The four operations, in one sentence each.
- BTUT — reduce. Turn an O(N³) PDE coordination problem into an O(N) kernel-weighted contraction.
- Crystara — crystallize. Turn a fixed JEPA predictor into a runtime-grown family of H₀/H₁/H₂ modules, typed by persistent homology.
- NIV — signal. Turn dozens of macro variables into a single, transparent, economically interpretable, walk-forward validated scalar.
- PDE — ingest. Turn arbitrary human submissions into an auditable, vector+FTS-retrievable, approval-logged knowledge substrate.
The composition.
Picture the Latent Ocean as a wide, shallow plane of operations. Every primitive is a region of the plane; every edge is a piece of data flowing between them.
- BTUT lowers the floor. Coordination is no longer a compute bottleneck. One million agents, one node, under ten seconds. This is the assumption the rest of the stack depends on.
- Crystara spends that surplus. Because coordination is linear, inference-time compute can be used to explore blank spaces in the predictor’s energy landscape, run persistent homology on the trajectories, and grow typed modules out of the stable features. The predictor grows into the shape of the data.
- NIV is what crystallized structure looks like when projected to a scalar. The same logic — known primitives, transparent weights, orthogonal information — is the recipe for every external signal the Ocean will emit. NIV is the first instance; monetary, energy-grid, and supply-chain signals follow the same template.
- PDE keeps the substrate alive. Submissions flow in through Submit → Moderate → Thin → Crystallize. The approval log makes the provenance chain queryable. The knowledge base is not a snapshot; it is a living document.
Data flow.
[ submissions ] ─► PDE ─► chunked, embedded,
approval-logged corpus
│
▼
[ BTUT-coordinated workers crystallize modules ]
│
▼
Crystara: H₀ / H₁ / H₂ predictors
│
▼
projections → scalar signals (NIV class)
│
▼
external systems
(policy, macro, grid, …)Why this is “horizontal.”
The dominant scaling story of the last five years is vertical. Bigger transformers, more tokens, more chips. That story is real, and it works, and it is also overwhelmingly capital-constrained — the only people who can play are the people who already have the cluster.
Horizontal intelligence is the other direction. A primitive earns its place by being composable, auditable, and reducible. A stack of four such primitives — coordination, structure, signal, ingestion — reaches the surface area of a platform without ever pretending to reach the scale of a frontier lab. The compute bill is small because the primitives do the work.
Inference-time substrate.
Latent Ocean is not a model. It is an inference-time substrate: coordination is on demand, module crystallization is on demand, signals are emitted on demand, the knowledge base is fed on demand. The heavy training runs live inside Crystara and inside the NIV walk-forward framework, but every external interface is a fast, narrow, well-typed contract.
The interesting consequence is that the substrate does not belong to any one model class. It is language-model-agnostic, it is planner-agnostic, it is observer-agnostic. The four primitives are the interface.