Models can reason.They cannot govern.

Indwel treats probabilistic intelligence as a governed faculty of Work, not an institutional office. Models can interpret information, generate and critique, plan, judge, and operate across modalities without acquiring standing over Evidence, authority, permitted Effects, or Settlement.

Authority remains outside inference.

Conventional AI applications often place context, policy, tool access, and terminal state around the same model call. Probabilistic output can then carry decisions that properly belong to the institution.

Indwel keeps those authorities outside the model. External policy engines that authorize individual agent tool calls contribute a useful part of this protection; Indwel extends it across the undertaking, including Work Definition, Evidence, model and venue admission, budgets, human offices, capabilities, Effects, continuity, and Settlement. More capable models can improve the quality of cognition within those conditions without acquiring the authority to redefine them.

Model contributionStanding retained outside inference

Interpret or classify informationEvidence standing and admissibility

Generate, critique, or synthesizeWork Definition, Cognitive Contract, and governing constraints

Propose a tool or external actionCapability and Effect authority

Report a likely outcomeObservation, reconciliation, and Settlement

Recommend what should happen nextInstitutional and human decision authority

From selection to answer admission

Cognitive MachineryCognitive Executive

selects

Cognitive MachineryFaculty Registry and Governed Faculty Execution

can use

Cognitive MachineryGoverned Inference Boundary

does not supersede

Cognitive MachineryIndependent Critique and Answer Verification

guards

Cognitive MachineryWriter and Answer Admission

Explore governed cognitive machinery

Governed bounds

The Inference Boundary turns changing model and provider capability into governed cognitive capacity. The Cognitive Contract and applicable institutional and platform policy establish the permitted solution space before runtime selection begins. A resolved configuration expresses the eligible candidates; adaptive resolution can choose only within that admitted set.

Set the governing envelope.

Cognitive role, candidate pools, provider and venue admission, data treatment, region, required capabilities, independence, budgets, stopping rules, and fallback authority are resolved before execution.

Resolve eligibility.

Privacy, commercial, residency, capability, operational, independence, and budget constraints remove ineligible candidates before optimization begins.

Select within authority.

The runtime can then optimize for capability fit, cognitive quality, reliability, latency, cost, placement, independence, or a declared combination.

Receipt selection and execution.

Route, provider, model, attempt authority, fallback, usage, cost source, and terminal disposition remain attributable to the governed Run.

Model roles

Indwel composes models around the cognitive role required by the Work. The governed act defines each contributor’s role, the independence required between contributors, the stopping rule, and the treatment of material disagreement.

StrategyGoverned purpose

Specialist sequenceDifferent models serve different stages, each with an explicit input and output boundary.

Independent verificationA second contributor checks a proposition without first inheriting the first conclusion.

Adversarial reviewA critic is deliberately exposed to another result in order to challenge it.

Provider-diverse panelContributions span declared provider or model-family boundaries before synthesis.

Private-first escalationCognition remains inside the private boundary unless an authorized escalation is necessary.

Consensus with dissentConvergence can be summarized while material disagreement remains part of the record.

Human decision gateModel contributions can inform a consequential choice without becoming the terminal authority.

Independence is evaluated across model family, origin provider, execution venue, upstream lineage, shared infrastructure, and exposure to prior conclusions. When the governing conditions require independent verification, unknown lineage can make a candidate ineligible. Agreement among models remains inference; it does not become Evidence merely because several models produced it.

Cognitive Governors

Model composition can improve cognition without making models the institution’s decision offices. Cognitive Governors sit outside probabilistic execution and apply the resolved Cognitive Contract at the boundaries they own. A Governor can refuse a source, require independent cognition, deny an inadmissible provider or venue, stop at a budget ceiling, require a named human authority, prevent fallback from widening permissions, hold Work for an unmet Evidence burden, or refuse Settlement when the observed Effect remains indeterminate.

When the Cognitive Contract reserves judgment to a person or office, interrogatory authority lets the undertaking ask that authority directly, preserve the question and answer as governed state, wait if necessary, and resume the same Work. Human judgment enters at a named decision boundary rather than as an improvised escalation after the model runs out of confidence.

Governed boundaryWhat it controls

Cognitive GovernorWhether a governed transition is permitted under the resolved office, delegation, Evidence burden, and Cognitive Contract in force.

Interrogatory authorityWhich person or institutional office may supply judgment, clarification, approval, or refusal when the Work requires it.

Continuity and timeWhen Work may wait, recur, reopen, or resume without allowing a clock, retry, or scheduler to create new power.

Economics and budgetsHow much inference, acquisition, tool execution, parallel cognition, or human effort may be consumed before the Work must stop, narrow, or escalate.

Inference Firewall

Inference cannot confer standing.

The Inference Firewall keeps probabilistic output in its proper status: inference. Confidence, structure, or agreement among models does not give that output standing as Evidence, institutional authority, Effect permission, or Settlement.

“This source proves the claim.”Does not admit Evidence.

“The approver authorized the action.”Does not create authority.

“Call this system with these arguments.”Does not grant Effect permission.

“The external operation succeeded.”Does not prove the intended Effect occurred.

“The matter is complete.”Does not create Settlement.

The Evidence Firewall establishes what material is entitled to support reasoning, while the Inference Firewall maintains the boundary between a model’s output and an institutionally warranted conclusion. That separation allows the organization to use increasingly capable probabilistic intelligence without transferring its judgment or decision rights to the model.

Models and venues

A model family can be available through several execution venues: directly from its developer, through a cloud model service, through another marketplace, through a private endpoint, or in a local runtime. Indwel keeps model identity, immutable version, execution venue, region, and operational certification distinct.

That separation keeps institutional semantics above the provider layer. Teams can change model, route, provider, or placement without recreating the Work Definition, Evidence obligations, Cognitive Contract, Work history, or Settlement conditions around it.

Frontier providerDirect access where policy, commercial terms, data posture, and resolved configuration permit it.

Cloud model serviceA model may execute through a governed cloud venue such as Amazon Bedrock without becoming equivalent to the same model elsewhere.

Private endpointPlacement can remain inside a controlled network boundary while the same higher-order Work semantics persist.

Local or isolated runtimePrivate or air-gapped inference can remain a faculty of the same governed undertaking when an eligible runtime is available.

Provider substitution can therefore remain an infrastructure decision rather than a change in institutional meaning.

Fallback

An alternate model or provider must already be eligible under the resolved governing conditions. Retries do not reset the Run’s attempt ceiling, and callers may narrow fallback, timeout, or provider access without silently expanding them.

Attempt authority: Physical model invocations are bounded before dispatch and tied to the same governed Run.
Cancellation: Deadlines, timeouts, abort signals, and cancellation remain part of the governed execution boundary.
Fallback: Declared alternatives preserve the same authority boundary; degraded availability does not authorize broader egress or capability.
Receipts: Selected route, provider, model, attempt, usage, cost source, fallback, and disposition remain inspectable without projecting raw sensitive content.

Execution spine

Indwel routes generative, judging, arbitration, retrieval-planning, code, and multimodal cognition through a governed inference façade with explicit cognitive roles, route and venue controls, eligibility and certification, bounded fallback, attempt authority, cancellation, budgets, and receipts. The model catalog and resolved configuration enumerate candidates eligible under governing conditions; adaptive selection can optimize only inside that admitted set.

Authority and scope

The next boundary asks a narrower question: when cognition predicts what is likely, what gives that prediction standing to govern what the institution does?

Prediction & AuthoritySeparate what is likely from what is allowed →