Technology
The foundation between your data and your model
Axiom Vertex is a Deterministic Spatial Compute Architecture (DSCA). It provides the compute fabric and the stack for deterministic, asynchronous reasoning by any AI model, token-based or non-token. Every record is validated, and the context an AI model reads is drift-free, bit-for-bit reproducible and re-derivable at whatever scale of computation the model requires.
Why it matters
What Axiom Vertex helps you solve
Context cost
Agents re-read the same context on every call. Context windows grow, memory fills, and context is compacted, stored and offloaded again and again. Axiom Vertex supplies each request with only the validated records it needs, which lowers operating cost and the hardware a team has to buy.
Drift
Small errors compound across long agent runs. The AI model reasons over drift-free context on every run, so the same question returns the same context, bit for bit.
The black box
The interior of a model remains hard to inspect. Axiom Vertex governs the evidence side, recording what the model saw, where it came from and that it is unchanged on the Hypergeometric Ledger. This is the record that EU AI Act Article 12 logging asks for.
Zero trust and sovereignty
Each Virtual Organization is an isolated tenant on its own dedicated compute fabric. The data stays in the user’s own storage, and every record carries a validation proof.
Principles
Data stays with the user
Records remain in your own storage. Axiom Vertex holds coordinates, validation proofs, ledgers and indexes.
Drift-free context
Every validated record has fixed coordinates that never drift, so the same question returns the same context, bit for bit, on every run.
Far less data moved
Hundreds to thousands of times less data moved per operation, depending on model size and hardware infrastructure.
One fabric per user
Each user gets an isolated tenant, a Virtual Organization, with its own hyperdimensional compute fabric.
Any AI stack
Token-based and non-token models, on CPUs, GPUs and deterministic hardware.
Where it sits
Between the caller and the model
The compute fabric stands between the client and the AI model, and the data remains with the user.
Architecture
Seven layers, one deterministic core
Select a layer to see what it does. Every layer answers to the Hub at the bottom.
SDK, API and observability
One surface for every caller
- REST, gRPC and WebSocket
- SDKs in six languages
- Metrics and tracing
Cross-cutting services
Shared by every layer above
- Geodesic Context Engineassembles context for an AI model from the validated records of a Virtual Organization
- Expresswaya physically separate high-throughput network fabric for accelerators
- AI model gatewayvalidates each request before forwarding it and attests each response
- MCP gateway
- Control plane
- Authentication
- WASM runtime
- Event bus
Execution and management
Runs each user's Virtual Organization
- Conductor
- Relational Reasoning Model
- A gateway for each Virtual Organization
- Client services
Spokes
Services on the compute fabric
- Axiom, the core validation and proof pipeline for federated, AI-generated and device-generated data
- Atlas, spatial intelligence for AI models, with deterministic, reproducible context retrieval without drift
- Cortex, lifecycle management for autonomous and IoT devices, including validation of device-generated data for AI reasoning
- QI Encoder, mathematically encodes and seals validation proofs, with seal receipts that verify independently and a zero-knowledge proof bridge for auditors and regulators
- VX Convergence, multi-domain consensus reasoning across all projects and services
M-Verse infrastructure
Allocates and tracks the fabric
- A dedicated hyperdimensional compute fabric for each Virtual Organization
- Isolated tenants, on premises or air-gapped
Compute fabric
M-Verse, validated coordinates and the record of validation
- HGL, the Hypergeometric Ledger
- HyperVector Ledger
- Hyperdimensional compute
Hub
The Hub's root system laws enforce determinism at every layer
- Root system laws: validation enforcement
- Root system laws: dimensional enforcement
Reasoning on the fabric
How an AI model reasons on the compute fabric
Everything before the model is deterministic, and the AI model reasons over the validated context.
Each record enters through an ingest gateway built for its kind of data.
A validation proof, an unforgeable, immutable mathematical identity committed to the Hypergeometric Ledger
Recorded on the Hypergeometric Ledger, with the record kept in your storage.
- 01Readiness checkPlatform
A quarantined Virtual Organization is refused, as is a model that is quarantined, suspended or not yet attested
- 02ChunkPlatform
Long text splits into slices, each within the platform's embedding limit
- 03Validate and commitPlatform
Each slice is validated; a slice that passes is committed with its validation proof, and a slice that fails is refused
- 04Resolve wordsPlatform
Each word receives its coordinate from the vocabulary; a word the vocabulary lacks is resolved from its parts and its nearest neighbours, and a word still unresolved is validated and added to the Virtual Organization's vocabulary
- 05ReducePlatform
Slice coordinates merge into one question coordinate
- 06Reasoning bridgesPlatform
The reasoning bridge that serves the question resolves its coordinate across its runners, and a question across the organization is resolved by all six; the Geodesic Context Engine assembles the validated context
- 07The model reasonsCustomer
Attention, hidden states, logits and decoding stay with the customer
- 08Output checked and attestedPlatform
The reply is checked against the model's baseline; the platform then validates it and commits it to the model-output attestation ledger when the Virtual Organization's policy selects it
RRM Harness
Six asynchronous reasoning bridges
The Relational Reasoning Model (RRM) Harness integrates and orchestrates your AI model, with guardrails and support for agentic workflow. One question coordinate, for text of any length, is resolved by the reasoning bridge of the sector that serves it, and a query across the organization is recomposed from the bridges that contributed. The six reasoning bridges share one runner budget.
Each reasoning bridge coordinates the resolution of queries for one sector of a Virtual Organization. It decomposes a query into sub-tasks, routes them to its runners, and recomposes their outputs into one validated coordinate whose proof is committed to the Hypergeometric Ledger.
Asynchronous
Each query proceeds independently of the other bridges.
Deterministic
Recomposition does not depend on the order in which bridges finish.
Scalable
Runner capacity follows the hyperdimensional tier and the memory available to the tenant.
Breadth, depth, efficiency
Wide, deep, and light to move
131,072 dimensions
Six hyperdimensional tiers, from 4,096 to 131,072 dimensions. A registered AI model is bound to a tier, and the width of its embeddings is reconciled with the tier's vector size.
834 depth layers
Hierarchical depth for context ordering. Records are ordered by lineage depth and sibling order, uniform across every tier.
Hundreds to thousands of times less data moved per operation
Depends on model size and hardware infrastructure. Conventional retrieval compares a full embedding vector per record, 4 bytes per dimension; Axiom Vertex compares one 48-byte coordinate. Over 834 depth layers that is up to 437 MB against 40 KB.
Model bootstrap
Your model joins in six steps
The platform steps run separately for each tenant. Bootstrap runs once for each AI model, and every later question reasons over drift-free context.
- 01The model stays putCustomer edge
The AI model stays where the customer runs it, with its weights, attention and decoding
- 02Register and attestPlatform
The AI Model Integration Service records the model, holds its credential and checks its identity on a schedule
- 03BootstrapEdge and platform
A vocabulary is built once for the model, and an adapter checkpoint is trained at the customer's edge and sent to the platform
- 04Reasoning mapModel integration
Our platform model, Qwen, writes the integration map once, at registration, when a model's output shape is unfamiliar; the map is validated and cached
- 05RRM HarnessPlatform
Integrates and orchestrates the AI model and its bootstrap jobs, with bindings that act as guardrails and support for agentic workflow
- 06ReasoningPlatform
The reasoning bridge that serves the question resolves its coordinate across its runners, and a question across the organization is resolved by all six; the Geodesic Context Engine assembles the validated context for the model
Vocabulary
Your registered AI model embeds the training text, and a sealed projection maps each embedding into platform coordinates that any party can re-derive. Every token holds one coordinate. A word found during a question that the vocabulary lacks is resolved from its parts and its nearest neighbours, and a word still unresolved is validated and added to the Virtual Organization's vocabulary.
Adapter checkpoint
You train a low-rank adapter on your own base model, at your own edge, with LoRA, QLoRA, TinyLoRA or a related parameter-efficient method. Only the adapter checkpoint crosses, once and in one direction, in the Hugging Face PEFT format. Each row receives a coordinate, is validated, and is committed with an attestation record.
One signed bundle
A trained model leaves with one signed attestation bundle that binds its parameters, its training data and its training program, so any party can verify them independently. The vocabulary release is signed separately, and once the model is in service its identity, output entropy and fingerprint are checked.
Details
Parameters are attested by a hash of the weights file, signed with ML-DSA-65; training data by a Merkle root over the validated documents; the training program is recorded for replay.
Assurance
Verification and sovereignty around the model
Your data stays sovereign
Weights, optimizer state, gradients and training compute remain on your infrastructure, and inference remains where you run the model. The trained adapter checkpoint is the only model artifact that crosses, one way. Training data is chosen by you and validated on ingest.
Verification outside the model
An identity probe sets a baseline for each registered model; repeated failures place it in quarantine. Replies pass the output contract and checks before they are committed, and an entropy baseline can withhold a reply that departs from the model's own history.
Proofs committed to the ledger
Each validation proof is an unforgeable, immutable mathematical identity, committed to the Hypergeometric Ledger and sealed by its Merkle leaf. Its coordinates never drift, so the same coordinates always return the same content.
Any model class
Token-based decoder LLMs and SLMs, and non-token graph, convolutional and diffusion models, which enter as vectors or raw bytes. Identity is checked for every class, through the fingerprint canary for non-token models, and the same accountability controls, the same six reasoning bridges and the same runners serve every class.