Architecture

Architecture designed for enterprise authority, continuity, and proof.

The architecture separates control-domain authority, regional model capacity, site-local execution, endpoint participation, enterprise systems, and approved external services.

Baseline v1.0.0

One architecture across evidence, decisions, workflows and action.

Each mutable record has one accepting authority. Models and structured decisions provide proposals; Identity, Policy, EL-Flow and Action Fabric retain their respective admission and execution roles.

Deployment qualification binds the exact release, hardware, models, workload and enabled profile. Published architecture does not by itself establish production availability, performance or certification.

Request the Architecture Guide

Reference architecture

Central governance. Regional scale. Site-level continuity.

The Manager and control domain govern identity, policy, configuration, release evidence, fleet state, and global reporting. Site and regional services carry the low-latency workload after they receive valid policy, artifacts, identities, and capacity.

Control domain

Manager, operator experience, Identity, Policy, fleet, configuration, release evidence, Studio, entitlement, and global reporting.

Regional AI cell

Shared inference, model routing, retrieval hotsets, telemetry aggregation, artifact service, and regional continuity.

Branch / site edge

Local context, cache, retrieval, qualified models, bounded workflows, evidence outbox, network participation, and offline continuity.

Endpoint estate

Device identity, outbound secure channel, telemetry and posture, encrypted local buffering, and device-bound action leases.

Enterprise systems

ERP, CRM, ITSM, CMDB, identity, network, security, finance, documents, SaaS, and line-of-business systems remain sources of record.

Approved external services

Eligible models, SaaS APIs, publishers, and partner services are explicit routes with data, provider, budget, and evidence controls.

Your data

Where your data lives and how it is treated.

Three rules shape everything here. Your systems remain the sources of record. Every kind of record inside EL-Ai-NE has exactly one accepting owner. Withdrawing a right outranks any reuse.

Where it lives

DataWhere it livesHow it is treated
Your enterprise records: tickets, documents, configurations, device stateYour source systemsRead through admitted connectors under the source's own permissions. They remain the system of record.
Source snapshots and prepared evidenceThe Brain Trust store in the control or regional services; a hot set at the site where placement allowsStored with origin, rights and retention. Invalidated when the source changes or rights are withdrawn.
Search, vector and graph indexesSite and regional hot sets; control servicesCan always be rebuilt. The same access rules apply as to the source.
Accepted operational state: entities, assertions, contradictions, casesThe operational record in the control services; a readable cache at the siteEarlier and conflicting assertions are preserved, never overwritten by a dashboard.
Workflow progress, waits and human tasksThe workflow journalSurvives restarts and client disconnects.
Action requests, approvals, executions and verificationsThe action journalImmutable receipts. An unknown outcome is reconciled, not retried.
Policy and its versionsPolicy services; signed, expiring copies cached at sitesA site can decide locally only inside the lease of its signed copy.
Governance evidence, receipts and tombstonesThe governance ledgerSigned and append-only. Tombstones outrank every cache, index and backup.
Prompts and responsesThe memory of the model runtime, for the life of the requestNot kept by default. Kept for 0 to 30 days only when justified.
Answer and context cachesSite and regional nodesScoped to permissions and policy version. A correction or revocation invalidates dependent entries.
Models and their supporting filesThe artifact store; a cache at the siteFixed by digest and signed. Live business facts are never baked into a model.
Usage, cost and outcome factsCost and outcome storesExact and append-only. A closed period is corrected only by a restatement.
Metrics, traces and logsA local agent at the site; a telemetry store in the region or control servicesIdentifiers, sizes and result codes only. No prompts, documents or secrets.
Secrets and keysYour approved key management service or HSMIssued as a narrowly scoped, short-lived lease at the moment of use.
BackupsA restricted, encrypted location with separate key custodyChecked before restore. Current tombstones are applied before restored data can be read.

What happens to it, stage by stage

  1. STAGE 1AdmitA connector or a request brings data in, with its origin and rights. External content is treated as untrusted evidence.
  2. STAGE 2Classify and bind rightsA class, purpose, locality and retention are attached before any use.
  3. STAGE 3Store with one ownerThe accepting owner records it. Everything else receives a reference.
  4. STAGE 4Use the minimumA context is compiled for one task from what this caller may see now. Rights are rechecked at use.
  5. STAGE 5Derive with lineageCaches, indexes, summaries and views stay linked to their sources and inherit the strictest rights.
  6. STAGE 6Retain or holdRecords are kept by purpose and type. A legal hold blocks deletion but allows no new use.
  7. STAGE 7Revoke and tombstoneA signed tombstone stops serving everywhere first. Cleanup follows.
  8. STAGE 8Purge and proveManaged copies are erased, backups expire, and deletion evidence is kept without the content.

Corrections, revocations and tombstones travel to every derivative: caches, indexes, generated pages, decisions and restored backups.

Rights are separate permissions

Being able to read something is not permission to do anything else with it. Capture, transient processing, transcription, translation, regional routing, external routing, storage, disclosure, memory, evaluation, training, synthetic derivation and cross-tenant contribution are each granted and checked on their own. No model is trained because a document was readable for answering a question.

How long it is kept

RecordPlanning starting point
Raw prompts and responsesNot kept by default; 0 to 30 days only when justified.
Detailed site telemetry7 to 30 days, kept locally.
Aggregated operational telemetry90 to 365 days, kept centrally.
Operational record assertions and cases1 to 3 years, or your own lifecycle.
Action, approval and governance evidence1 to 7 years, according to consequence and regulation.
Model, evaluation and deployment evidenceThe life of the artifact plus the rollback and audit period.
Cost and invoice evidenceYour accounting retention policy.
Operational backupsA short rolling window locally; typically 30 to 90 days centrally, plus an approved archive.

Configurable starting points from the baseline, not compliance requirements. Your retention policy is an owner decision; until it is set, the affected capability stays disabled.

Deletion you can verify

A deletion is reported as one of seven honest states instead of a single "done": accepted by the owner, propagation in progress, local copy invalidated, physical deletion complete, backup expiry pending, blocked by a legal hold, or failed. A restore can never bring deleted or revoked content back, because current tombstones are applied before restored data can be read.

Status at Baseline v1.0.0. The governance ledger, audit record, tombstones and usage metering are Available now. Database-enforced tenant separation, data classification, retention bundles and signed backups are Included, enabled per qualification. On current installs, chat sessions are also saved to disk, and the knowledge store does not yet accept your own documents; plan a proof of concept accordingly.

Authority chain

A request never moves directly from natural language to a consequential tool.

Every material request moves through a structured chain that prevents a model, connector, UI, or endpoint client from becoming an accidental control plane.

01IdentityActor, workload, tenant, site, delegation, and session.
02PolicyPurpose, data rights, route, consequence, approvals, and budget.
03ContextOperational record, source assertions, permissions, freshness, and memory.
04WorkflowEL-Flow journal, tasks, waits, retries, and compensation.
05ActionTyped target, parameters, single execution, and credential boundary.
06VerificationIndependent observation of the actual resulting state.
07CommitOperational truth, evidence, exact cost, and outcome facts.
08ImproveDrift and optimization candidates routed through qualification.

Governance authority integration

Governance strengthens the same chain. It does not create a second path around it.

Governance records inform what must be enforced. Identity, Policy, EL-Flow, Action Fabric, the operational record, Studio, release integrity, Privacy, Cost, Outcome, and Observability continue to own and enforce their existing technical state.

InventoryAI inventory
RegulatoryApplicability
ImpactAssessment
OversightHuman control
Third partiesSupplier AI
TransparencyNotices
Post-marketMonitoring
Governance obligations and evidence
IdentityWho and what acts
PolicyWhat is permitted
EL-FlowDurable process state
Action FabricConsequential effects
Operational recordAccepted operational truth
Studio and releaseQualification and release
No governance dashboard, classifier, framework mapping, or assessment can directly grant runtime access, execute an action, promote a model, or qualify a release.

Design principles

Capability can degrade. Authority cannot widen.

During disconnection, source gaps, unqualified models, stale policy, or capacity pressure, EL-Ai-NE narrows, delays, refuses, or marks the result unknown. It never silently expands permissions to keep an automation moving.

Authority before autonomy

Models and agents remain proposal components.

No data is not green

Missing or stale telemetry remains unknown or degraded.

Verification before success

Execution acknowledgement does not prove the target state.

Evidence before claims

Production and outcome claims bind to exact artifacts, profiles, tests, and operating windows.

Structured Decision Intelligence Planned

Narrow judgments with explicit evidence and uncertainty.

Registered questions

An immutable decision definition sets the question, candidate meanings, required evidence and permitted consequence.

Qualified evaluation

Use deterministic rules when sufficient, then eligible local or regional models. Review or abstain when evidence or calibration is insufficient.

Controlled consumption

EL-Flow checks current evidence, rights, qualification and policy before using a result. A decision cannot authorize an action.

Structured decision cascade with deterministic rules, qualified local models, review and validated signed results
Accepted results enter validation; unresolved judgments escalate only within the admitted scope and deadline.

Governed Web Experiences Planned

Freshness and access travel with every generated view.

A page composer assembles approved components from versioned evidence. A freshness coordinator handles freshness and invalidation. The interface serves the accepted view only after current audience, source-right and dependency checks. A failed refresh never renews freshness, and restored content must pass current revocation and tombstone checks.

Persistent Agent Workspace retains workspace preferences and routine definitions while EL-Flow owns durable execution. Each trigger and collaboration handoff rechecks authority.

Ask EL-Ai-NEPublic website guide
Ask about capabilities, governance, privacy, Brain Trust, risk, deployment, pricing, comparisons, or public resources.