts: "turning left because ISO 10218 §5.3 requires X given sensor state Y; here is the trace."
The move
V: delayed → full. Runtime assurance + decision-time citation.
Occupied neighbors
Drone w/ Lyapunov controller (active, delayed). Legal RAG (reactive, V₂). Both exist.
First inhabitant
Combine Simplex runtime assurance with citation anchoring. Neither is new; nobody has fused them.
Live-versioned regulatory LUT ⟨ Z₁↔Z₂, persistent, L₀, {lookup, temporal}, tokens, full, facts ⟩
A temporally-indexed database of what the rules are, at every point in time — queryable as "what did §1031 say on 2023-04-15." Every Z₂ legal, tax, and compliance synthesizer should rest on this instead of regenerating "the tax code says…" from weights (L₂, unverifiable, possibly hallucinated, definitely stale). This is the cache-with-invalidation, productized.
The move
R: facts-in-weights → navigation-over-versioned-substrate.
Occupied neighbors
Curated case-law database (persistent, V₂). Tax engine + UI (Z₁↔Z₂). Both exist.
First inhabitant
Boring infrastructure, not a paper. Which is exactly why it's uncontested — and why whoever builds it owns the substrate every regulated-AI product licenses.
Grounded generation ⟨ Z₂↔Z₃, reactive, L₂, {gen, verify}, tokens, citation, navigation ⟩
Today's Z₃ generation runs until a token budget expires — it doesn't know why it stopped. The empty cell is generation that carries its own stopping condition: generate until [a human accepts | a Z₁ verifier passes | a consequence loop closes | an aesthetic threshold is met]. Constitutional methods are a baby step. The full cell is generation budgeted against its grounding — which is also the lowest-cost Z₃, because it stops the instant the grounding is satisfied.
The move
V: none → citation. Budget the generation against a named verifier.
Occupied neighbors
LLM + tool use (reactive, V₂). Constraint-bound search + verifier. Both exist.
First inhabitant
Couple a generator to a verifier with a budget. Lower-cost and more controllable than budget-exhaustion generation — same move.
Constraint-bound at scale ⟨ Z₁, reactive, L₁, {search, verify}, tokens, full, grammar ⟩
The constraint-bound shape — energy-based search with a verifier — exists at demo scale (Sudoku, individual proofs). The empty cell is constraint-bound synthesis as a first-class production mechanism: dependency resolution, scheduling, type inference, formal verification, EDA, all routed to a hardened verifier substrate with the same operational maturity as an LLM endpoint — receipts, SLAs, caching, audit.
The move
Productize search-with-verifier. Operational parity with the LLM endpoint.
Occupied neighbors
SAT solver (reactive, V₁). Theorem prover (grammar). Energy-based reasoning models. All exist.
First inhabitant
Wrap a verifier in the operational envelope the field built for LLMs. The reasoning is solved; the productization isn't.
Part 04 · Why they stay empty
The field competes on capability. These cells reward verifiability.
Every cell in the catalog is V₁ or V₂. None is reachable from a V₄ baseline without rebuilding the mechanism — and rebuilding the mechanism means abandoning the architecture the incumbents are optimizing. The scoreboard the field plays on is benchmark capability. There is no benchmark for "auditable, lower-cost, deployable in a regulated market," so nobody optimizes for it, even though that's the valuable thing.
That's not a knowledge failure. Everyone knows brute-force AI is wasteful and unverifiable. It's a coordination failure: the incentives all point at bigger models, and the empty cells require a different mechanism, not a bigger one. The cells stay empty because filling them is uninteresting to the scoreboard and uncontested in the market — the same fact, read two ways.
Whoever fills one owns a market the V-blind incumbents structurally cannot enter. Not because the incumbent lacks capability — because entering means rebuilding from a verifiable substrate, and the incumbent's whole advantage is the substrate they'd have to abandon. The empty cell is the moat capability can't dig around.
Part 05 · How to read this
A research agenda, not a roadmap.
This is not anyone's product plan. It is the field's open territory, made legible by the Stack. The cells are described by their coordinates, not by who intends to build them. Anyone with the two adjacent disciplines and the will to bolt them together can be the first inhabitant.
The catalog is not exhaustive. The tensor has more teal cells than four, and every projection of the seven axes surfaces different predictions. These four are the ones whose neighbors are closest and whose first-inhabitant path is shortest — the ones where the structure is most confident.
Mendeleev didn't mine gallium. He predicted it, and named what it would be. The Stack predicts these. Someone should build them.
The map terminates here, at the edge of what's been built. Everything past this line is open. The thesis named the structure; the cost function named the method; the receipt made the claims falsifiable; the frontier names what's unbuilt — and tells you, before you start, what shape it has.
Continue
The edge of the built.
See the cells on the tensor. Read the axis that gates them. Watch the loop that fills them.