Skip to content

Architecture

This is the curated map of the architecture/ tree (37 pages). It groups every page by what it's actually about — the commit model, the analytics/reasoning plane, distribution and scaling, performance, media/voice/visualization, the robotics/GPU tail, and security/access hardening — so you can go straight to the layer you need instead of scanning a flat file list. Pages marked are the release-defining contract pages tracked directly in site navigation and by the documentation-contract gate (scripts/check_documentation_contract.py); everything else here is just as real, only demoted out of the sidebar.

For the engine's single guiding design principle, see North Star: Seamless — every cross-modal read/write path (a write that touches OWL, vector, graph, and time-series together) is implemented at every wire surface (RPC, every SQL wire, SPARQL, GraphQL, …), never merely flagged at the one it was first built for.

Core commit & governance model

The transactional backbone: one durable commit point, one structural property schema, one served modality identity, and the substrate they all sit on.

Page What it covers
The master-of-all engine The deep architectural reference: durability, cross-modal ACID, RDF/OWL mapping, the RLS request path, streaming/CDC, federation, multi-Raft + cross-shard 2PC, tenant lifecycle.
Verified Request Authority The eg2. request envelope every call carries — request id, graph, method, digest, timestamp, nonce, idempotency key, ACL agent, tenant, audience, policy version, roles, scopes, delegation chain, trace context.
Authoritative MutationBatch The engine's durable mutation currency — commit-before-ack staging, durable idempotency/status, ordered projection outbox, the native WorkItem state machine.
Governed ChangeEnvelope One native transaction for external graph/object material, policy, lineage, evidence, typed content versions/cursors, and the outbox.
Canonical structural properties Why structural meaning uses one key per concept — property blobs are open MessagePack, but structural fields aren't silently promoted through aliases.
Governed modality serving Universal Artifact/Occurrence/Rendition/Segment/Feature/EvidenceLocus identities, the ServedModality ingest/query/lifecycle service, and the one cursor-driven KnowledgeStream protocol shared by every modality.
Subsystems (C4 containers) How the broker, observability stack, KV-cache, and every wire adapter compose on the one GraphCore + redb-authoritative store + unified RowSet planner.
One build, opt-in layers The feature-composition map: the one main full-featured build plus the opt-in cluster (HA Raft) and full-extras (GPU/ROS2) layers.
Epistemic OS Hardening — capability catalog The code-verified, line-anchored catalog of everything the hardening program (Phase 0 → Phase 3 + "exceed" tracks) shipped.

Analytics, reasoning & optimization

The numeric/analytics kernel, the incremental-reasoning plane, and self-improving program optimization.

Page What it covers
Analytics Program — one kernel, two surfaces The BLAS/LAPACK-free Rust numeric kernel (eg-numeric) that serves both Python-side array math and in-database analytics over engine-resident data.
Numeric kernel The kernel foundation of the Analytics Program — compiled kernel, Agent Utilities numeric surface, and native engine operators.
Distributed analytics & incremental reasoning Verified coordinator RPCs for leased/fenced remote workers, durable typed results, and outbox/cursor-driven TMS, conflict, causal, and materialization maintenance.
Native program optimization The submit_program_optimization contract, 13 Rust-native optimizer families, evidence across all 14 modalities, governed runtime plan steps, evaluation-gated promotion.
Lakehouse LTAP interop The LTAP (Lakehouse-Transactional-Analytical Processing) superset: external engines read the store as open Parquet + Delta/Iceberg with zero ETL.
Data mining graph_mine / /api/mining — pattern and structure mining over engine-resident graphs.
Graph learning graph_learn / /api/graphlearn — learned representations over the same substrate.
Analytics in UQL (EG-353) Analytics operators reachable directly from the Unified Query Language.

Distribution & scaling

Raft clustering, resharding, admission control, and the correctness harness that gates every distributed/durability claim.

Page What it covers
Engine Scaling Program (M1/M2/M3) The claim-status vocabulary (DESIGNED/IMPLEMENTED/UNIT-PROVEN/LAB-PROVEN/LIVE) and the milestone map for the whole scaling program.
Scale claim register The machine-readable evidence-state register backing the M1/M2/M3 docs.
Multi-Raft cluster status (M2) Current-main status and handoff for M2 Raft hardening.
Catalog-driven resharding (M3) Current-main status handoff for catalog-driven resharding.
M3 cross-node elasticity planner The policy-only planning layer for cross-node graph/shard elasticity — proposals only; placement/reshard stays the sole mutator.
Cluster deployment & migration Running the engine as an HA Raft cluster across runtime-selected nodes and converting the current authoritative node into the cluster seed without data loss.
Bounded shard/Raft drain contract The side-effect-free state machine a membership owner uses before removing a node.
Per-graph write coalescer Turning N concurrent single-op writes to one graph into one topology-lock acquisition per batch.
Reserved read-admission lane The admission lane reserved for reads/queries so an interactive request is never shed behind an ingestion write firehose.
Unified IndexManager One registry/seam over the engine's secondary indexes, so the planner consults one place instead of each index individually.
Correctness + load harness The standing proof-engine gating every distributed/durability claim: multi-Raft, M2 redb durability, distributed transactions, PITR/replication.
Cgroup-aware capacity resolution The single resource-capacity seam deriving queue/concurrency/runtime/memory/ASR limits from host + cgroup observation.

Performance & complexity

Page What it covers
Hot-path complexity ledger Algorithmic bounds (not benchmark claims) for the hottest request paths.
D-OP-1: RLS projection cache Caching project_core()'s RLS projection per (actor, graph version).
Push-gate execution evidence ci_gate_replica.py — the single producer for workflow-derived heavy checks in the pre-push gate.

Media, voice & visualization

Page What it covers
Native ASR provider (GOC-33) Native Rust speech-to-text, reachable over the engine's own wire protocol as Method::Asr.
Native Piper-ONNX TTS (GOC-34) The native text-to-speech provider behind the tts.* wire contract.
Native visualization (D-VZ-1) The LOD-native visualization stack (declarative chart IR, columnar store, decimation/density kernels) in place of a matplotlib-style library.
Remote KV-cache HTTP backend The kvcache-server feature exposing the shared, content-addressed KV-cache backend so parallel vLLM/LMCache instances share blocks by token-hash.

Robotics, GPU & the distribution tail

Page What it covers
Distribution, Robotics & GPU tail (EG-3.x) The last features of Program B — all feature-gated into the opt-in full-extras layer; the default build links none of the heavy deps.
The unified in-process engine (PyO3) — design Embedding the Rust engine in-process via PyO3 for the single-binary deployment shape.

Security, access & hardening

Page What it covers
RBAC action unification — Phase 1 Blast-radius map for unifying two RBAC action vocabularies. Phase 1 verdict: not yet safe to fully unify; no behavior changed by the document itself.
Native capacity & WorkItem admission authority The additive native protocol for the GOC-21 capacity lease and GOC-19 WorkItem submit surfaces; the engine's redb tables are authoritative, schedulers are projections.

Design principle: design for a network boundary

Every out-of-process invocation crosses a process boundary — serialize, socket round-trip, deserialize. A call is not a cheap function call. Batch, never per-element: ship work into a single round-trip over data already resident in the graph (one all-pairs op, not a Python loop), and keep tight per-element math in-process. The Rust Compute Guide explains how this shapes every caller.