DataFabPlatform Deeper Dive
DataFab
Platform

PlatformDeeper Dive

This document answers the four areas requested in your note.

Most platform overviews answer segregation by describing how tenants are isolated inside a shared store. Ours is answered on two different axes: your records remain in your source systems and are read in place, and what the platform derives from them is governed twice — by where it resides, and by attribute-based access controls carried on the knowledge graph itself.

Released under mutual NDA · Confidential
Part 1 — Architecture

A control plane that never holds your records.

The placement rule is simple and applied uniformly: a component's placement is set by the most sensitive class of data it touches. Anything that touches raw records or holds derived customer data sits with your data; anything that holds only definitions and control may sit in the managed control plane.

Control plane · definitions and control only

What it holds

The catalog index, policy and ontology, identity and entitlement configuration, agent and workflow definitions, orchestration, and telemetry.

Catalog indexPolicy & ontologyEntitlement configAgent definitionsOrchestrationTelemetry
It never holds your records
↓   placement set by the most sensitive class of data touched   ↓
Data plane · where the work happens

What it does

Connects to your systems, resolves what they describe into a knowledge graph, executes agents and rule chains, and stores the evidence for every execution. From the Dedicated edition upward, the entire data plane — including graph and inference — runs inside a boundary you designate.

ConnectorsEntity resolutionKnowledge graphAgents & rule chainsEvidence storeInference

The Knowledge Fabric

The data-plane component that turns your existing estate into a single, governed knowledge state — without migrating it. Select a stage.

Source-system health is monitored continuously — connectivity and credential validity every five minutes, freshness continuously, schema drift daily. If a source is unavailable, the affected subgraph degrades to a clearly flagged stale state rather than failing silently.

Hosting editions

Foundation
Data planeManaged, in-region
InferenceManaged, in-region
Key custodyDataFab-managed
Source recordsRemain in source
Dedicated
Data planeYour tenancy
InferenceYour boundary
Key custodyCustomer (BYOK)
Source recordsRemain in source
In-boundary
Data planeYour boundary
InferenceSelf-hosted
Key custodyCustomer-held
Source recordsRemain in source
Sovereign
Data planeAir-gapped
InferenceSelf-hosted only
Key custodySplit duty, your HSM
Source recordsRemain in source

Integration model

Connections are read-oriented and MCP-first. The platform discovers, catalogs, and resolves; it does not ingest your estate into a warehouse.

System classExamplesAccess modeWrite-back
Document managementiManage, NetDocuments, SharePointRead, catalog, resolveOptional, governed
Case / practice managementSector systems of recordRead, catalog, resolveOptional, governed
Client relationshipInterAction, Dynamics, SalesforceRead, catalog, resolveOptional, governed
Billing and ledgerFinance systems of recordRead, catalog, resolveNo
Mail and collaborationMicrosoft 365, Google WorkspaceRead, catalog, resolveNo
Data platformsExisting lake or warehouseRead in placeNo
External registersCompanies House, screening sourcesReadNo
Stated plainly

Write-back is off by default

Where enabled it is a separately approved capability: each write is an explicit governed action with its own authority record, conflict resolution, and cascade control.

Stated plainly

The source is the outer wall

Connectors authenticate with read-only, source-granted credentials — the source system's own permission model is the outer wall.

Stated plainly

Connectors are supply chain

A connector built in the platform's own tooling enters the same chain of custody as our artifacts — built, signed, approved, and run sandboxed with the same read-only credential model.

Third parties, in three categories

a · Sub-processors

Managed control plane

Cloud platform, identity, observability, notification, artifact registry. Each under a DPA; the list is published, with 30 days' notice before engaging any new sub-processor.

b · Model providers

Edition-dependent

Contractual training opt-out in every case; not present at all in In-boundary and Sovereign editions, where inference is self-hosted.

c · External data

The egress broker

OSINT, adverse media and sanctions screening are external by nature. Every outbound query passes a single audited broker enforcing an allow-list, field-level minimization, tokenization, rate limits, and a full audit of every term sent.

For higher-assurance deployments the pattern inverts: screening lists and licensed feeds are imported and matched inside your boundary, so no customer-derived attribute leaves. Sovereign deployments use feed-in and offline import only.

Part 2 — Delivery and tenancy

A tenant is a deployment, not a partition.

Tenancy depends on where the data plane runs — not on isolation controls layered over shared infrastructure.

Foundation is a managed, single-tenant deployment in a region you select; from Dedicated upward, the data plane runs in your own tenancy or boundary, with no shared runtime at all. This is a deployment-time decision made jointly at onboarding, alongside custody and residency.

Changes by edition

— Where the data plane and inference run

— Who holds keys, under what separation of duties

— Whether execution depends on our infrastructure

— Isolation: logical, infrastructure, or physical

Constant across every edition

— Source records remain in their systems, read in place

— The control plane holds definitions and control only — never records or derived knowledge

— Governance applied at discovery, traversal, action and evidence

— Every access and execution audited

Managed editions run under a published shared responsibility model. DataFab is responsible for platform hosting, patching, application security, encryption, and platform monitoring; you are responsible for data-classification decisions, role assignment and access policy, IdP configuration, and residency and retention requirements. Incident response is jointly defined. Customer auditors receive read-only access on approval, and quarterly access-review reports support your own recertification cycles.

Part 3 — Segregation

What stays in place, and what the platform derives.

Your operational records are not copied into a shared store, migrated, or warehoused. What the platform persists is a derived knowledge state — and we prefer to name its contents precisely rather than claim it holds nothing.

Persisted by the platform
  • Entity-resolution mappings — entity A in system X = entity A in system Y
  • Cross-system relationships discovered during resolution
  • Provenance, including the cited source passages evidencing each fact
  • Analyst annotations, tags and classifications
  • Derived insights and agent conclusions, with their evidence
  • Temporal snapshots for investigation history
Remains only in your source systems
  • The records themselves — names, addresses, financial data
  • Documents and case files
  • Communication logs — mail, messages
  • All authoritative operational data

That derived state is customer data, and we treat it as such: boundary-resident from the Dedicated edition upward, and encrypted under your keys from that point on.

Segregation on two axes

Axis one

Placement and custody

The edition determines where the derived knowledge state physically resides and who can decrypt it. From Dedicated upward, the graph, the evidence store and the audit store all sit inside your boundary under customer-managed keys; in Sovereign, key custody is split-duty under your HSM.

Exit is correspondingly simple: removing the deployment removes the derived state; your source systems are unchanged.

Axis two

Governance attributes on the graph

Within a deployment, access is segregated by attributes carried on the graph itself — so one knowledge state serves many roles without replicating it per audience.

SensitivityFive levels, Public to Highly Restricted, customizable to your taxonomy
ProvenanceSource, to character offset — trust, citation and source-permission alignment
Validity periodWhether the fact is current for the asking context
ConfidenceWhether the fact may be asserted, and how it must be qualified
Permission scopeWho may traverse the relationship — connection-level, not just record-level

A role may be permitted to know two entities exist without being permitted to see that they are connected. Connections are governed objects in their own right.

Enforcement is at query time, in the data plane. The policy decision and enforcement points sit with the data — they fail closed and do not depend on the control plane being reachable. Each request is evaluated against the caller's identity context, resolved fresh from your IdP every turn, so a revoked entitlement takes effect on the next interaction, not the next login.

No per-role copies

One governed graph

Views are derived at query time from a single governed graph, so there is no second store to drift, leak, or reconcile. Every access decision — subject, resource, action, outcome — is logged and retained.

The practical consequence

Expressed once, holds everywhere

Your ethical walls and need-to-know boundaries are expressed once, as attributes and policy over the graph, and hold across every surface — search, dialog, agents and reporting — rather than being re-implemented per application.

Schemas, metadata, residency

Schemas

Derived, not hand-built

Business-domain discovery reads your own documents and systems and proposes the schemas, which your teams review and refine. No schema-migration program precedes value.

Metadata

Catalogued in place

We treat source metadata as sensitive — your schema names and structures describe your business. From Dedicated upward it resides in your boundary alongside the graph.

Residency & exit

A configuration you own

UK-only, EU-only, or a specified region, for both data plane and control-plane metadata. On exit the deployment and its derived state are removed; nothing was taken from your source systems, so nothing needs returning.

Part 4 — Models and AI governance

Two governed forms of execution.

Deterministic rule chains

The regulated decision paths

A versioned DAG of explicit rules with defined scoring and decision routing, including hard overrides that no model output can override.

Given the same inputs, policy and versions, the same path executes — and the full chain trace, every rule, parameter, score and source record, is persisted as evidence.

Where a model assists inside a chain it does so as a bounded step whose output is validated, cited and carried with its confidence — with a deterministic fallback.

Schema-bounded agents

Analytical and conversational work

Each agent combines a language model with a structured query plan over governed sources and — the load-bearing property — a defined output schema.

The agent cannot emit fields or values outside its declared schema. Inputs are equally bounded: the query plan names the datasets, connectors and sub-agents it may touch, subject to the same graph-attribute governance as any human caller.

An agent acting for a user sees only what that user may see.

One fact base, many facets

Because every agent output conforms to a declared schema, a single resolved fact base can be rendered as multiple governed facets, each carrying its own definition and citations. Headcount — a measure that famously differs by who is asking. Select an audience.

All three are correct; they answer different questions. The reconciliation between them is computed and evidenced — so “which number is right?” becomes a governance answer, not an argument.

Where inference runs

EditionInference runsDo prompts containing customer data leave your boundary?Provider training
FoundationManaged, in-regionGoverned content reaches the in-region managed endpoint under DPA, via structured prompt templates with field-level minimizationContractually opted out
DedicatedYour boundary (default: self-hosted)No, by default. An external provider is possible only by your explicit sign-off, via the allow-listed egress pathContractually opted out
In-boundaryYour boundaryNoNot applicable
SovereignYour boundary, self-hosted, air-gappedNoNot applicable

Models are substitutable without changing agent definitions — the logic lives in the query plans, schemas and rules, not the model. The model version in force at execution is pinned and recorded with the execution.

Governance applied throughout execution

Not as a policy check on a finished answer.

Discovery
Catalog and sensitivity classification determine what may be seen at all
Traversal
Graph-attribute permissions govern each relationship
Action
Authority scope determines what an agent may do, not only read; tools and connectors allow-listed per playbook
Output
Defined output schemas — no fields outside schema; grounding validation requires citations back to source
Human gate
Execution halts at configured gates until a person acts
Evidence
Every output carries the reasoning path, sources and confidence that produced it
Audit
Immutable record of inputs, execution, outputs and governance decisions
Versioning
Which rule, agent, schema and model ran, in which version, on which day — reconstructable after the fact

Adversarial content is treated as a first-class threat. Agents read documents and mail, and treat content in those sources as data, never as instructions: the schema-bounded output contract means injected text cannot add fields, redirect tools outside the allow-list, or bypass a gate. Grounding validation rejects assertions without cited sources; consequential actions sit behind risk-classed human gates.

Human oversight — five modes

Configured per workflow, by you — a deployment decision with policy limits, not a runtime suggestion.

0
Fully supervised — every step reviewed
First deployment, high-risk domains
1
Human approves before any external effect
Regulated actions
2
Human approves at defined checkpoints
Routine governed work
3
Exception-based review, with escalation
Mature, measured workflows
4
Autonomous within a bounded authority scope, post-hoc audit
Low-risk, high-volume

Risk framework and assurance

Mapped to the NIST AI RMF functions so your risk team can assess us in a structure it already uses; ISO/IEC 42001 certification in progress, building on the existing ISO 27001 certification.

AreaPosition
CertificationISO 27001 and SOC 2 Type II — certified May 2026, audited annually. Architecture was recently Pen Test by Big4 for their own deployment
Security testingVulnerability scanning daily; third-party penetration test at least annually, report available; patch SLAs: critical 24–72 h, high 7 days, medium 30 days
Incident response24/7 SOC; P1 response within 1 hour with customer notification within 1 hour; regulator notification per requirement; post-incident report within 5 business days
ResilienceKnowledge graph RPO 1 h / RTO 4 h; API services RTO 15 min; audit logs RPO 15 min; region failover 4 h. Because sources remain the master, the resolved state is additionally re-derivable from your estate — loss and corruption are treated as distinct failure modes
Supply chainRelease artifacts signed and verified before execution; software bill of materials per release; customer approval gate before any release is applied to your deployment
TransparencyPublished sub-processor list with 30-day change notice; monthly service and quarterly security reporting; quarterly access-review reports
Part 5 — Deployment playbook

The architecture is fixed. Seven choices are not.

A reusable, interactive deployment playbook — six views from the logical reference through network, resilience, release and key topology, each edition-aware and each component inspectable. Deployment-specific values appear as placeholders rather than fixed numbers, because this is a reference pattern, not a network drawing.

01Reference deployment
02Governed utility
03Network topology
04Resilience & DR
05Build & release
06Identity & keys
07Deployment journey
Open the interactive playbook Switch editions to morph any view · click a component for detail

The seven deployment-time decisions

Everything else — the boundaries, the gateway, the model rule, the enforcement points — is fixed by the architecture and does not change per deployment. That is what makes these seven the whole of the deployment-time surface.

DecisionSet byWhen
EditionCustomer + DataFabDesign
Region / enclaveCustomerDesign
CIDRsCustomer infrastructureSetup
KMS / HSMSecuritySetup
SIEM targetSOC teamSetup
RPO / RTOBusiness + ITDesign
Source listData ownerDiscovery

Enforcement points C1–C5 are proven in the engineering evidence pack. Values shown as [set at deployment] are populated by engineering per environment.

Deployment Architecture Playbook Edition-aware · interactive