Intelligence-Authority Separation
Intelligence-Authority Separation is a control architecture for letting fallible, non-deterministic intelligence participate in consequential operations while durable authority governs truth, permission, lifecycle, and consequence.
Its central rule is:
Intelligence may contribute judgment. Authority must remain in governed structure.
The pattern applies when an intelligent process can perceive, infer, classify, draft, recommend, or request action, and where its output requires mediation before becoming operative state.
The opposite of separation is collapse.
In this context, collapse means that distinctions needed for accountable consequence are lost under operational pressure. The most dangerous form is cognitive-action collapse:
model judges X
therefore system does X
Intelligence-Authority Separation exists to prevent judgment from becoming consequence without mediated authority.
1. Positioning
Intelligence-Authority Separation names the institutional boundary required when non-deterministic judgment enters real operations.
The claim is structural: intelligence and authority are different functions.
Intelligence supplies candidate meaning. Authority determines what counts, what may change, what may be done, who or what may do it, and how consequence is confirmed.
Compactly:
fallible intelligence + durable boundaries + governed consequence
= accountable operation
Narada is an instantiation of this pattern. It places AI charter evaluation inside a control plane where facts, work, decisions, intents, execution, and reconciliation have durable identity and explicit authority boundaries.
Related concepts:
- Progressive De-Arbitrarization: the underlying discipline for exposing decision-relevant hidden arbitrariness.
- PDA Organizational Transfer: the organizational form of closure persistence and structural differentiation.
- IAS Organizational Transfer: the organizational projection of IAS onto roles, responsibilities, authorization, execution, and audit.
- Plural Embodiment, Singular Authority: the structural rule that multiple operational embodiments may request or assist while one declared locus owns each governed mutation.
- Governed Crossing: the more primitive boundary transition of which intelligence entering consequence is one specialized case.
- Constructive Coherence Coordinates: the development-posture coordinate system in which IAS collapse appears as a cross-coordinate failure.
- Universal Systems Constructor: the construction calculus that can build IAS-preserving systems.
- Constructive Universalization by Re-instantiation: the abstraction discipline under which IAS is accepted as a reusable invariant only when re-instantiated.
2. Core Type
Given an observed operational state S, an intelligent evaluator I, and an authority structure A, the pattern has the shape:
IAS : (S, I, A) -> (S', E) | R
Where:
| Symbol | Meaning |
|---|---|
S |
current durable operational state |
I |
fallible intelligence producing judgment or proposals |
A |
authority structure: policy, lifecycle, permission, review, effect boundaries |
S' |
transformed durable state after governed admission and execution |
E |
evidence of judgment, decision, execution, and confirmation |
R |
residual: rejected proposal, blocked action, ambiguity, failed validation, or unconfirmed effect |
The intelligent output is an input to authority-bearing transformation, not the transformed state itself.
3. Separation State
A separation state can be represented as:
IAState = (F, C, W, V, D, N, X, Q, O)
| Component | Meaning |
|---|---|
F |
facts: durable truth boundary derived from observation |
C |
context: bounded materialization presented to intelligence |
W |
work: explicit lifecycle-managed operational unit |
V |
evaluation: intelligence output stored as evidence, not authority |
D |
decision: governed admission, rejection, or review requirement |
N |
intent: durable representation of permitted consequence |
X |
execution: mechanical effect performed by authorized worker |
Q |
confirmation: reconciliation of effect against the outside world |
O |
observation: inspectable view of durable and derived state |
A state is authority-preserving when every transition from V onward is performed by an authority-bearing component rather than by the intelligent evaluator.
4. Operational Boundary
The operational loop is:
observe
-> normalize
-> fact
-> context
-> work
-> evaluation
-> decision
-> intent
-> execution
-> reconciliation
-> observation
This loop is a sequence of boundaries. Each boundary preserves a specific distinction:
| Boundary | Preserved distinction |
|---|---|
| observation to fact | observed world vs prompt memory |
| fact to context | durable truth vs bounded model context |
| context to work | attention vs lifecycle-managed task |
| work to evaluation | inference vs mutation |
| evaluation to decision | judgment vs permission |
| decision to intent | approval vs effect request |
| intent to execution | authorized purpose vs mechanical action |
| execution to reconciliation | attempted effect vs confirmed effect |
| reconciliation to observation | durable state vs inspectable view |
The pattern is present when these boundaries remain durable, reviewable, and separately owned.
4.0 Collapse Boundary
Separation preserves decision-relevant distinctions across execution.
Collapse erases them.
The anti-collapse chain is:
evaluation
-> evidence
-> authority
-> intent
-> effect attempt
-> confirmation
Each arrow is a governed boundary. If any boundary disappears, an upstream function starts owning a downstream consequence.
Common collapses:
| Collapse | Meaning |
|---|---|
| Cognitive-action collapse | Judgment becomes action without mediated authority |
| Judgment-permission collapse | Evaluation is treated as authorization |
| Permission-effect collapse | Approval immediately performs the effect without durable intent |
| Attempt-confirmation collapse | API success or worker completion is treated as world confirmation |
| Trace-authority collapse | Observability records become control authority |
The pattern does not deny that one implementation may host multiple functions. It requires that the functions remain distinct in durable state, authority, evidence, and review.
Draft-first delivery is the mailbox-specific expression of this anti-collapse rule. Intelligence may produce or recommend text; sending requires a separate governed effect path and later confirmation.
4.1 Stratified Boundary Geometry
The first geometric form is stratification. Each stratum has a distinct authority function. The separation is preserved when information may pass between strata only through the named boundary transformations.
The diagram is not a stack of implementation modules. It is a stack of authority boundaries. An evaluator can produce Evaluation; it cannot skip downward to Intent, Execution, or Reconciliation.
4.2 Control Loop Geometry
The second geometric form is a control loop. Intelligence participates inside the controller, but sensing, authority, actuation, and confirmation remain separate functions.
The control loop representation shows why API success is insufficient as final truth. The actuator changes or attempts to change the world; reconciliation observes the world again and updates durable facts or confirmations.
5. Admissible Intelligence Contribution
An intelligent contribution v is admissible only if:
context(v) is bounded
authority(v) is separately governed
output(v) is persisted as evidence
proposals(v) are typed
policy(v) is external to v
effects(v) require durable intent
confirmation(v) is external to v
The separation fails when the intelligent process directly mutates durable truth, grants permission, executes effects, or confirms its own success.
6. Authority Loci
The architecture preserves separate loci of authority.
| Locus | Function |
|---|---|
| Truth authority | Determines what is durably known |
| Work authority | Determines what operational units exist |
| Evaluation authority | Produces judgment under bounded context |
| Policy authority | Determines allowed action, posture, review, and rejection |
| Intent authority | Converts admitted decisions into durable effect requests |
| Execution authority | Performs effects without inventing purpose |
| Confirmation authority | Reconciles the outside world after execution |
| Observation authority | Presents inspectable state without becoming control |
A concrete system may combine implementations. The loci remain conceptually distinct.
7. Relation To Institutions
Institutions let individual judgment contribute to collective action through offices, records, procedures, review, authorization, and appeal. Intelligence-Authority Separation is the computational form of the same structure.
The intelligent process writes the memo, proposes the draft, classifies the case, or recommends the action. The institution opens the case, records the memo, checks authority, creates the approved instrument, performs the effect, verifies the result, and keeps the file.
7.1 Organizational Transfer
In organizational design, Intelligence-Authority Separation is the structural-differentiation branch of Progressive De-Arbitrarization applied to judgment-bearing roles.
The full transfer is developed separately in IAS Organizational Transfer.
Its core chain is:
analysis
-> recommendation
-> authorization
-> instrument
-> execution
-> audit
This is the form often hidden inside complaints about “roles and responsibilities.” The surface complaint names people and boxes. The deeper failure is usually collapse in the transition from knowledge to authorized consequence.
7.2 Relation To Separation of Duties
IAS overlaps with separation of duties, maker-checker workflows, four-eyes review, audit trails, delegated authority matrices, and internal controls.
It is narrower and sharper in one respect: it names the high-risk seam introduced by judgment-producing intelligence.
Separation of duties asks generally:
Which powers should not be held by the same unchecked locus?
IAS asks specifically:
How does judgment enter consequence without becoming consequence?
This matters for AI systems because model output is fluent enough to look like action policy, but unstable enough that treating it as authority collapses accountability.
7.3 Organizational Examples
| Domain | Judgment | Authority | Execution | Confirmation |
|---|---|---|---|---|
| Trading | trader or model identifies opportunity | risk limits / compliance / mandate | order execution | settlement, P&L, risk report |
| Medicine | physician diagnoses and prescribes | prescribing authority + pharmacy checks | medication administration | charting, labs, follow-up |
| Software | engineer proposes change | reviewer / policy / CI gate | deploy system | telemetry, incident review |
| Sales | sales proposes terms | legal / finance approval | contract execution / delivery | accounting, renewal, customer outcome |
| Support | agent diagnoses customer issue | refund/send/escalation policy | refund, reply, or escalation | audit trail, customer response, reconciliation |
The point is not maximum separation. The point is preserved distinction. A small organization may combine roles in one person, but the record must still show which function was being exercised.
7.4 Organizational Closure
IAS inherits closure persistence from PDA.
An authority decision remains binding until explicitly reopened. A later local actor may not silently convert prior judgment, exception, or temporary workaround into new policy.
An IAS-preserving organization therefore records:
- what judgment was offered
- what authority admitted or rejected
- what instrument was created
- what execution occurred
- what confirmation showed
- what would reopen the decision
This is how judgment can decentralize without consequence becoming arbitrary.
8. Relation To USC
USC governs construction work by preserving intent, arbitrariness, task structure, execution, evidence, and review.
Intelligence-Authority Separation governs operational work by preserving the boundary between judgment and consequence.
The relation is:
USC: how systems are constructed under admissible transformation.
IAS: how intelligence participates in operations under separate authority.
A USC can construct systems that instantiate IAS. A system such as Narada can also use USC-like construction discipline to evolve its own authority boundaries.
9. Relation To Constructive Universalization
Constructive Universalization by Re-instantiation lifts concrete mechanisms into generic substrates only when the originating case can run through the lifted form.
Intelligence-Authority Separation is a candidate invariant extracted from systems where AI is useful and consequence must remain governed.
The originating pressure is practical:
- models can reason usefully
- model outputs are unstable
- real operations require accountability
- direct execution by model output obscures responsibility
- durable systems require replay, review, and reconciliation
Narada re-instantiates the invariant in the mailbox vertical: an AI support steward may draft and recommend, but drafts, sends, confirmations, retries, and audit records are owned by the control plane.
10. Guarantees
Theorem 1 - Mediated Intelligence
An intelligent evaluator creates operational consequence only after its output is admitted by a separate authority locus.
Theorem 2 - Consequence Mediation
Every real-world effect must pass through a durable intent or equivalent governed effect boundary before execution.
Theorem 3 - Confirmation Independence
An effect is not confirmed by the evaluator that proposed it or by the worker that attempted it. Confirmation requires reconciliation against an authority-bearing observation of the outside world.
Theorem 4 - Recoverable Judgment
Because evaluations are persisted as evidence rather than applied as hidden state, the system can inspect, reject, replay, or reinterpret them without losing the operational record.
Theorem 5 - Authority Preservation
Removing the intelligent evaluator degrades capability while preserving truth, permission, lifecycle, and audit structure.
11. Failure Modes
Intelligence-Authority Separation fails or degrades when judgment and authority lose their boundary.
| Failure mode | Description |
|---|---|
| Prompt authority | Model output directly mutates durable state |
| Sovereign inference | Model judgment is treated as sufficient authority |
| Hidden permission | The evaluator determines permission or lifecycle implicitly |
| Direct effect | The model or runner performs side effects without durable intent |
| Self-confirmation | The same component that proposes or executes also declares success |
| Context smearing | Unbounded operational state is treated as informal prompt material |
| Evidence loss | Judgment is used but not persisted as inspectable evaluation |
| Policy absorption | Operational policy lives inside prompt wording rather than external authority |
| Worker discretion | Mechanical executors invent reasons or alter purpose |
| Observation-control merge | Read surfaces acquire mutation power |
12. Scope and Limits
The pattern still requires capable intelligence. It controls how intelligence enters consequence.
Its limits include:
- poor facts or incomplete observations
- missing policy for real decisions
- insufficient review predicates
- effectors that cannot expose confirmation evidence
- environments whose state cannot be reconciled
- institutions unwilling to preserve separate authority loci
A system can satisfy the separation structurally and still make bad decisions if its facts, policies, or evaluators are bad. The separation provides accountability and recoverability rather than omniscience.
13. Summary
Intelligence-Authority Separation is the architecture of mediated operational judgment.
Its primitive act is to store judgment as evidence. Its central transformation is governed admission of that judgment into decision, intent, execution, and confirmation. Its invariant is that consequence belongs to durable authority rather than inference.
In shortest form:
Judgment contributes.
Authority admits.
Intent mediates.
Execution acts.
Reconciliation confirms.