Coordination Closure

Three placements of authority in AI-mediated institutions

Institutional systems do not act on information directly. They act after information has passed through some mechanism that makes action admissible.

A signal may become a fact.
A fact may support a judgment.
A judgment may justify authorization.
Authorization may produce an intent.
An intent may be executed.
Execution may later be reconciled into state.

AI does not remove these distinctions. It increases the number of components capable of producing plausible intermediate outputs.

A model can observe, summarize, classify, infer, recommend, plan, draft, invoke, monitor, and reconcile. None of those capabilities by itself says where authority lives.

This essay uses coordination closure for the point at which an underdetermined input becomes admissible action.

There are three useful closure strategies:

1. Cognitive Closure — ambiguity is handled by an intelligent runtime.
2. Ontological Closure — ambiguity is handled by canonical operational structure.
3. Invariant Closure — ambiguity is handled by explicit authority-preserving transitions.

These are not mutually exclusive components. A system may use models, ontologies, workflows, logs, interfaces, and humans at the same time. The distinction is which layer is allowed to turn uncertainty into consequence.

1. Coordination closure

A message arrives. It is not yet work.

A model emits a recommendation. It is not yet authority.

A workflow step completes. It is not yet proof that the world changed.

A dashboard shows state. It is not necessarily the authority that produced that state.

A coordination system must decide which transformations are allowed to connect these stages.

Example:

customer email
→ support classification
→ risk judgment
→ refund recommendation
→ refund approval
→ payment instruction
→ payment execution
→ ledger reconciliation

A model may assist at several points in this chain. But the system still has to distinguish:

classification
judgment
authorization
intent
effect
reconciled state

A closure strategy answers:

Where is unresolved ambiguity handled before action becomes admissible?

The three answers considered here are cognition, ontology, and invariants.

2. Cognitive Closure

Cognitive Closure places closure in the intelligent runtime.

The basic assumption is:

the model/runtime can resolve much of the residual ambiguity at use time

The runtime interprets context, fills gaps, maps informal input onto tools, chooses workflows, invokes APIs, drafts responses, routes exceptions, and repairs underspecification.

This is a natural direction for frontier-model companies because they begin with a general cognitive asset. To expand enterprise value, they tend to move outward:

model
→ assistant
→ agent
→ workflow runtime
→ organizational interface
→ operating substrate

At each step, more institutional ambiguity is handled by the model-adjacent layer.

Cognitive Closure is attractive when the environment is poorly formalized. The institution may not yet have clean object models, stable workflows, explicit exception handling, or well-defined authority boundaries. A model can still produce useful continuations.

Its strength:

high tolerance for underdetermination

Its capture mechanism:

users return to the runtime because it resolves ambiguity that surrounding systems have not made explicit

Its failure mode:

judgment-authority collapse

The model judges.
The system acts.
The boundary between recommendation and consequence becomes hard to inspect.

This is not guaranteed, but it is a pressure internal to the strategy. If cognition is the main place where ambiguity closes, cognition tends to move closer to authority.

Cognitive Closure can be summarized as:

admissible action follows from intelligent interpretation

3. Ontological Closure

Ontological Closure places closure in canonical operational structure.

The basic assumption is:

the institution needs explicit operational objects before action can be governed reliably

This is the Palantir-like direction, though not unique to Palantir.

The central object is not the model. The central object is the operational ontology: entities, relations, permissions, provenance, actions, workflows, constraints, and state.

The model is useful, but it operates against an already structured institutional world.

The movement is:

raw data
→ operational objects
→ governed relationships
→ permitted actions
→ institutional state

Ontological Closure does not ask the model to infer the institution from context on every pass. It tries to make recurring operational meaning explicit.

A shipment is not merely a string in a table.
A customer is not merely an embedding neighborhood.
A risk is not merely a generated summary.
An approval is not merely a conversational utterance.

Each becomes an object, relation, action, or state transition in a governed graph.

Compared with Cognitive Closure, this reduces some arbitrariness. Repeated meaning is not rediscovered each time by the model. It is stabilized in a structure that tools, humans, and models can all reference.

Its strength:

stable action over named operational objects

Its capture mechanism:

the platform becomes expensive to replace because workflows, permissions, objects, and state are encoded through it

This is a deeper form of dependency than access to a model alone. A model may be swapped more easily than the operational structure into which models are inserted, although in practice model replacement may still be costly because prompts, evals, tools, permissions, and workflows co-adapt.

Its failure mode:

centralized semantic control

Reality is made more explicit, but the explicit structure may become owned by a platform. The organization then depends on that platform not just for software execution but for the representation through which operations are understood.

Ontological Closure can be summarized as:

admissible action follows from canonical operational structure

4. Invariant Closure

Invariant Closure places closure in explicit authority-preserving transitions.

The basic assumption is:

no interpreting component should by itself own consequence

This is the Narada direction.

The system does not start by asking which central substrate should define operational reality. It asks which distinctions must be preserved so that consequence remains inspectable.

Important distinctions:

judgment ≠ authority
authority ≠ embodiment
embodiment ≠ mutation locus
intent ≠ effect
effect ≠ reconciliation
artifact ≠ governing structure
projection ≠ source of truth

A model may classify, propose, summarize, infer, or draft. Its output remains evidence until admitted by authority.

A UI may display. A clone may assist. A daemon may execute. A chat interface may mediate. None of these embodiments automatically owns mutation authority.

A task file may exist. The file is not necessarily the task lifecycle.

A recommendation may be correct. Correctness alone does not make it authorized.

Invariant Closure does not require one total ontology of the institution. It requires that transitions across components preserve the distinctions needed for audit, substitution, and recovery.

The movement is:

fact
→ governed admission
→ durable intent
→ authorized execution
→ reconciled effect

Closure occurs when the transition has preserved the invariants required for action to be admissible.

Its strength:

substitutability without loss of authority structure

Models can be replaced because models do not own authority.
Interfaces can multiply because interfaces do not own mutation.
Ontologies can remain local because cross-boundary consequence passes through explicit transitions.
Systems can be audited because judgment, admission, intent, execution, and reconciliation are not collapsed into one event.

Its weakness:

slower initial formation

Invariant Closure asks for distinctions before they always feel necessary. That can make early demos less fluid than Cognitive Closure and less platform-shaped than Ontological Closure.

It also reduces some forms of lock-in. If admissible interaction is explicit, value moves away from hidden interpretation and toward boundary correctness, auditability, replayability, authority preservation, interoperability, and local autonomy.

Those can still be commercialized through tooling, hosting, certification, governance, and operations. But they are less naturally dependent on one interpreter.

Invariant Closure can be summarized as:

admissible action follows from authority-preserving transition

5. Same parts, different authority

All three regimes may contain the same visible parts:

models
agents
tools
workflows
data stores
ontologies
interfaces
permissions
logs
humans

The difference is not the inventory.

The difference is which layer is allowed to close ambiguity into consequence.

Mega-AI and Palantir-like systems disagree about the primary closure substrate.

Mega-AI emphasizes cognition.

ambiguous situation
→ intelligent interpretation
→ action

Palantir-like systems emphasize ontology.

ambiguous situation
→ operational object model
→ action

Narada emphasizes authority-preserving transition.

ambiguous situation
→ evidence
→ governed admission
→ intent
→ execution
→ reconciliation

6. Capture and arbitrariness

Market capture tends to appear where unresolved arbitrariness remains operationally necessary.

If workflows are poorly defined, a cognitive runtime can capture value by resolving them dynamically.

If operational objects are poorly defined, an ontology platform can capture value by formalizing them.

If authority cannot be preserved across systems, boundary infrastructure becomes valuable.

The first two strategies produce clearer dependency surfaces.

Cognitive Closure captures because users rely on the runtime for interpretation.

Ontological Closure captures because institutions rely on the graph for operational representation.

Invariant Closure is structurally different. It reduces dependency on any single interpreting layer by making admissible interaction explicit.

This does not mean it cannot be captured commercially. It means the capture surface shifts.

Less value comes from:

owning the interpreter of ambiguity

More value comes from:

implementing, verifying, operating, and governing authority-preserving transitions

7. A PDA ordering

The three closure strategies can be arranged as a PDA sequence.

Not a historical inevitability. A structural ordering.

First: Cognitive Closure

When the space is highly underdetermined, cognition is often the cheapest closure mechanism.

No one has written the workflow.
No one has named the entity.
No one has specified the exception.
No one has defined the boundary.

The model can still produce a plausible continuation.

So Cognitive Closure often appears early.

Second: Ontological Closure

Repeated continuations create pressure for structure.

If the same kinds of objects recur, if the same decisions recur, if the same workflows recur, then repeated cognitive resolution becomes wasteful and risky.

The system begins to ask:

What are the objects?
What are the relations?
What actions are admissible?
Who may mutate what?
What state is authoritative?

This creates pressure toward Ontological Closure.

Third: Invariant Closure

At interoperation boundaries, ontology ownership can itself become arbitrary.

Which graph owns the object?
Which platform defines the action?
Which representation is authoritative across zones?
What happens when two ontologies interoperate?
What happens when authority must cross institutional boundaries?
What happens when the model, ontology, interface, and executor disagree?

At that point, the problem is not only better cognition or a larger ontology.

The problem is authority-preserving transition.

This creates pressure toward Invariant Closure.

The ordering:

Cognitive Closure
absorbs arbitrariness into intelligence

Ontological Closure
formalizes arbitrariness into operational structure

Invariant Closure
removes arbitrariness from authority boundaries

8. The Palantir / mega-AI overlap

Palantir-like systems occupy part of the space that mega-AI firms move toward after raw model access becomes insufficient.

Mega-AI begins above the institution:

general intelligence
→ enterprise interface
→ workflow penetration
→ operational substrate

Palantir-like systems begin inside the institution:

operational substrate
→ workflow control
→ ontology authority
→ model absorption

The overlap is the layer where institutional objects become actionable.

Mega-AI wants to move downward because model access alone is weak capture.

Palantir-like systems want to absorb models because operational ontology becomes more useful when intelligence can act over it.

Narada differs from both by treating that actionable layer as a boundary problem rather than a sovereign platform problem.

9. The distinction

Institutions will use models.
Institutions will use ontologies.
Institutions will use workflows, tools, interfaces, audits, and humans.

The useful distinction is where admissible consequence is produced.

Cognitive Closure:

admissible action follows from intelligent interpretation

Ontological Closure:

admissible action follows from canonical operational structure

Invariant Closure:

admissible action follows from authority-preserving transition

The first two concentrate closure in a substrate.

The third tries to make closure a property of explicit transitions between substrates.