VIEW THIS AS

Auto mode follows the Route Engine until you choose a viewpoint.

YOU ARE HERE

ROUTE CHECK

CONNECTED TO

WHAT NEXT

Use the canonical route for this room, or HELP if you are unsure.

How to Create a Reliable Source of Truth for Super Intelligence | Systems of Record, Authority and Reconciliation

Three students work together at a classroom table, reviewing open textbooks and notes.

How do you create a reliable source of truth for Super Intelligence? Define which system, record, document or owner is authoritative for each important fact; make that authority explicit and current; preserve provenance; resolve conflicts through known precedence; and require SI to retrieve or verify against that source before consequential action.

This eduKateSG article completes the workplace knowledge and context layer. Earlier articles explain context engineering, SOPs, knowledge bases, document connections, enterprise search, structured and unstructured information, freshness, permissions, and memory. This page owns the authority question: when multiple sources disagree, which one is truth for the workflow?

In this series, Super Intelligence is the practical machine-intelligence layer commonly described as artificial intelligence, generative AI, assistants, copilots, agents and connected automation. A model can synthesise many sources fluently, but fluency does not decide institutional authority. The organisation must decide which facts come from which source.


Source of Truth Does Not Mean One Database for Everything

Most organisations do not have one universal source of truth. They have domain-specific authoritative systems: CRM for customer account state, finance for ledger figures, HRIS for employee records, version control for code, project systems for milestones, approved policy repositories for rules.

The goal is not to force every fact into one system. It is to make authority clear enough that SI knows where to verify each fact.

The Four Truth Layers

  1. Authoritative operational state: live systems of record.
  2. Authoritative policy and rules: approved documents or governed rule systems.
  3. Authoritative decisions: confirmed decision records, contracts or approvals.
  4. Derived interpretation: summaries, analysis, recommendations and memory built from the authoritative layers.

The first three establish organisational state. The fourth helps people understand or act on that state but should not silently replace it.

System of Record vs Source of Truth

A system of record is the authoritative application that owns a class of operational state. A source of truth is the authority for a specific fact in a workflow. They often coincide, but not always.

For example, CRM may be the system of record for customer accounts, while a signed contract may be the authoritative source for one contractual obligation. The workflow should know which source wins for each fact.

Single Source of Truth vs Reliable Truth Architecture

The phrase “single source of truth” is useful but can become misleading. Complex organisations need a reliable truth architecture: each important fact has one clear authority, even if different facts live in different systems.

Super Intelligence can then assemble a cross-system view without pretending the assembled summary is itself the authority.

The Fact Ownership Model

Truth becomes manageable when ownership is assigned at the level of fact categories. Customer contact details may be owned by CRM; product pricing by commerce or pricing systems; employment status by HRIS; policy by the relevant policy owner.

  • Fact category
  • Authoritative source
  • Owning team
  • Refresh expectation
  • Permitted downstream uses
  • Conflict rule
  • Historical source if needed
  • Change notification

This registry is one of the most useful assets for workplace SI because it turns source selection into an explicit operating rule.

The Field-Level Truth Model

A single object can contain fields with different authorities. A customer record may store contact information, billing status, support entitlement and contract metadata, while the signed contract remains authoritative for legal commitments.

Field-level authority prevents the organisation from treating one convenient database as authoritative for facts it merely copies.

The Document Truth Model

For policies, contracts, procedures and approved standards, authority should include version, effective date and owner. Drafts, examples and historical versions need a different status.

Retrieval should prefer the current authoritative document while preserving historical versions when the workflow legitimately needs them.

The Decision Truth Model

Important decisions should leave a durable record: decision, owner, date, evidence, assumptions and any trigger for reconsideration.

A meeting transcript or email thread may contain the discussion, but the confirmed decision record should become the authoritative state.

The Human Authority Model

Some truth comes from authorised human judgment rather than a database field. A legal opinion, medical determination, management approval or policy exception may require a qualified or authorised person.

The system should record who made the decision and under what authority rather than treating model output as equivalent.

The Truth Precedence Ladder

  • Live authoritative system for current operational facts.
  • Current approved document for policy and rules.
  • Signed or confirmed record for commitments and decisions.
  • Versioned reference for historical or technical context.
  • Derived summary for navigation and explanation.
  • Persistent memory for continuity and preference.
  • Model internal knowledge as background, not workplace authority.

The exact order varies by fact, but the principle remains: derived context should defer to explicit organisational authority.

The Conflict Rule

When two authoritative-looking sources disagree, SI should not silently choose one. The workflow should surface the conflict, apply the known precedence rule or escalate to the source owner.

Conflict is a state to resolve, not an invitation for fluent synthesis.

The Source Registry

A source registry lists important fact domains and their authority. It can be a table, catalogue or configuration service.

  • Domain
  • Fact type
  • Authoritative source
  • Owner
  • Access method
  • Freshness class
  • Version model
  • Conflict rule
  • Downstream workflows
  • Fallback source

The registry allows humans and SI systems to reason about where to look before retrieval begins.

The Provenance Rule

Every material derived claim should be traceable to its source. Provenance can include source ID, document version, record timestamp, query time or tool response.

Provenance makes verification and incident repair possible.

The Timestamp Rule

Truth has time. A customer entitlement that was correct yesterday may be wrong today. Record when the fact was retrieved or effective and refresh according to volatility.

The Effective-Date Rule

Policies and contracts can become effective in the future or remain relevant historically. Effective date is different from file-modified date.

SI should use the version applicable to the case’s time and jurisdiction where relevant.

The Snapshot Rule

Some workflows need a snapshot of state as it existed when a decision was made. Preserve the snapshot or source references so later review can reconstruct the decision context.

Do not overwrite history simply because the live system changed.

The Current-State Rule

For present action, refresh volatile facts close enough to execution. A cached summary may be acceptable for briefing but not for a payment, permission change or customer commitment.

The World-Return Rule

After SI executes an action, the destination system’s return becomes the new authoritative state. The agent’s intention or narrative is not enough.

This is critical for messages, tickets, calendar events, payments, record updates and deployments.

The Reconciliation Rule

When systems legitimately hold overlapping state, define reconciliation. Which field wins? Is the update one-way or bidirectional? Who resolves unmatched records?

Super Intelligence can help identify discrepancies, but the reconciliation policy must come from the organisation.

The Duplication Problem

The same fact is often copied into documents, spreadsheets, messages and dashboards. Copies become stale. Mark which representation is authoritative and treat the rest as derived views.

The Spreadsheet-as-Truth Problem

Spreadsheets can be legitimate systems of record for small workflows, but they often become uncontrolled parallel truth when copied repeatedly. Assign ownership, permissions, schema and version if the spreadsheet is truly authoritative.

The Email-as-Truth Problem

Email contains commitments and decisions but is difficult to govern as current operational state. Move accepted decisions and actions into the appropriate system.

The Chat-as-Truth Problem

Chat creates fast informal state. Important outcomes should return to project, CRM, decision or knowledge systems so later workflows do not depend on searching old threads.

The Memory-as-Truth Problem

Memory is valuable for continuity but can become stale and hard to inspect. It should remember what to retrieve and how to work, not silently own volatile organisational facts.

The Model-as-Truth Problem

A model can generate a plausible answer without a source. In workplace workflows, unsupported model knowledge should not override current internal systems, approved policy or qualified human authority.

The Search-Result-as-Truth Problem

Search ranking identifies likely relevance, not authority. The top result may be obsolete, unofficial or contextually wrong. Authority metadata should influence retrieval and verification.

The Dashboard-as-Truth Problem

A dashboard is a view over underlying data. If the dashboard refreshes slowly or applies transformations, the underlying system may remain the true source for certain decisions.

The Data-Warehouse-as-Truth Problem

Warehouses are powerful analytical sources but may lag live operational systems. Define whether they are authoritative for reporting, analysis or current action.

The Data-Lake Problem

A data lake may contain many versions and raw sources. It is not automatically a source of truth. Curated models, lineage and ownership are needed before SI can treat its contents as authoritative.

The API-as-Truth Problem

An API exposes a system’s view; it does not establish whether that system owns the fact. Tool design should preserve source identity and timestamp.

The Source-of-Truth Checklist

  • Fact category is defined.
  • Authoritative source is named.
  • Owner is named.
  • Version and effective date are visible.
  • Freshness expectation is known.
  • Permissions are enforced.
  • Provenance is preserved.
  • Conflict rule exists.
  • Historical state is preserved where necessary.
  • World-return updates current state.
  • Derived summaries are labelled as derived.

What This Article Owns

This page owns source-of-truth architecture: which source is authoritative for which fact, how provenance, currentness, conflicts, reconciliation and world-return create reliable state for Super Intelligence.

The preceding articles own knowledge freshness, permissions and memory. Together they complete the context and workplace knowledge layer.

The Core Truth Rule

Super Intelligence should synthesise across sources, but it should not invent authority between them.

Reliable workplace SI knows where each important fact comes from, how current it is, who owns it and which source wins when reality disagrees with memory or summary.


Truth Architecture Starts With Definitions

Before choosing systems, define the facts the organisation cares about. Terms such as active customer, booked revenue, approved supplier, current employee, open incident or completed project may have several informal meanings.

Super Intelligence will amplify ambiguity if definitions are not explicit. A reliable source of truth therefore includes a semantic layer: what each important fact means, how it is calculated and which source owns it.

The Semantic Contract

  • Fact or metric name
  • Definition
  • Unit
  • Calculation or rule
  • Authoritative source
  • Owner
  • Effective date
  • Refresh cadence
  • Permitted use
  • Known exceptions

The semantic contract gives both humans and SI a common interpretation before retrieval begins.

The Entity Identity Problem

Different systems may refer to the same customer, supplier, employee or project with different names or identifiers. SI can help match records, but entity resolution should use stable identifiers where possible.

Without a shared identity model, the system may mix two similar entities or treat one entity as several.

The Golden Record

For some entities, organisations create a golden record that reconciles several source systems into one governed representation. The golden record can be authoritative for selected fields while retaining provenance back to contributing sources.

Super Intelligence can consume the golden record more safely than guessing which duplicate record is current.

Master Data

Master data includes relatively stable entities such as customers, products, suppliers, locations and employees. Reliable SI benefits when master-data ownership and identifiers are clear.

A model should not invent a new product name or customer identity because it cannot reconcile inconsistent copies.

Transactional Data

Transactions represent events such as orders, payments, tickets or deployments. The transaction system usually owns the event state, while analytical systems may hold replicated or transformed copies.

For consequential action, SI should prefer the operational source or another explicitly approved current representation.

Reference Data

Reference data includes code lists, classifications, currencies, categories and standard mappings. These small datasets can have large effects on automated decisions.

Assign ownership and versioning so SI does not use obsolete categories or inconsistent definitions.

Derived Data

Derived fields and metrics are calculated from other facts. The organisation should record the formula or transformation and distinguish derived truth from raw source state.

This is especially important when a model explains metrics. The explanation should be grounded in the actual definition rather than a generic interpretation of the metric name.

Operational Truth vs Analytical Truth

Operational systems optimise for current transactions. Analytical systems optimise for reporting and historical analysis. Both can be authoritative for different questions.

A warehouse may be the approved source for monthly reporting even if the CRM is the live source for current customer status. The workflow must ask the right question of the right source.

Operational Truth

Operational truth answers questions such as: what is the current order status, who owns this ticket, is this account active, has this deployment completed?

It usually requires low latency and strong state integrity.

Analytical Truth

Analytical truth answers questions such as: what was revenue last quarter, how did churn change, which segment has the highest support load?

It may rely on curated transformations and a deliberate reporting cutoff rather than the live transaction system.

Historical Truth

Historical truth preserves what was known or true at a previous time. It is critical for audits, decisions, contracts and incident reconstruction.

Do not overwrite historical truth merely because current state changed.

Policy Truth

Policy truth comes from current approved policy or rule sources. It includes effective dates, scope and exceptions.

An email, chat message or remembered practice should not override approved policy unless an authorised exception has been recorded.

Contractual Truth

Signed agreements, orders or approved amendments can establish obligations that differ from default policy. SI should know when contract-specific terms override general guidance.

Decision Truth

Confirmed decisions should leave durable records. A recommendation, draft or meeting discussion remains different from an approved decision.

Decision truth is particularly important for agents that later act on prior approvals.

Identity Truth

Identity systems determine who a person or service account is and what roles or groups they currently hold. SI should not rely on remembered role information for access decisions when the identity system can be queried.

Permission Truth

Current permissions belong to the access-control system, not to model memory. An old memory that a user once had access cannot justify present retrieval.

The Truth Registry

Build a truth registry for the facts most important to SI workflows. The registry can start small and grow from real use cases.

  • Fact domain
  • Fact name
  • Authoritative source
  • Source owner
  • Identifier
  • Currentness requirement
  • Historical source
  • Conflict rule
  • Access rule
  • Dependent workflows

The registry prevents every team from rediscovering source authority independently.

The Truth Graph

Large organisations can think of truth as a graph: facts connect to sources, sources to owners, facts to workflows and workflows to actions. This makes dependencies visible.

A policy change can then identify which SI workflows require refresh or regression testing.

The Data Lineage Layer

Lineage shows how a fact moved from source through transformations into a report or model context. It answers: where did this number come from, what changed it, and when?

Lineage is especially valuable when SI explains metrics or when an agent acts on derived data.

The Calculation Lineage

For derived metrics, preserve formulas, filters, time windows and units. A model cannot reliably explain “conversion rate” if different teams calculate it differently.

The Document Lineage

For policies and documents, preserve draft → review → approval → effective → superseded relationships. Retrieval can then choose the correct status.

The Decision Lineage

For decisions, preserve evidence → recommendation → approval → action → outcome. This allows later review of both the decision and the information available at the time.

The Agent Action Lineage

For agentic workflows, preserve objective → context → proposed action → approval → tool call → tool return. The final tool return establishes external state.

The Truth Conflict Types

Timing Conflict

Two systems disagree because one is fresher. Apply the currentness rule appropriate to the fact.

Definition Conflict

Two teams calculate the same-sounding metric differently. Resolve the semantic definition before asking SI to reconcile values.

Ownership Conflict

Two departments each believe they own the field. Governance must designate authority or define a reconciliation process.

Copy Conflict

A derived copy diverges from the source. Refresh or retire the copy.

Exception Conflict

A general rule conflicts with an approved exception. The case-specific authoritative record may override the default.

Historical Conflict

Current truth differs from the state that applied when an earlier decision was made. Preserve both with time context rather than treating one as wrong.

The Reconciliation Patterns

Source Wins

One system owns the field; downstream copies are overwritten from it.

Latest Valid Wins

The most recent verified event determines state, provided event ordering is trustworthy.

Owner Review

Conflicts route to the owning team for resolution rather than automatic merge.

Rule-Based Merge

Different sources own different fields and a deterministic rule creates the combined record.

Case-Specific Override

A documented exception overrides the general source for a bounded case and time period.

No Automatic Resolution

High-consequence conflicts remain visible until an authorised human decides.

The Reconciliation Queue

Unresolved conflicts should enter a visible queue with owner, age, affected workflows and consequence. SI can group similar conflicts and prepare evidence.

A hidden conflict resolved differently by each agent becomes organisational drift.

The Freshness Contract

Every authoritative source should have a currentness expectation. Some facts are valid for months; others must refresh seconds before use.

The source-of-truth registry and knowledge-freshness layer should share these expectations.

The Replication-Lag Problem

A downstream system may be authoritative for reporting but lag the live operational source. Make lag explicit so SI does not use analytical data for real-time action unless allowed.

The Cache Problem

Caches improve speed but can return old state. High-impact workflows should know when cached values are acceptable and when a live refresh is required.

The Event-Ordering Problem

Distributed systems can receive events out of order. A later-arriving event is not always the latest real-world state. Use event timestamps, sequence numbers or source logic where necessary.

The Time-Zone Problem

Dates and deadlines can become ambiguous across regions. Authoritative time fields should include time zone or a clear standard.

The Unit Problem

The same value can mean different things if units differ. SI should preserve currency, measurement unit, tax treatment and scale when synthesising across systems.

The Null Problem

Missing data should remain missing rather than being filled with a plausible value. A reliable truth architecture distinguishes unknown, not applicable, zero and not yet entered.

The Default-Value Problem

Default values can look like real observations. Track whether a field was explicitly supplied or defaulted if the distinction affects decisions.

The Manual-Override Problem

Humans may override system values for legitimate reasons. Define where overrides are recorded, how long they apply and what source becomes authoritative during the override.

The Duplicate-Entity Problem

Two records for the same person or organisation can lead SI to combine or compare the wrong histories. Entity-resolution rules and stable identifiers reduce this risk.

The Deleted-Record Problem

A deleted or deactivated record may remain in search indexes, caches or memory. Source lifecycle should propagate to downstream SI context where appropriate.

The Soft-Delete Problem

Soft-deleted records remain stored for recovery or audit but should not appear as active truth. Status metadata must remain visible to retrieval.

The Archived-Document Problem

Archived material can be valuable for history but dangerous for current policy questions. Retrieval should use status and effective date to distinguish archive from authority.

The Draft-Document Problem

Drafts may contain newer text than the approved document. Recency alone should not outrank approval status.

The Signed-Document Problem

Signed contracts or approved orders can establish case-specific truth. SI should use the executed version rather than a later working draft unless the workflow is explicitly preparing an amendment.

The Human-Memory Problem

Experienced employees may remember how things usually work. Their knowledge is valuable, but repeatable authoritative facts should be externalised into governed sources where practical.

The Model-Memory Problem

Persistent personalisation should never become the hidden authority for policy, pricing, permissions or live business state. Memory can preserve continuity and preferences, then retrieve the truth.

The Derived-Summary Problem

An SI-generated summary can become widely copied and treated as fact. Label important summaries as derived and preserve links to the underlying sources.

The Truth Verification Step

Before consequential action, identify which material facts require re-verification and retrieve them from their authoritative source. The system need not re-check every low-risk detail every time.

The Pre-Action Truth Packet

  • Target object
  • Current authoritative state
  • Source and timestamp
  • Relevant policy or decision
  • Pending conflicts
  • User or agent authority
  • Proposed action
  • Expected new state

This packet gives both human reviewers and agents a consistent basis for action.

The Post-Action Truth Packet

  • Action attempted
  • Destination system
  • Tool return
  • New record or transaction ID
  • New authoritative state
  • Any partial failure
  • Required follow-up
  • Rollback state where relevant

The post-action packet turns execution into verified organisational state.

The Source Change Notification

When an authoritative source changes structure, ownership, definition or access method, downstream SI workflows should be notified or discovered through dependency management.

A silent source change can break retrieval even when the model itself remains unchanged.

The Source Retirement Process

  1. Declare replacement or end-of-life.
  2. Identify dependent workflows.
  3. Update source registry.
  4. Update retrieval and connectors.
  5. Expire caches.
  6. Archive historical content if needed.
  7. Run regression tests.
  8. Monitor for stale references.

The Source Migration Process

When moving from one system of record to another, define cutover time, reconciliation, historical access and rollback. During migration, both systems may appear current; SI needs an explicit transition rule.

The Source-Outage Process

If an authoritative source is unavailable, decide whether the workflow should wait, use a last-known snapshot, switch to a fallback source or require human approval.

Fallback behaviour should depend on consequence and staleness tolerance.

The External-Source Problem

Some workplace truth comes from external sources such as regulators, vendors, market data or public records. Record source identity, retrieval time and domain-specific authority.

A web search result is not automatically authoritative merely because it is recent.

The Human-Decision Source

When a qualified human decision is the authority, preserve the decision and scope in a durable system rather than relying on oral memory or a private chat.

The Truth Architecture for Agents

Agents need explicit tools for reading authoritative state, resolving identifiers, checking permissions and writing changes. The model should not be forced to infer truth from unstructured history when a structured source exists.

Agent plans should reference authoritative facts rather than carrying stale values through long chains without refresh.

The Truth Architecture for Multi-Agent Systems

Multiple agents should share explicit state or query the same authoritative systems. One agent’s summary should not automatically become another agent’s source of truth.

The Truth Architecture for Personal Assistants

Personal assistants can remember projects and preferences, but calendars, tasks, customer records and documents remain the authoritative sources for live work.

The Truth Architecture for Enterprise Search

Search should rank current authoritative sources above obsolete copies and preserve permissions, status and provenance.

Answers should make it easy to inspect the supporting source.

The Truth Architecture for Analytics

Analytics needs a curated semantic layer, calculation definitions, data lineage and reporting cutoff. SI explanation should be tied to the governed metric definition.

The Truth Architecture for Documents

Documents should carry status, owner, version and effective date. Search and memory should treat draft, approved, superseded and archived states differently.

The Truth Architecture for Decisions

Decision logs should identify authority, evidence, date, assumptions and later changes. This prevents future agents from interpreting informal discussion as approval.

The Truth Architecture for Actions

Every important external action should return state to a system of record. The workflow closes only when the destination acknowledges the change.

The Source-of-Truth Metrics

  • Facts with no registered authority
  • Source conflicts
  • Average conflict age
  • Obsolete sources still retrieved
  • Unresolved entity duplicates
  • Derived metrics without definition
  • Workflows acting on cached state beyond tolerance
  • World-return failures
  • Source changes without downstream retest
  • Memory conflicts with current truth

The Source-of-Truth Audit

  • Critical facts are defined.
  • Authority is registered.
  • Owners are known.
  • Identifiers are stable.
  • Definitions are consistent.
  • Freshness is explicit.
  • Version status is visible.
  • Permissions are enforced.
  • Provenance is preserved.
  • Conflict rules exist.
  • Historical state is available where needed.
  • World return closes external actions.
  • Memory and summaries defer to authoritative sources.

The Truth Operating Standard

A reliable workplace SI system can answer: What fact am I using? Which source owns it? When was it current? How was it derived? What happens if another source disagrees?

Those questions turn source-of-truth architecture from a data-engineering concept into an operating requirement for machine intelligence.


Domain Truth Map: Customer

Customer truth rarely lives in one place. CRM may own identity, account owner and opportunity state; billing may own payment status; support may own open cases; a signed contract may own service terms.

The SI workflow should assemble these views while preserving field-level authority. A customer brief can be a useful synthesis without becoming the system of record.

Domain Truth Map: Product

Product truth can include catalogue, pricing, inventory, technical documentation, roadmap and support status. Different systems may own each layer.

Customer-facing SI should use approved current product claims and live availability where commitments depend on it. Roadmap material should not be confused with released capability.

Domain Truth Map: Finance

Finance truth includes ledger, subledgers, invoices, payment systems, budgets and analytical reporting. Exact figures should remain traceable to governed financial systems and calculation definitions.

SI can explain movements and prepare commentary, but narrative explanation does not become the financial source of truth.

Domain Truth Map: HR

HRIS may own employment status, role and organisational assignment; payroll owns pay processing; learning systems own training completion; formal HR records own specific decisions.

Workplace SI should not infer employment truth from email or memory when governed records exist.

Domain Truth Map: Legal

Legal truth can include signed contracts, approved advice, matter records, current law and formal decisions. Draft clauses and model summaries remain derived artefacts.

The applicable jurisdiction and effective date can be as important as the text itself.

Domain Truth Map: Engineering

Version control owns code state, CI systems own test state, deployment systems own release state and monitoring owns production health. A generated engineering summary should link back to these systems.

Domain Truth Map: Operations

Operations may depend on inventory, ticketing, telemetry, schedules, runbooks and incident records. Real-time state should come from live systems where action is consequential.

Domain Truth Map: Education

Curriculum, assessment results, attendance, learning plans and safeguarding records may have different authoritative sources. SI can organise them for educational work while preserving appropriate access and professional judgment.

Domain Truth Map: Research

Research truth includes source publications, datasets, protocols, extraction tables and analysis code. Synthesis is derived and should preserve provenance to the original evidence.

Domain Truth Map: Procurement

Supplier master data, approved vendor status, contract terms, purchase orders, delivery records and payment state can live in different systems. SI should not treat one proposal PDF as current supplier truth after a contract or amendment supersedes it.

Domain Truth Map: Projects

Project plans, milestones, decisions, task state, risks and budgets may span several tools. The project system can own current execution state while decision logs and contracts own specific commitments.

Casebook: Twelve Source-of-Truth Decisions

Customer Delivery Date

The email thread says Friday, the order system says Monday and the salesperson remembers Thursday. The workflow should identify which source legitimately owns the delivery commitment. If the order system is operational truth but a later approved exception exists, that exception must be recorded explicitly rather than inferred from memory.

The key is to decide authority by fact type and workflow, not by whichever source is easiest for the model to retrieve.

Product Price

A marketing document lists an old price, CRM contains a copied figure and the pricing system has the current approved amount. The pricing system wins for a new quotation. The old marketing document should be retired or clearly marked historical.

The key is to decide authority by fact type and workflow, not by whichever source is easiest for the model to retrieve.

Employee Role

A project note says someone leads the team, but HRIS shows a role change and the organisation chart has not refreshed. The workflow should use the designated authoritative employment or organisational source and treat project notes as context.

The key is to decide authority by fact type and workflow, not by whichever source is easiest for the model to retrieve.

Contract Obligation

CRM notes summarise a promise, but the signed contract contains different language. For contractual obligation, the executed agreement is authoritative unless an approved amendment exists.

The key is to decide authority by fact type and workflow, not by whichever source is easiest for the model to retrieve.

Project Deadline

A meeting summary says March 15 while the project system shows March 22 after approved change control. The project system should own current schedule; the meeting record remains historical evidence of the previous state.

The key is to decide authority by fact type and workflow, not by whichever source is easiest for the model to retrieve.

Incident Severity

Chat participants use several severity labels, but the incident-management system has an official declared severity. That official state should control escalation and reporting until authorised incident command changes it.

The key is to decide authority by fact type and workflow, not by whichever source is easiest for the model to retrieve.

Deployment Status

An agent says deployment completed, but the deployment platform reports partial failure. The platform return wins. The agent must treat the action as incomplete and follow the recovery path.

The key is to decide authority by fact type and workflow, not by whichever source is easiest for the model to retrieve.

Invoice Payment

An email says payment was sent, but the payment system shows pending. The payment system owns transaction state. Communication should reflect the actual status rather than the sender’s intention.

The key is to decide authority by fact type and workflow, not by whichever source is easiest for the model to retrieve.

Policy Requirement

A colleague remembers the previous travel limit, while the current approved policy has a new threshold. The approved effective policy wins and the old memory should be corrected.

The key is to decide authority by fact type and workflow, not by whichever source is easiest for the model to retrieve.

Research Finding

An SI synthesis says evidence supports a conclusion, but the cited studies differ materially. The original sources and analysis method remain authoritative; the synthesis should represent uncertainty rather than become truth.

The key is to decide authority by fact type and workflow, not by whichever source is easiest for the model to retrieve.

Customer Preference

A CRM note records a customer’s preference for written updates. A later direct customer request changes that preference. Update the governed customer record if appropriate; do not let old memory continue to drive communication.

The key is to decide authority by fact type and workflow, not by whichever source is easiest for the model to retrieve.

Software Configuration

Documentation says one setting is enabled, but the live configuration service reports another. For current operation, the live configuration state wins; documentation should be corrected if it claims current state.

The key is to decide authority by fact type and workflow, not by whichever source is easiest for the model to retrieve.

The Truth Conflict Queue

When a workflow cannot resolve a source conflict automatically, create a visible queue. The queue should identify fact, competing sources, owners, affected workflow, consequence and age.

SI can prepare the evidence and group recurring conflict patterns, but source owners should resolve institutional ambiguity.

Conflict Severity

  • Low: cosmetic or explanatory difference with no action impact.
  • Medium: changes internal analysis or workflow routing.
  • High: affects customer commitment, financial state, employment, security, legal obligation or production action.
  • Critical: conflicting truth could cause broad or irreversible harm.

Severity can determine whether the workflow continues with a fallback or stops until resolution.

The Conflict Age Metric

Long-lived conflicts are evidence of unclear ownership or broken synchronisation. Track the age of important conflicts and the systems that generate them.

The Conflict Recurrence Metric

If the same field repeatedly conflicts, fix the data contract, integration or ownership rather than reconciling every case manually.

The Source Owner Contract

Every authoritative source should have an owner who can define the fact, approve changes, resolve conflicts and communicate material updates.

  • Source name
  • Fact domains owned
  • Technical owner
  • Business owner
  • Freshness expectations
  • Version model
  • Conflict process
  • Downstream dependencies
  • Retirement process

The Data Contract

When one system supplies another, define the fields, types, units, null semantics, update frequency and error handling. SI integration benefits from the same explicitness.

A model should not have to guess whether an empty field means zero, unknown or not applicable.

The Tool Contract

Agent tools should state what source they query, what state they return, timestamp, identifier and whether the response is authoritative or derived.

This allows the agent to reason about evidence quality rather than treating every tool result equally.

The Citation Contract

For material claims, the SI output should preserve source references in a form the user can inspect. Citations are useful only when they point to the actual supporting source and the user is authorised to access it.

The Metric Contract

For business metrics, define numerator, denominator, filters, period, currency or unit and owner. This reduces the common problem where two dashboards show different numbers under the same label.

The Entity Contract

Define stable identifiers and matching rules for customers, employees, suppliers, projects and products. AI-assisted entity resolution can help, but ambiguous matches should remain reviewable.

The Event Contract

For event-driven workflows, define event type, identifier, timestamp, source, ordering and retry semantics. This is especially important when agents react automatically to external changes.

The Truth Change Log

Material changes to source ownership, definitions, versions or precedence should be recorded. The log helps explain why the same question can legitimately produce a different answer after a policy or system change.

The Truth Notification Layer

Dependent SI workflows should be notified or flagged when critical source definitions change. The system may need re-indexing, cache invalidation, prompt/configuration updates or regression testing.

The Truth Regression Set

Keep representative questions and actions that depend on important sources. After source migration or rule change, confirm the workflow retrieves the correct authority and handles conflicts.

Source Migration: Dual-Run

During migration, old and new systems may operate in parallel. Define which source is authoritative at each phase and how discrepancies are reconciled.

Do not let SI choose based on whichever system responds faster.

Source Migration: Cutover

At cutover, update the truth registry, connectors, retrieval, caches, documentation and dependent workflow tests. Old sources may remain available for historical use but should lose current authority.

Source Migration: Rollback

If the new system fails, define whether authority returns to the old source or the workflow pauses. Rollback truth rules should be explicit before migration begins.

Source Retirement

A retired source can continue to influence SI through indexes, memory, copied documents and caches. Retirement should propagate through all active context layers.

Truth and Backups

Backups preserve historical copies but should not become active current sources. Restoration procedures should know how to re-establish current state after recovery.

Truth and Disaster Recovery

During an outage, a last-known snapshot may be the best available source. Label it clearly with timestamp and restrict consequential action if staleness exceeds the workflow’s tolerance.

Truth and Offline Work

Offline users or agents may work from cached data. Define what they may do without live verification and what actions must wait until connection returns.

Truth and External Web Research

For external current facts, identify authoritative or high-quality sources and retrieval date. Search ranking and model synthesis should not obscure source quality.

Truth and Vendor Documentation

Current vendor documentation may be authoritative for product behaviour, but local configuration can differ. SI should combine vendor guidance with the organisation’s actual environment.

Truth and Regulatory Sources

Regulatory, legal and standards information should preserve jurisdiction, effective date and source authority. Professional interpretation may still be required.

Truth and Human Expert Knowledge

Experts often know exceptions not written anywhere. Capture recurring legitimate knowledge into governed sources rather than relying indefinitely on personal memory.

Where judgment cannot be reduced to a rule, preserve the expert decision and authority rather than pretending a database can replace it.

Truth and Model-Generated Content

Generated content can become a new input to workflows. Label whether it is draft, recommendation, approved output or authoritative record. Status should be explicit.

Truth and Synthetic Data

Synthetic data can support testing and training, but it should not be confused with observed business facts. Keep origin metadata so SI knows whether a record represents reality or simulation.

Truth and Simulation

Scenario models and forecasts are conditional outputs, not present truth. Store assumptions and distinguish forecast state from observed state.

Truth and Predictions

A predicted churn risk or failure probability is derived intelligence. It can inform decisions but should not overwrite actual customer or machine state.

Truth and Recommendations

Recommendations belong in the decision-support layer. They may be accepted or rejected; acceptance should create a separate decision record.

Truth and Exceptions

Approved exceptions should be recorded in a governed place with scope, owner, effective period and relationship to the general rule.

An exception remembered only in email or human memory is difficult for SI to apply reliably.

Truth and Approvals

Approval status should come from an approval system or durable decision record. The model should not infer approval from supportive language.

Truth and Signatures

Where signatures establish commitment, use the executed signed artefact and verification process appropriate to the domain. A draft with the same wording is not equivalent.

Truth and Status Labels

Labels such as active, complete or urgent should have operational definitions. SI should not infer them from narrative tone when the workflow has explicit state rules.

Truth and Business Rules

Where a business rule can be represented deterministically, keep the rule explicit. SI can explain the rule or gather inputs without becoming the hidden rule engine.

Truth and Semantic Layers

A semantic layer standardises definitions across dashboards, analytics and SI. It is especially useful for metrics with multiple historical definitions.

Truth and Knowledge Graphs

Knowledge graphs can connect entities, facts and provenance, but the graph still needs ownership, currentness and authority. Structure alone does not make every edge true.

Truth and Vector Retrieval

Vector similarity finds semantically related content; it does not establish authority. Combine semantic retrieval with metadata such as status, owner, date and permission.

Truth and Reranking

Reranking can improve relevance, but current approved sources should receive appropriate authority signals so a more semantically similar obsolete document does not outrank the valid source.

Truth and Memory

Memory should know the fact domain and source-of-truth path. A remembered customer preference can be useful; a remembered account balance should trigger retrieval of current state.

Truth and Personalisation

Personal preferences can shape presentation without changing organisational facts. A user may prefer a different report style, but the same authoritative figures and policy still apply.

Truth and Permissions

A source can be authoritative and still inaccessible to a particular user. The workflow should not reveal restricted truth merely because it knows where the source lives.

Truth and Freshness

Authority without currentness can still produce wrong action. The source registry should include freshness expectations and update triggers.

Truth and Workflow Mapping

Workflow maps should show where authoritative state enters and leaves. This prevents derived summaries from becoming hidden substitutes for the original source.

Truth and Handoffs

Handoff packets should carry source references and current state, not merely narrative interpretation. The receiver should know which facts can be trusted without reconstruction.

Truth and Pilots

A pilot should identify source-of-truth before measuring SI quality. Otherwise, reviewers may disagree because they are checking different sources.

The Truth Case Review

When a material wrong answer occurs, ask whether the model failed, retrieval selected the wrong source, the source itself was wrong, the source was stale, definitions conflicted or the organisation never assigned authority.

These are different repairs.

The Truth Incident Review

  1. Preserve the output and action.
  2. Identify every source used.
  3. Identify the authoritative source.
  4. Compare timestamps and versions.
  5. Check permissions.
  6. Check conflict and precedence rules.
  7. Correct the source, retrieval or workflow.
  8. Check dependent systems.
  9. Add regression tests.
  10. Update the source registry.

The Truth Governance Meeting

For important domains, owners can review unresolved conflicts, stale sources, definition changes, migration status and workflows acting on derived state.

Governance should focus on exceptions and high-impact facts rather than manually reviewing every record.

The Truth Portfolio

As SI expands, maintain a portfolio of critical truth domains and dependent workflows. This reveals concentration risk when many agents depend on one source or semantic definition.

The Truth Maturity Ladder

Stage 1 — Informal

Truth lives in people, documents and local systems with unclear precedence.

Stage 2 — Registered

Critical facts have named sources and owners.

Stage 3 — Connected

SI retrieves authoritative state with provenance and permissions.

Stage 4 — Reconciled

Conflicts, lineage, migrations and historical truth are managed explicitly.

Stage 5 — Operational

Agents refresh truth before action, world-return closes state and source changes trigger dependent workflow checks.

The maturity ladder describes capability, not a requirement that every fact receive complex infrastructure.

The Truth Final Architecture Rule

Every important fact should have one explainable authority at the moment it matters, even if the organisation uses many systems overall.

Super Intelligence becomes safer and more useful when it can synthesise broadly while remaining disciplined about where reality is actually recorded.


The Source-of-Truth Architecture

A reliable SI workplace separates three layers: authoritative systems, interpretation layers and interaction layers. Authoritative systems hold current state. Interpretation layers retrieve, compare and explain. Interaction layers present information to people and agents.

The mistake is letting an interaction layer become authoritative merely because it is convenient. A chatbot answer can be useful, but it should not silently replace the CRM, finance system, contract repository or current policy.

The Three-Layer Model

  1. Authority layer: systems of record, controlled documents and confirmed decision logs.
  2. Interpretation layer: retrieval, search, transformation, reasoning and validation.
  3. Interaction layer: chat, dashboards, agents, email drafts and workflow interfaces.

This model gives SI freedom to reason while keeping truth anchored elsewhere.

The Read Path

A read path describes how SI reaches authoritative state. It should identify the query, access control, freshness requirement, source response and provenance returned to the model.

For high-impact work, the system should not rely on a cached narrative when it can retrieve the current record directly.

The Write Path

A write path describes how changes enter the source of truth. The path should be more controlled than reading because it alters future state for everyone.

  • Who or what requests the change?
  • What validation applies?
  • Who has approval authority?
  • Can the action be reversed?
  • What audit evidence is recorded?
  • What downstream systems must be notified?

Many useful SI workflows can remain read-only or prepare writes for human approval.

The Reconciliation Path

When two sources conflict, the system needs a reconciliation workflow. Do not let the model decide based on wording, frequency or confidence alone.

The reconciliation path identifies source owners, determines the correct state, updates the authoritative system and removes or marks stale copies.

The Source Priority Matrix

Different tasks can require different source priorities. A customer-service response may need CRM plus current policy. A contract analysis may need the legal repository plus the signed agreement. A project brief may need the project tool plus decision log.

Document these priorities so SI does not guess which source deserves authority.

Freshness Classes

Real-time or near-real-time

Inventory, access state, payments, system health and live operational conditions may require refresh immediately before action.

Daily

Project status, pipeline state and some operational dashboards may be acceptable at daily freshness depending on consequence.

Version-driven

Policies, SOPs and contracts change when a new version becomes effective. The relevant control is version status rather than elapsed time alone.

Event-driven

Employee role, customer status, project ownership or approvals can change through events that should invalidate cached state immediately.

Freshness SLAs

For each truth domain, define the maximum age acceptable for use. This is a data freshness SLA. It should match the cost of acting on stale state.

The SLA can differ across workflows. A project dashboard may be fresh enough for a weekly review but not for an urgent deployment decision.

Truth State Labels

  • Current: approved for present use.
  • Pending: proposed but not yet authoritative.
  • Superseded: replaced by a newer state.
  • Historical: retained for context only.
  • Disputed: conflict exists and requires reconciliation.
  • Unknown: no reliable current source has been established.

Explicit state labels prevent old or provisional information from being used as current truth.

The Unknown State Is Legitimate

A mature SI system can say that authoritative state is unknown. This is better than synthesising a confident answer from inconsistent records.

Unknown should route to the responsible source owner or exception workflow.

Source Lineage

Lineage records how a fact moved from origin through transformations into the current system. This matters when the organisation needs to explain why a field has a certain value.

For high-impact workflows, SI should preserve lineage rather than presenting the transformed answer as if it appeared from nowhere.

Derived Truth

Some useful state is calculated from authoritative inputs. A risk score, forecast or eligibility result may be derived rather than directly stored.

Derived truth should record the source inputs, rule or model used, timestamp and confidence or validation status where appropriate.

Reference Data

Definitions, product codes, taxonomy and standard rates are often shared across many workflows. These can become small but powerful sources of truth when clearly owned.

Master Data

Customers, suppliers, employees, products and locations may have master records used across systems. SI should know which record owns identity and which downstream copies are secondary.

Transactional Truth

Orders, payments, tickets and events are transactional. Their status often changes quickly. Agents should query the current transaction state before taking dependent action.

Policy Truth

Policy truth needs effective date, version, owner and status. Search relevance alone cannot establish which policy governs the current case.

Contract Truth

Contracts require clear distinction among draft, negotiated, executed and amended versions. The signed or otherwise authoritative contract state should be explicit.

Decision Truth

A confirmed organisational decision can create new truth: approved budget, chosen vendor, revised launch date, policy exception. Decision logs should capture authority and effective state.

The Cache Problem

Caching improves speed but can create stale truth. Every cache should have invalidation or refresh logic appropriate to the underlying source.

The faster the source changes and the more consequential the action, the less the system should rely on old cached state.

The Snapshot Problem

Reports and exports are snapshots. They may be correct at creation time but not later. Label snapshots with date and purpose so SI does not treat them as live operational state.

The Data Warehouse Problem

A warehouse can be authoritative for analytics while operational systems remain authoritative for live transactions. Different truth domains can legitimately use different systems.

The Search Index Problem

An enterprise search index is a retrieval layer, not automatically the source of truth. It may contain stale copies or delayed sync. The system should retain links back to original authoritative content.

The Vector Database Problem

Embeddings and vector stores help semantic retrieval but do not create authority. They should carry source IDs, version metadata and freshness signals so retrieved chunks remain traceable.

The RAG Problem

Retrieval-augmented generation can improve grounding, but retrieval quality depends on source governance. Feeding more documents into the index can reduce reliability if obsolete or conflicting sources remain active.

The API Problem

An API exposes data but does not by itself prove authority. The source behind the API, update latency and semantics still matter.

The Human Override Problem

Humans may know an authoritative system is wrong. The correct response is to use the governed correction process, not to create a permanent prompt exception around the error.

The Local Spreadsheet Exception

A local spreadsheet can be the right authoritative source for a small team if ownership and update discipline are strong. The question is not enterprise scale; it is whether the truth is explicit, current and controlled.

The Source-of-Truth Registry

A compact registry can document the truth map across the organisation.

  • Domain
  • Authoritative system
  • Owner
  • Read method
  • Write method
  • Freshness
  • Versioning
  • Conflict process
  • Fallback source
  • Known limitations

The registry can later become machine-readable so agents know which tool to call.

The Truth Router

A mature SI system can route questions to the correct authority. Policy questions go to the governed knowledge source; account state goes to CRM; financial state goes to finance; project state goes to the project tool.

This is more reliable than one undifferentiated retrieval pool.

The Truth Router and Context Engineering

Context engineering can use the truth registry to decide which sources belong in the context window for each task.

This makes context selection systematic rather than dependent on user memory.

The Truth Router and Agents

Agents can use the same registry to choose read tools and write tools. Read authority and write authority may be different even within one domain.

The Truth Router and Search

Enterprise search can use source authority as a ranking signal. A current official policy should outrank an old presentation that happens to contain similar words.

The Truth Router and Memory

Memory can remember the user’s usual workflow and preferred sources, but the truth router should still confirm the current authoritative location.

Worked Example: Customer Service

The customer asks whether a refund is permitted. The workflow needs current order state from CRM or commerce, current refund policy from the governed policy source and any authorised exception record. A previous email or remembered preference is secondary.

Worked Example: Finance

A manager asks why expenses increased. The ledger or finance system owns the numbers. SI can retrieve the figures and supporting transactions, then generate narrative analysis. The narrative is interpretation, not financial truth.

Worked Example: HR

An employee asks about leave entitlement. Current employee status and policy both matter. HRIS supplies employee state; the current policy supplies the rule. SI combines them without inventing either.

Worked Example: Procurement

A procurement agent prepares a purchase request. Supplier identity, approved terms, budget availability and authorisation each come from different truth domains. The agent should not treat the proposal document alone as sufficient.

Worked Example: Engineering

A coding agent needs the current branch, tests, issue state and deployment environment. Repository and CI systems own these facts. Documentation can explain the architecture but not replace current build status.

Worked Example: Education

An educator asks for a learner plan. Current assessment evidence and curriculum are authoritative inputs; an old lesson note may be historical context. SI can combine the sources while the educator retains pedagogical judgment.

Worked Example: Legal

Legal review should use the current contract version and current applicable authority. An old summary can support history but should not silently govern the analysis.

Worked Example: Operations

An incident agent should read live telemetry and current runbook versions before recommending action. A cached status report may be useful context but cannot replace live system health.

The Truth Conflict Queue

When conflicts appear, route them into a visible queue with source owner, affected workflows and resolution priority.

This turns data disagreement into maintained work rather than repeated downstream confusion.

Conflict Severity

  • Low: formatting or descriptive difference with no effect on action.
  • Moderate: discrepancy affects reporting or workflow routing.
  • High: discrepancy affects money, rights, security, safety or external commitment.

Resolution priority should reflect consequence.

The Truth Correction Loop

  1. Detect issue.
  2. Identify owner.
  3. Confirm correct state.
  4. Correct authoritative source.
  5. Invalidate caches and derived copies.
  6. Notify affected workflows.
  7. Retest material use cases.

The Truth Migration Problem

When systems are replaced, the source-of-truth mapping must change deliberately. During migration, both old and new systems may coexist, creating ambiguity.

Define cutover dates and which system wins at each stage.

The Truth Merger Problem

Organisational mergers can create duplicate sources for customers, employees and policies. SI should not choose automatically until authority and reconciliation rules are established.

The Truth SaaS Problem

Third-party software can be the authoritative source if the organisation has decided that it owns the relevant domain. Vendor boundaries do not change the need for provenance, permissions and continuity.

The Truth Continuity Problem

If a critical source is unavailable, define whether the workflow waits, uses a last-known-good snapshot, falls back to another system or switches to manual handling.

Fallback state should be clearly marked so stale information is not mistaken for live truth.

The Last-Known-Good Rule

A last-known-good value can preserve continuity during outages, but the system should disclose its timestamp and limit actions that require fresh state.

The Source Availability Metric

Track whether authoritative sources are available when workflows need them. An SI system cannot be reliable if its truth layer is frequently inaccessible.

The Source Latency Metric

Some integrations return current data slowly. Latency can affect user experience or cause timeouts in agentic workflows. Measure it separately from model latency.

The Source Quality Metric

Track completeness, consistency and known correction rates for material fields. Source-of-truth governance does not mean pretending the source has zero errors.

The Provenance Coverage Metric

Measure what percentage of material SI claims can be traced to an authoritative source or controlled derivation.

The Conflict Rate Metric

Track how often two candidate sources disagree. Repeated conflict indicates architecture or ownership debt.

The Correction Latency Metric

Measure how long it takes to fix a wrong authoritative value after detection. Long correction times encourage downstream workarounds.

The Source Ownership Metric

Track how many critical sources lack a clear owner or review process.

The Stale-Use Metric

Measure how often a workflow used a superseded or out-of-date source. This is one of the most direct quality indicators for retrieval-heavy SI.

The Write-Safety Metric

For agents writing to systems of record, track rejected writes, rollbacks, conflicts and unauthorised attempts.

The Source-of-Truth Audit

  1. Inventory material truth domains.
  2. Name canonical sources.
  3. Confirm owners.
  4. Map read paths.
  5. Map write paths.
  6. Define freshness.
  7. Define versions and effective dates.
  8. Find duplicates.
  9. Test conflict handling.
  10. Test outages and fallback.
  11. Test SI provenance.
  12. Review permissions.

The Source-of-Truth 30-Day Build

Week 1 — Map

Choose one workflow and list every material fact it uses. Name current sources and conflicts.

Week 2 — Govern

Assign owners, define freshness and mark obsolete copies. Establish correction process.

Week 3 — Connect

Give SI read access to the approved sources with provenance and version metadata.

Week 4 — Test

Run stale, conflicting and missing-source cases. Confirm the system escalates rather than fabricates certainty.

The Source-of-Truth Promotion Gate

Before a source becomes machine-authoritative, confirm owner, versioning, freshness, access, correction and conflict behaviour.

The Source-of-Truth Demotion Gate

Demote or retire a source when ownership disappears, updates stop, a new authoritative system replaces it or conflict rates become unacceptable.

The Source-of-Truth Deletion Test

If a duplicate can be removed with no loss of authoritative state, remove or archive it. Reducing active copies improves retrieval quality.

The Source-of-Truth Simplification Rule

Prefer one explicit authority per fact class rather than complex precedence among many overlapping copies.

The Source-of-Truth Security Rule

Authority does not imply universal access. Sensitive systems should remain permissioned according to legitimate role and purpose.

The Source-of-Truth Final Standard

A workplace source of truth is reliable when humans and SI agree on where current state lives, how it changes, who owns it and what happens when sources disagree.

Once this layer is stable, instruction design becomes simpler because the system can be told not only what to do, but where to obtain the facts it is allowed to use.

The Source-of-Truth Checklist Before SI Acts

  • Is the relevant truth domain identified?
  • Is the source current enough for this action?
  • Is the source authoritative rather than merely relevant?
  • Can the user or agent access it legitimately?
  • Is the version or timestamp visible?
  • Are important fields complete?
  • Do other candidate sources conflict?
  • Does the action require a second source or approval?
  • Will the external write return observable state?
  • Can the source be corrected if the value is wrong?

This checklist can be embedded into high-impact workflows so SI verifies the truth layer before it reasons or acts. It is particularly useful when currentness changes quickly or when several systems appear relevant.

The Source-of-Truth and Human Judgment Boundary

Authoritative facts do not eliminate human judgment. A system can establish the current contract, balance, project state or policy, while people still decide what those facts mean for strategy, negotiation, fairness or professional advice.

A source of truth should therefore make judgment better informed rather than pretend every decision can be reduced to database lookup.

The Source-of-Truth and Explanation Boundary

Super Intelligence can explain an authoritative record in plain language, compare it with prior state and surface implications. The explanation remains secondary to the record itself. If the explanation conflicts with the source, the workflow should resolve the interpretation instead of rewriting the source through generated text.

The Source-of-Truth and Documentation Boundary

Documentation can explain how systems work and what users should do. Systems of record preserve live operational state. A strong workplace SI design links the two: procedures tell the system how to interpret and use the state, while the state tells the system what is true now.

The Source-of-Truth and Decision Log Boundary

Some decisions create new authoritative state before a database has been updated. A confirmed decision log can temporarily carry authority if the workflow defines that role and later reconciles the operational system.

The dangerous condition is informal discussion pretending to be authoritative without confirmation, owner or effective date.

The Source-of-Truth Exception Packet

When truth cannot be established, the system should return an exception packet instead of a guessed answer. The packet can include the conflicting sources, timestamps, owners, affected workflow and the exact fact that requires reconciliation.

This gives the responsible person a focused task rather than a vague warning that the model is uncertain.

The Source-of-Truth Operating Rule

Every material SI action should be able to answer: Which source made this fact authoritative, how current was it, and what would override it?

When that answer is visible, generated language, agents and automation can move faster without forcing people to guess where the underlying truth came from.

The Source-of-Truth Reliability Tests

Before a truth source is used in an important SI workflow, test it against four conditions: authority, availability, currentness and correction. Authority asks whether the organisation really treats this source as definitive. Availability asks whether the workflow can reliably reach it. Currentness asks whether its state is fresh enough for the decision. Correction asks whether a discovered error can be repaired at the source instead of patched downstream.

A source that fails one of these tests may still be useful as reference material, but it should not quietly carry the full weight of authority.

The Source-of-Truth Fallback Standard

When a critical source is unavailable, the workflow should know whether to wait, use a clearly labelled last-known-good value, switch to a secondary controlled source or escalate. The fallback should never masquerade as live truth.

For consequential actions, stale fallback data may be suitable only for preparation rather than execution. The system should disclose the age and limitation of the information so the human or agent can choose the correct next step.

The Source-of-Truth Review Meeting

For important domains, review the source map periodically with the data or process owner and the teams that consume it. Ask whether the designated source still reflects real operating practice, whether duplicate copies have appeared, whether update latency changed and whether downstream SI systems still receive the right metadata.

This review is especially useful after system migrations, organisational restructures, policy changes or the introduction of agentic automation.

The Final Reliability Standard

A source of truth is reliable when authority is explicit, current state is observable, conflicts are resolvable and corrections propagate back into the system rather than accumulating as local exceptions.

That standard gives Super Intelligence a dependable factual floor. Once facts are anchored, the next challenge becomes instruction: telling the system exactly what work to perform, what constraints to respect and what output will count as acceptable.

The Source-of-Truth Floor

Before any workplace SI workflow is allowed to treat a fact as settled, it should be able to identify the authoritative source, the owner of that source, the time or version at which the fact was valid, and the process for correcting it. These four properties form the minimum factual floor for reliable automation.

When the floor is missing, lower autonomy. Let SI search, compare and prepare questions, but do not let it convert unresolved organisational ambiguity into confident external action. Reliable intelligence begins with reliable state.


Build the Truth Registry From Real Work

Do not begin by cataloguing every database and document in the company. Start with the workflows already using or planning to use Super Intelligence. List the material facts those workflows depend on, then assign authority.

This use-case-first approach keeps the source registry small enough to maintain while still covering the facts that can change decisions or actions.

The Minimum Truth Registry

  • Fact: the specific state or rule needed.
  • Definition: what the fact actually means.
  • Authority: the system, document or authorised human that owns it.
  • Owner: who maintains or resolves it.
  • Freshness: how current it must be.
  • Identifier: how the relevant entity or record is found.
  • Conflict rule: what happens when another source disagrees.
  • Historical rule: where previous state is preserved.
  • Permission rule: who may access it.
  • Dependent workflows: where the fact is used.

The registry becomes the map that context engineering, retrieval, memory and agents consult before material action.

The Truth Registry Should Be Versioned

Authority can change. A company may migrate CRM, replace a policy repository or redefine a business metric. Version the registry so historical decisions can be interpreted using the source rules that existed at the time.

The Truth Registry Should Have Change Triggers

Source migration, policy change, merger, system redesign, ownership transfer, new data warehouse, new region or incident can all change the truth model. These events should trigger review of affected fact domains and dependent SI workflows.

The Truth Registry Should Be Queryable

For larger deployments, agents and developers can benefit from a machine-readable source registry. The workflow can ask where to retrieve current price, employee status, account owner or policy rather than hard-coding every source path.

Machine readability should complement human ownership, not hide the governance decision inside code.

The Semantic Layer as a Truth Contract

Business terms often fail before data does. One team defines active customer as a purchase in 90 days; another defines it as an open contract. SI cannot reconcile the numbers until the meaning is reconciled.

A semantic layer makes terms, metrics and categories explicit. This improves dashboards, search, analysis and SI reasoning simultaneously.

The Metric Dictionary

  • Metric name
  • Business definition
  • Formula
  • Numerator
  • Denominator
  • Filters
  • Time window
  • Currency or unit
  • Owner
  • Source tables or systems
  • Refresh cadence
  • Historical changes

When SI explains a metric, it should use this definition rather than the ordinary-language meaning of the label.

The State Dictionary

Operational states such as Open, Active, Approved, Complete, Escalated and Closed should have explicit definitions and transition rules.

Agents become easier to govern when they act on observable states rather than vague prose.

The Identifier Dictionary

Stable identifiers reduce entity confusion. Record which ID joins the customer, supplier, employee, product, project or incident across systems.

Where one stable ID does not exist, define matching rules and a review path for ambiguous matches.

The Truth Map for Customer Service

  • Customer identity → CRM or identity system
  • Entitlement → contract/subscription system
  • Open case state → support platform
  • Product policy → approved knowledge base
  • Refund authority → policy plus approval system
  • Message sent → communication platform return

The support brief can synthesise all six without becoming authoritative for any of them.

The Truth Map for Sales

  • Account ownership → CRM
  • Opportunity stage → CRM
  • Current approved pricing → pricing system
  • Contract terms → executed agreement
  • Product claims → approved product documentation
  • Customer commitments → confirmed meeting/email record returned to CRM

The Truth Map for Finance

  • Ledger figures → accounting system
  • Payment status → payment system
  • Budget → approved planning system
  • Management metric definitions → semantic layer
  • Variance commentary → derived analysis
  • Approval state → governed approval record

The Truth Map for HR

  • Employment status → HRIS
  • Payroll state → payroll system
  • Approved role → HRIS/organisation record
  • Performance decision → formal HR process
  • Training completion → learning system
  • Policy → approved HR policy repository

The Truth Map for Engineering

  • Code → version control
  • Build/test state → CI system
  • Deployment state → deployment platform
  • Production health → monitoring
  • Incident state → incident system
  • Architecture standards → approved engineering documentation

The Truth Map for Legal

  • Executed obligation → signed agreement
  • Current matter state → matter-management system
  • Approved legal position → formal advice/decision record
  • Current law/regulation → authoritative legal source
  • Template clause → approved precedent library
  • Draft language → non-authoritative working document

The Truth Map for Procurement

  • Approved supplier → supplier master
  • Purchase order → procurement system
  • Contract terms → executed agreement
  • Delivery state → logistics/operations system
  • Payment state → finance
  • Evaluation commentary → derived decision support

The Truth Map for Education

  • Curriculum → approved curriculum source
  • Assessment result → assessment system
  • Attendance → school/learning system
  • Learning plan → authorised educational record
  • Generated feedback → derived instructional artefact
  • High-stakes decision → authorised institutional process

The Truth Map for Research

  • Source claim → original publication or dataset
  • Extraction → evidence table with citation
  • Analysis → code/notebook and method
  • Interpretation → researcher synthesis
  • Protocol → approved research plan
  • Final conclusion → published or approved research output

Truth Architecture for Small Businesses

A small business may legitimately use a spreadsheet, shared drive and email as core systems. The same principles still apply: designate one master file, one current pricing sheet, one customer system and one approved policy location.

Reliability comes from explicit authority and ownership, not from enterprise software size.

Truth Architecture for Large Enterprises

Large organisations need domain ownership, data contracts, semantic layers, lineage and portfolio-level dependency maps. Shared SI infrastructure should consume these governed layers rather than build its own hidden parallel truth.

The Truth Conflict Playbook

  1. Stop consequential action if the conflict matters.
  2. Identify the fact in dispute.
  3. Identify competing sources.
  4. Check timestamps, versions and definitions.
  5. Apply the registered precedence rule.
  6. Escalate to the fact owner if no rule resolves it.
  7. Record the resolved state.
  8. Repair duplicate or stale sources.
  9. Check other workflows exposed to the same conflict.

The Truth Gap Playbook

Sometimes no authoritative source exists. Treat that as a gap, not permission for the model to invent one. Assign an owner, decide whether a new field, policy or decision record is required, and keep the workflow human-led until the gap is resolved.

The Truth Unknown State

Unknown should be a valid explicit state. A reliable SI system can say the organisation does not currently know rather than replacing missing truth with plausible language.

The Truth Disputed State

Disputed means credible sources or owners disagree. Preserve the disagreement and route it instead of collapsing it into one generated answer.

The Truth Pending State

Pending means an authoritative process has not completed. A pending approval should not be treated as approved, and an attempted transaction should not be treated as settled.

The Truth Superseded State

Superseded facts remain useful historically but should not drive current action. Retrieval should know the difference between historical relevance and active authority.

The Truth Invalid State

Invalid records should be excluded from ordinary use but may be retained for audit. The status should remain machine-readable.

The Truth Provenance Packet

  • Fact value
  • Source
  • Source record ID
  • Timestamp or effective date
  • Version
  • Transformation if derived
  • User or workflow that retrieved it
  • Permission context
  • Confidence/status
  • Conflict state

The packet does not need to be shown in every response, but it should be available for material decisions and incidents.

The Truth Pre-Action Gate

Before a consequential action, refresh the facts that can change the action’s legitimacy: target identity, current amount, current permission, policy, approval state and live system status.

This avoids acting on stale context collected earlier in a long session.

The Truth Post-Action Gate

After action, verify destination state and update the workflow. If the tool returns uncertainty, do not invent a successful state. Reconcile before continuing.

The Truth Human-Approval Gate

Where a human decision establishes authority, capture the decision explicitly. The model should not infer approval from silence, attendance or supportive conversation.

The Truth Automation Gate

Automate only when source authority, identifiers, currentness and conflict handling are stable enough that the machine can distinguish normal from exceptional state.

The Truth Agent Gate

Agents need a truth map for every tool-driven decision. If the agent cannot identify the current source for a critical fact, it should retrieve, ask or escalate.

The Truth Multi-Agent Gate

Multiple agents should not create a chain of summaries where each agent treats the prior summary as truth. Preserve links to the original source and shared state.

The Truth Memory Gate

Persistent memory can guide retrieval but should not bypass authoritative state. Volatile facts should be refreshed before action.

The Truth Search Gate

Search results should include authority, version and date signals. Semantic relevance alone is not enough for workplace truth.

The Truth Dashboard Gate

Dashboards should disclose reporting cutoff, definitions and source. SI should not compare two dashboard metrics until it knows they use compatible definitions.

The Truth Spreadsheet Gate

If a spreadsheet is authoritative, give it owner, access rules, schema, version and change process. If it is only a working copy, mark it as derived.

The Truth Document Gate

If a document controls behaviour, its status and effective date should be machine-readable. Draft and approved should be impossible to confuse.

The Truth Decision Gate

If an action depends on a previous decision, retrieve the confirmed decision record, not merely meeting notes or a remembered summary.

The Truth Migration Checklist

  • Old and new sources identified
  • Fact domains mapped
  • Cutover rule defined
  • Historical access preserved where needed
  • Conflict reconciliation completed
  • Connectors updated
  • Indexes and caches refreshed
  • Dependent workflows tested
  • Rollback plan available
  • Old source authority retired

The Truth Source-Retirement Checklist

  • Replacement or end-of-life confirmed
  • Downstream workflows identified
  • Search index updated
  • Memory conflicts checked
  • Caches invalidated
  • Documentation updated
  • Permissions removed where appropriate
  • Historical archive labelled
  • Regression tests passed

The Truth Failure Catalogue

  • No authoritative source assigned
  • Two systems claim authority
  • Source is authoritative but stale
  • Draft outranks approved version
  • Derived copy mistaken for source
  • Entity mismatch
  • Metric definition mismatch
  • Unit mismatch
  • Historical state mistaken for current
  • Memory overrides current state
  • Agent assumes tool success
  • Approval inferred rather than recorded
  • Deleted source remains indexed
  • Migration leaves dual authority
  • Permission blocks user from inspecting evidence

The Truth Repair Sequence

  1. Identify the material fact.
  2. Find every source currently supplying it.
  3. Assign or confirm authority.
  4. Correct definitions and identifiers.
  5. Resolve current conflicts.
  6. Repair retrieval and integrations.
  7. Retire misleading copies.
  8. Add provenance and currentness checks.
  9. Run regression cases.
  10. Update the source registry.

The 30-Day Source-of-Truth Build

Week 1 — Fact Inventory

Choose three important SI workflows and list the facts they require.

Week 2 — Authority

Assign source and owner for each material fact. Identify gaps and conflicts.

Week 3 — Retrieval and Reconciliation

Test whether SI retrieves the authoritative state and handles conflicts or stale copies correctly.

Week 4 — Operationalise

Create the registry, add regression cases, set change triggers and link source owners to dependent workflows.

The 90-Day Truth Architecture Build

Days 1–30 — Critical Facts

Cover only high-impact facts used by current SI workflows.

Days 31–60 — Shared Semantics

Standardise metric definitions, identifiers and key state labels across teams.

Days 61–90 — Lifecycle Control

Add lineage, migration procedures, conflict queues, source-change notifications and portfolio monitoring.

The architecture should grow from workflow evidence rather than attempt to catalogue the universe before value exists.

The Source-of-Truth Owner Triangle

  • Business owner: defines what the fact means and why it matters.
  • System owner: maintains the platform or data source.
  • Workflow owner: ensures the SI use case retrieves and applies the fact correctly.

One person may hold several roles in a small organisation, but the responsibilities should remain understandable.

The Truth Portfolio Dashboard

  • Critical facts without authority
  • Source conflicts
  • Conflicts older than SLA
  • Sources past freshness tolerance
  • Source migrations in progress
  • Retired sources still retrieved
  • Dependent SI workflows
  • World-return failures
  • Memory-source conflicts
  • Metric-definition changes

The Truth Review Meeting

For large SI portfolios, review exceptions rather than every source. Focus on conflicts, stale authority, migration, incidents and source changes affecting high-impact workflows.

The Truth Transfer Test

Before reusing a source-of-truth pattern in another workflow, confirm that the new workflow needs the same fact, currentness and authority. Reporting truth and real-time action truth may differ.

The Truth Deletion Test

Ask whether a derived copy, local spreadsheet or remembered summary can be removed without harming the workflow. Fewer redundant truths reduce reconciliation cost.

The Truth Simplification Rule

Do not create a new master database if the existing system of record is already reliable. Source-of-truth architecture is about clear authority, not centralisation for its own sake.

The Truth Human-Agency Rule

Users should be able to inspect important sources, challenge conflicts and understand when the system is uncertain. A truth layer should make reality more legible, not more hidden.

The Truth Final Checklist

  • Important facts are defined.
  • Fact authority is registered.
  • Source owners are named.
  • Identifiers are stable.
  • Definitions are shared.
  • Freshness requirements are known.
  • Permissions preserve inspectability.
  • Provenance is available.
  • Conflict rules are explicit.
  • Historical state is preserved.
  • Derived outputs are labelled.
  • Memory defers to current sources.
  • Agent actions return authoritative state.
  • Source changes trigger dependent workflow review.

Frequently Asked Questions

Does Super Intelligence need one single source of truth?

No. Different fact domains can have different authoritative systems. The important requirement is one clear authority for each material fact in the relevant workflow.

Can a knowledge base be the source of truth?

Yes for approved knowledge such as policy, procedures or product documentation, provided ownership, version and effective date are governed. It may not own live transactional state.

Can a CRM be the source of truth for everything about a customer?

Usually not. CRM may own account state while contracts, billing and support systems own other facts.

Can memory be a source of truth?

Memory is best for continuity and preferences. Volatile facts and institutional authority should remain in governed sources.

What should SI do when sources disagree?

Apply the registered precedence rule if one exists; otherwise surface the conflict and route it to the source owner rather than guessing.

What is provenance?

Provenance is the trace from a claim or value back to its source, version, timestamp and any transformations.

How fresh must truth be?

It depends on the workflow. Monthly reporting and real-time payment action have different freshness requirements.

What is world return?

World return is the verified response from the destination system after an action. It establishes what actually happened rather than what the agent intended.

What if no authoritative source exists?

Treat it as a knowledge or governance gap. Assign an owner and keep consequential decisions human-led until authority is established.

What comes next?

The next part of the workplace SI series moves from knowledge and context into delegating work to Super Intelligence: instructions, prompts, structured outputs, constraints, review and recovery.

The Final Source-of-Truth Principle

Super Intelligence may combine many sources, but every consequential fact should remain traceable to one legitimate authority for that moment and purpose.

That is the foundation that lets retrieval, memory, agents and automation become more powerful without allowing fluent synthesis to replace organisational reality.