Identity and Policy
Resolve the principal, workload, tenant, site, purpose, delegation, data rights, route eligibility, consequence, approval, and budget.
Platform
EL-Ai-NE coordinates existing systems rather than replacing them. It provides the authority, context, orchestration, execution, verification, and evidence layer that makes AI operationally accountable.
How it works
Models and agents reason and propose. A small set of named authorities accepts facts, grants permission, runs workflows, performs actions and records evidence. Nothing can grant itself a permission, accept its own fact or approve its own action.
| Building block | What it does | Where it runs |
|---|---|---|
| Manager | The control plane: node inventory, configuration, policy distribution, dashboards and administration. It stays out of the request path. | Control services |
| Site node | Enforces policy, serves the cache, mediates AI requests, holds local evidence and queues work for the region. | Each site |
| AI Gateway | The only way to reach a model. Every AI request passes the same eight stages. | Site and region |
| Operational record | Accepted cross-system operational truth, with contradictions and unknowns kept visible. Your own systems remain the sources of record. | Control services |
| Company Brain Trust | Governed knowledge and memory: stages, verifies, retrieves and tombstones evidence. | Site, region and control |
| EL-Flow | The durable workflow engine. Work survives restarts, waits and disconnects. | Site, region or control |
| Action Fabric | The only admitted path for a change to another system: request, approval, execution and verification. | Enforced locally |
| Governance ledger | Signed, append-only evidence and tombstones. | Control services |
| Studio | Evaluation, promotion, canary and suspension of models and other artifacts. | Control services |
| Cost and Outcome | Exact usage and cost facts, and accepted outcome measures. | Control services |
A request for an answer passes eight fixed AI Gateway stages. A denial, a missing authority or an unknown state stops the request at that stage. Nothing fails open.
An action is anything that changes something outside EL-Ai-NE: opening a ticket, refreshing a device, changing a setting. Every action follows the same path, whoever or whatever proposed it.
| Consequence class | Meaning | Default handling |
|---|---|---|
| C0 | No material external effect. | Proceeds within policy. |
| C1 | Low: easily reversible, low value, narrow scope. | Follows the approved action lane. |
| C2 | Moderate: a meaningful effect with a bounded blast radius. | Explicit human approval by default. |
| C3 | High: substantial financial, security, customer, compliance or availability impact. | Explicit human approval by default, with strong authentication and separation of duty. |
| C4 | Critical: safety-critical, broadly irreversible, legally controlled or systemic. | Explicit human approval by default. |
The default ceiling for a deployment is C0. Raising it is an owner decision. The scale is an operating control and does not assign a legal risk category.
Platform authorities
Components may propose, cache, project, or display state. They cannot create a competing private truth or bypass the authority chain.
Resolve the principal, workload, tenant, site, purpose, delegation, data rights, route eligibility, consequence, approval, and budget.
Compile permission-filtered retrieval, graph context, source assertions, approved memory, language, and workflow evidence.
Keep durable workflow progress, waits, retries, deadlines, approvals, compensation, and resumable multi-agent tasks outside model context.
Admit typed mutations, issue short-lived credentials or endpoint leases, enforce single execution, verify the resulting state, and compensate when qualified.
Preserve accepted operational assertions, contradictions, source lineage, governance receipts, and correction history.
Select qualified local, regional, or approved external models by quality, privacy, latency, capacity, reliability, and economics.
Join site, path, application, device, telemetry, posture, and source completeness with the business workflow.
Keep exact cost and accepted business outcomes as separate facts, then publish reports with freshness, unknowns, lineage, and restatement state.
Agent and workflow operations
An EL-Ai-NE agent can interpret evidence, create a plan, call permitted models or tools, and propose an action. The durable task, permissions, side effects, and success state remain owned by the platform authorities.
Governance and assurance plane
Eleven bounded governance services connect inventory, impact, oversight, vendor governance, transparency, monitoring, redress, frameworks, audit, and resource impact to the existing technical authority chain.
The inventory and applicability services establish what exists, who owns it, where it operates, which roles and obligations may apply, and when reassessment is required.
The impact and oversight services record affected populations, mitigations, residual risk, intervention design, AI literacy, competence, and oversight effectiveness.
The transparency and post-market services govern system cards, instructions, notices, monitoring, incidents, complaints, challenge, remedy, appeal, and corrections.
The framework, audit and resource-impact services maintain crosswalks, audit, corrective action, management review, and material resource-impact evidence.
Model strategy
EL-Ai-NE uses deterministic logic and typed APIs first, permission-aware retrieval for current company knowledge, qualified models for reasoning, and model adaptation only when a stable behavioral gap justifies it.
Rules, schemas, validation, calculations, and APIs remain deterministic.
Current facts, policies, inventory, cases, and source evidence remain externally updateable.
Use the smallest sufficient local, regional, or approved external model.
Train only with an explicit grant, immutable lineage, evaluation, canary, and rollback.
Structured Decision Intelligence Planned
Return a permitted candidate with source references and the qualified probability of that choice.
Keep false, unknown and abstention distinct. Ordinal labels follow a registered rubric rather than an invented score.
Missing evidence, unsupported inputs or stale qualification lead to explicit abstention, a permitted fallback or human review.
Capability status
This site describes EL-Ai-NE as designed in Baseline v1.0.0. The labels below say where each capability stands in the baseline build.
Switched on by default in the current baseline build. You can see it working in a proof of concept.
Built into the platform and shipped switched off. An accountable owner turns it on for a deployment once it is qualified for that deployment's profile.
Specified in Baseline v1.0.0 and on the delivery path. Not yet part of a deployable build.
| Capability | Status | Note |
|---|---|---|
| Governed core | ||
| Sign-in with local accounts and roles | Available now | Enterprise single sign-on is Planned. |
| Manager console: fleet, configuration, health and features | Available now | |
| Site node enrollment | Available now | A site node refuses requests until a policy authority is configured. |
| AI Gateway with fixed, policy-first stages | Available now | Every AI request passes the same stages. |
| Operational record with contradictions and unknowns kept visible | Available now | |
| Signed, append-only governance ledger and audit record | Available now | |
| Erasure and tombstones | Available now | A tombstoned item is never served again. |
| Emergency stop | Available now | Asserted and cleared by a verified person, with a reason. |
| Usage metering by request and route | Available now | Money figures need a rate card. |
| Database-enforced tenant separation, consequence classes, workload identities, data classification, retention bundles, signed backups | Included, enabled per qualification | Each engages with an owner-signed decision. |
| Enterprise single sign-on | Planned | Specified as the production identity path. |
| AI serving and knowledge | ||
| Local model serving and chat | Available now | One local model server. |
| Answer cache and context budget | Available now | |
| Knowledge search | Available now | Loading your own documents is not yet enabled. |
| Evaluation suite and model registry | Available now | |
| AI routing console: route policy, budgets and simulator | Included, enabled per qualification | |
| Durable assistant memory | Included, enabled per qualification | |
| Studio evaluation, canary and promotion | Included, enabled per qualification | |
| Local speech | Included, enabled per qualification | |
| External model providers | Planned | Off in every profile at this baseline. |
| Structured Decision Intelligence: typed decisions with abstention | Planned | |
| Governed Web Experiences and Persistent Agent Workspace | Planned | |
| Knowledge Wiki and content fabric | Planned | |
| Workflow and action | ||
| Agent and job runs with read-only tools; recommendations | Available now | Reliability depends on the local model you choose. |
| Action Center: propose, approve, run and roll back | Included, enabled per qualification | Every action is refused until a signed action allowlist is installed. |
| Connector catalog, outbound webhooks, computer use | Included, enabled per qualification | |
| Write connectors, change verification and active probes | Planned | |
| Network and endpoints | ||
| Network Truth: path, intent, exposure and health questions | Included, enabled per qualification | Read-only. |
| Endpoint estate and telemetry | Included, enabled per qualification | Partial at this baseline. |
| Full endpoint client fleet | Planned | |
| Trust, governance and reporting | ||
| Trust Center, evidence bundle and privacy registers | Included, enabled per qualification | |
| Assurance views and governed reporting | Included, enabled per qualification | |
| AI governance services: inventory, impact, oversight, transparency and post-market monitoring | Planned | |
| Financial close and chargeback | Planned | Usage metering is Available now. |
| Distributed operation | ||
| Cross-site state convergence and regional cells | Planned | |
| Disconnected operation with offline leases | Planned | No offline authority exists until a lease duration is approved. |
| High-availability production topology | Planned | Needs its own acceptance evidence. |
| External risk and assurance services | ||
| Assurance Exchange: trust portal and questionnaires | Planned | |
| Supplier risk and external exposure | Planned | |
| Shadow AI discovery and Shadow AI Guard | Planned | |
| Installation | ||
| Installation controller and signed release bundle | Included, enabled per qualification | Partial at this baseline. |
Labels describe the Baseline v1.0.0 build, not a production approval. Production use of any capability depends on qualification evidence for the accepted release and deployment profile.