← Frameworks
Position paperDesign synthesisVersion 1.0

Knowledge Architecture Beyond the Single Consumer

When knowledge becomes an interface

A platform now explains itself to developers, agents, auditors, and software that acts before a human can repair an ambiguity. Each consumer needs a different interface. All of them need compatible beliefs about what the system does, where the claim applies, who stands behind it, which version it describes, and how to check it.

Knowledge architecture is the closed-loop design of that relationship. Responsible sources represent relevant system state as scoped assertions and project them for declared consumers. Recorded outcomes then supply evidence that can support, qualify, contest, or supersede those assertions.

Working thesis

Reliable knowledge systems preserve material meaning across different projections and reconcile those projections against observed outcomes.

Evidence status

This page presents a design synthesis and an illustrative capability. Its claim envelope, projection invariants, and measures are implementation proposals awaiting comparative validation.

01 · The problem

One system fact, three chances to drift

Consider one illustrative rule: the v2 upload endpoint rejects decoded payloads above 10 MB, defined here as 10,000,000 bytes after transport decoding. A developer needs a plain-language limit and recovery path. An agent needs a typed boundary before tool selection. An assurance reviewer needs the requirement, owner, release, and test result. The material fact is shared; its useful form changes with the decision.

When teams maintain those forms as independent artifacts, each can begin accurately and still diverge after a release. A guide says 10 MB, a manifest still declares 5 MB, and a control checklist omits the version. Duplicate prose creates a maintenance problem. Divergent truth conditions create an epistemic one.

Independent copies after a version change

Developer guide

10 MB

current

Agent manifest

5 MB

stale

Assurance check

10 MB

version missing

“Single-consumer coupling produces rebuild cycles” is shorthand here for a conditional mechanism. Translating a new interface from artifacts written for a previous decision context can produce a rebuild cycle. Under that condition, each new consumer may trigger another rebuild. A shared assertion layer changes the work: teams model the material claim once, preserve its coordinates, and govern each projection against that record.

Existing information architecture already supports modularity, reuse, metadata, and multiple renderings. DITA explicitly supports different deliverables from shared content (OASIS DITA Technical Committee 2026). OpenAPI describes interfaces for people and computers (OpenAPI Initiative 2024). The new threshold appears when a knowledge artifact participates in operational selection, execution, assurance, or recovery. Such an artifact needs interface-grade semantics, versioning, provenance, failure states, and observation.

When a knowledge surface can cause action, it becomes part of the system's control surface.

Authorization remains with policy and execution systems. The knowledge surface shapes what a consumer believes those systems permit, require, and produce.

02 · The model

The unit of architecture is a claim envelope

Pages are delivery surfaces. The durable unit beneath them is a scoped assertion: a proposition with enough coordinates for another reasoner to interpret, evaluate, and trace it. A claim envelope gives that assertion an address and carries its conditions wherever it travels.

Epistemic interoperability lets different consumers form scoped, current, and traceable beliefs from shared assertions. Their projections are compatible when they do not contradict one another about the material invariants, even if consumers draw different task-specific conclusions.

Illustrative claim envelope

JSON
{
  "id": "files.upload.max-size",
  "type": "specification",
  "assertion": "POST /files rejects decoded payloads above 10,000,000 bytes with HTTP 413.",
  "scope": {
    "interface": "Public Upload API",
    "operation": "POST /files"
  },
  "authority": "API platform team",
  "validity": {
    "systemVersion": "api-v2",
    "effectiveFrom": "2026-08-20"
  },
  "constraints": [
    "10 MB means 10,000,000 bytes, measured after transport decoding"
  ],
  "verification": {
    "method": "conformance test",
    "record": "upload-limit-v2"
  },
  "status": "confirmed"
}

Identity

Which assertion is this?

A stable address lets every projection, test, and revision refer to the same object.

Assertion and type

What is being asserted?

Observation, specification, policy, and promise carry different evidentiary obligations.

Scope

Where does it apply?

Interfaces, actors, configurations, jurisdictions, and boundary conditions prevent accidental generalization.

Authority

Who stands behind it?

The responsible source and approval state remain visible when the assertion travels.

Validity

For which state and time?

System version, effective interval, and supersession links bind meaning to the state being described.

Constraints

Which conditions change the meaning?

Limits, preconditions, exclusions, failure semantics, and uncertainty stay beside the assertion they govern.

Verification

How can it be checked?

A test, record, observation, or cited source gives the consumer a route from assertion to support.

Status and conflict

What is its epistemic state?

Confirmed, disputed, unknown, withheld, superseded, and withdrawn are explicit states rather than silent gaps.

The envelope is an intermediate model, not a replacement for DITA, OpenAPI, MCP, JSON Schema, W3C PROV, or domain-specific assurance formats. Teams can serialize these fields into existing standards and retain local extensions where the decision requires them.

03 · Consumer projections

Different interfaces can preserve the same material assertion

Every projection removes detail. A useful human explanation, compact tool schema, and assurance record emphasize different relations. Knowledge architecture decides which losses are safe for a declared decision and which qualifiers must survive.

A summary carries obligations. It may reduce resolution while preserving scope, constraints, uncertainty, and a route to deeper support. Projection fidelity names the correspondence between a source assertion and the material meaning represented in a view.

Human guide

Upload size

Help a developer prepare a valid request and recover from failure.

Assertion files.upload.max-size
The api-v2 Public Upload API accepts decoded payloads up to 10 MB (10,000,000 bytes). POST /files returns HTTP 413 above that decoded limit.
Authority: API platform team · Effective: 2026-08-20 · Status: confirmed
Verification: conformance record upload-limit-v2

Decision supported

Resize, split, or reject the file before submission.

Invariant across all views

files.upload.max-sizePOST /files10,000,000 decoded bytesapi-v2API platform team2026-08-20confirmedHTTP 413upload-limit-v2

Meaning

The material proposition and its qualifiers survive.

Coordinates

Scope, authority, system state, and time remain explicit.

Trace

The projection points back to its assertion and support.

Absence

Unknown, disputed, and withheld fields remain visible states.

Software engineering already treats consistency among multiple views as a problem that requires explicit relations and periodic restoration. Distributed work makes temporary inconsistency likely, which makes unrecognized inconsistency the dangerous state (Stevens 2020). Knowledge projections need the same discipline, with semantic invariants defined around the decisions each view supports.

04 · Closed-loop architecture

A knowledge system should reconcile with reality

Publishing completes a delivery step. Operational knowledge remains open until accountable teams can compare assertions and projections with recorded outcomes, investigate discrepancies, and update or dispute the responsible record. The loop makes that maintenance observable without treating any observation as self-interpreting.

01

System state

Accountable sources select relevant configuration, code, policy, and observations.

02

Claim envelope

A scoped assertion with authority, validity, constraints, and support.

03

Projection

A view shaped for a declared consumer and decision context.

04

Belief and action

Selection, interpretation, execution, review, or escalation.

05

Evidence and reconciliation

Recorded outcomes support, qualify, contest, or supersede the assertion.

Design consequence

Every omission delegates a decision to the consumer's default.

An explicit unknown invites a question, safe rejection, or escalation. A silent gap invites each consumer to supply its own assumption. Structured absence belongs in the model because it changes behavior.

05 · Agent application

DNCE turns the envelope into an execution lifecycle

Declaration, Negotiation, Contract, and Evidence describe the questions an agent must resolve from discovery through reconciliation. DNCE applies the shared claim layer across an agent execution lifecycle without creating a parallel knowledge system.

Negotiation has a literal role here. The consumer supplies task constraints. A provider or broker returns compatible options or structured mismatch reasons. A local orchestrator can apply the same comparison semantics without a network exchange.

D

Declaration

What is exposed?

Reads
Identity, assertion, scope, authority, validity
Produces
A discoverable capability with explicit boundaries
N

Negotiation

Does it fit this task?

Reads
Consumer constraints, cost, latency, permissions, alternatives
Produces
A fit decision or a structured mismatch reason
C

Contract

What governs execution?

Reads
Inputs, outputs, state changes, errors, retry and idempotency terms
Produces
An execution agreement the consumer can interpret
E

Evidence

What happened?

Reads
Result state, event record, validation output, causal identifiers
Produces
A checkable outcome and a route back into reconciliation

MCP already provides names, human-readable descriptions, input and output schemas, structured results, and behavioral annotations. Its specification also says that tool annotations are hints and should not drive decisions when the server is untrusted (Model Context Protocol 2025). DNCE begins at that boundary: declarations support discovery, while negotiation, contract, policy, and evidence determine whether and how an agent should act.

06 · Knowledge as a control surface

Knowledge surfaces govern through defaults and expectations

Language and metadata shape recognition, selection, recovery, and migration. Contracts establish expectations. Code and policy enforce authority. Keeping those levels distinct makes the governance claim useful without turning a description into an access-control mechanism.

SurfaceDecision shapedDesign obligation
DescriptionWhether a consumer recognizes and selects a capabilitySeparate purpose, authority, and risk in direct language.
Missing fieldWhich default fills the gapEncode unknown, inapplicable, withheld, and unsupported states.
Error modelWhether a consumer retries, revises, aborts, or escalatesPublish typed failure states and bounded recovery paths.
Version signalWhether prior assumptions remain validBind every material projection to system state and supersession.

Governance is distributed across several actors. Domain owners establish assertions, while knowledge stewards preserve coordinates and projection rules. Client applications and operators own trust policy, approval flows, and local logs. Servers enforce access controls. Outcome records should preserve identifiers that let investigators locate where a mismatch first appeared and test which layer introduced it.

07 · Put it to work

Start with one brittle decision

A team can test this architecture without rebuilding its publishing stack. A one-week pilot should follow a small set of consequential assertions from source through two projections and one observed outcome.

  1. 01

    Choose five consequential assertions

    Start where a wrong belief changes selection, cost, safety, compliance, or recovery. A page inventory can follow later.

    Done when: Each assertion has a named consumer decision and failure cost.

  2. 02

    Build the envelopes

    Extract identity, type, scope, authority, validity, constraints, verification, status, and known conflicts from existing sources.

    Done when: A reviewer can see both what is known and what remains unresolved.

  3. 03

    Produce two projections

    Render one human view and one machine view from the same envelopes. Preserve consumer-specific language and structure.

    Done when: Both views retain the material invariants required for their decisions.

  4. 04

    Add conformance checks

    Test envelope completeness, projection invariants, version binding, links, and any executable behavior that supports the assertion.

    Done when: A release fails when a material projection drifts from its envelope.

  5. 05

    Reconcile observed outcomes

    Capture mismatches among the record, the projection, consumer interpretation, policy, and runtime behavior. Route each mismatch to its owner.

    Done when: The next revision changes the responsible layer and records why.

Pilot question

Can a developer, an agent, and a reviewer correctly select, use, recover, and verify one capability from the same governed assertions without private context?

08 · Measurement

Measure correspondence instead of content volume

Page counts and word counts describe output. The loop needs measures that reveal whether material assertions carry their coordinates, survive projection, match deployed state, and recover from mismatch.

Envelope coverage

material assertions with complete envelopes / material assertions sampled

Shows how much decision-critical knowledge has explicit coordinates and support.

Projection fidelity

material invariants preserved / material invariants tested

Detects summaries, schemas, and views that change the meaning while changing the form.

Temporal alignment

operational projections correctly bound to the system state they describe / operational projections tested

Distinguishes a correct historical record from a stale operational interface.

Verification reach

assertions with a usable verification path / material assertions sampled

Measures whether consumers can move from published language to inspectable support.

Reconciliation latency

median time from detected mismatch to corrected or explicitly disputed envelope and projections

Measures how quickly the system corrects drift or records a legitimate disagreement.

These measures are proposed audit instruments. Their usefulness depends on whether they predict lower selection error, safer recovery, faster integration, or more accurate assurance. Field evaluation must isolate the source of failure among the knowledge record, projection, consumer, policy, and runtime.

09 · Boundaries

What the model preserves, and what it cannot settle

Structure can expose the conditions of a claim and improve the quality of downstream reasoning. It cannot manufacture truth, eliminate judgment, or predict every emergent outcome. The architecture earns trust by making those boundaries operable.

Formalization leaves a remainder

Judgment, experience, and emergent behavior can outrun a schema. The envelope should mark the need for interpretation, an owner, and an escalation path instead of manufacturing precision.

Provenance does not establish truth

Authority and lineage show where an assertion came from. Verification, corroboration, and observed correspondence determine what weight it earns.

Consistency does not require consensus

A shared substrate can carry competing assertions, different authorities, and unresolved conflicts. Epistemic interoperability depends on exposing those relations faithfully.

Shared semantics do not require central ownership

Federated teams can own separate records when they share identifiers, status semantics, provenance links, and rules for reconciliation.

10 · Convergence

One semantic substrate, plural interfaces

Technical writing, API description, agent metadata, disclosure, and assurance remain distinct practices. They converge on a shared engineering problem: preserving material assertions while their form, resolution, and decision context change.

Existing standards supply important parts. DITA separates reusable content from delivery. OpenAPI describes an API surface for people and computers. MCP supplies tool schemas and structured results while drawing a boundary around untrusted hints. W3C PROV provides a common model for provenance and responsibility (W3C Provenance Working Group 2013). Each solves a real layer of the problem.

DITA

Reusable topics and multiple deliverables from shared content

Delivery reuse alone does not test whether operational qualifiers survive every view.

OpenAPI

A formal API description consumable by people and computers

Source-state binding, evidence, and decision-specific projections still depend on implementation.

MCP

Tool schemas, structured results, and behavioral annotations

Annotations are hints; policy, trust, verification, and task fit remain distributed responsibilities.

W3C PROV

A common model for entities, activities, agents, and provenance

Provenance supports assessment without proving truth or specifying action semantics.

The new object

The object of knowledge architecture is the correspondence among source state, claim envelopes, consumer projections, and observed outcomes.

The page, manifest, and assurance record remain plural. Shared semantics and a reconciliation loop let them differ without quietly becoming different accounts of reality.

The practical next move is small: choose one brittle decision, model the assertions it depends on, produce two faithful projections, and observe what happens. Repeating that loop reveals which invariants reduce drift, improve recovery, and deserve to become operating requirements.

Sources and status

References

  1. Model Context Protocol. 2025. “Tools.” Specification revision November 25, 2025. modelcontextprotocol.io. Accessed August 28, 2026.
  2. OASIS DITA Technical Committee. 2026. “Darwin Information Typing Architecture (DITA) Version 2.0.” dita-lang.org. Accessed August 28, 2026.
  3. OpenAPI Initiative. 2024. “OpenAPI Specification, Version 3.1.1.” October 24, 2024. spec.openapis.org. Accessed August 28, 2026.
  4. Stevens, Perdita. 2020. “Maintaining Consistency in Networks of Models: Bidirectional Transformations in the Large.” Software and Systems Modeling 19: 39–65. https://doi.org/10.1007/s10270-019-00736-x.
  5. W3C Provenance Working Group. 2013. “PROV-O: The PROV Ontology.” W3C Recommendation, April 30, 2013. w3.org/TR/prov-o. Accessed August 28, 2026.

Available for reading as a working draft.

Version 1.0 · revised