Security for sovereign cognition
Indwel keeps enterprise authority intact while AI reasons across data, systems, people, and time. The security object is the undertaking itself: who is acting, what may count as Evidence, what cognition may infer, what capabilities may be exercised, what Effects may be accepted, and what the institution is finally prepared to settle.
Security begins with the undertaking.
Enterprise AI does not remain inside a model call. It reaches repositories and systems of record, retrieves confidential material, combines sources, calls tools, waits on people, resumes later, and may eventually change something outside itself. A model can behave perfectly and the undertaking can still be insecure.
The failure may be a credible answer built from the wrong source; a safe document used outside the purpose for which it is authorized; a valid credential exposed to the wrong Work; a technically available tool invoked without business authority; a successful API request mistaken for the intended outcome; or a retry that changes provider, cost, or destination after the original authority has expired.
Indwel therefore keeps the undertaking—not the prompt, model, agent, gateway, or workflow step—as the enduring security object. Purpose, principal, Evidence, Cognitive Contract, inference, capability, Effect, Settlement, and continuity remain attributable to the same Work even as the technologies participating in it change.
Existing controls secure parts of the act.
The enterprise security stack has become more capable as AI has entered everyday operations. Guardrails and AI firewalls inspect prompts and responses for content risk, injection, sensitive information, and grounding; agent gateways add identity and tool policy; and copilots inherit familiar permissions and data-loss protections. Trust, orchestration, and governance platforms contribute their own strengths, from masking and monitoring to coordinating participants and maintaining risk inventories. The remaining challenge is to make those protections coherent across the business undertaking as a whole.
Consider a vendor-risk review. Every familiar control can succeed: the user is authenticated; the repository is permitted; the prompt is safe; the answer is grounded; the connector call is authorized; the workflow completes; the interaction is logged. The institution can still be wrong if the policy was superseded, the source had no standing for the proposition, the model turned a finding into a determination it did not own, the external update was merely accepted rather than effected, or the record was closed before those conditions were settled.
Indwel does not displace those controls. It governs the handoffs between them as one institutional act. External authority therefore continues across Evidence standing, inference and venue admission, capability, human judgment, time and budget, observed Effect, continuity, and Settlement.
Inspect model inputs and outputs, content safety, sensitive data, and grounding. By themselves they do not establish whether retrieved material had authority for the proposition at issue.
Authorizes who or what may invoke a tool under declared conditions. By itself it does not establish Evidence standing, reconcile an external Effect, or create Settlement.
Coordinates steps, systems, agents, and people. The authority and proof of the cognitive act moving through those steps remain a separate problem.
Inventory, evaluate, monitor, and audit AI systems. Indwel also preserves the governed record of the undertaking itself: what governed it, what it relied on, what it attempted, what happened, and what settled.
Evidence and inference
Authority in execution
Cognitive Contracts resolve organizational, team, Work, source, and turn-specific authority into executable conditions before and during cognition. Cognitive Governors enforce those resolved conditions outside the model. They can permit, reject, qualify, redirect, abstain, require more Evidence, or route the Work to a competent person or office without becoming another probabilistic decision-maker or minting authority of their own.
Where a condition is determinable, Indwel can enforce it directly. Where the institution has reserved judgment, interrogatory authority lets the undertaking ask the person or office that owns the decision, preserve the question and answer under that authority, wait if necessary, and resume the same Work. Human judgment is therefore an addressable office inside the act, not a generic “human in the loop” interruption after the model becomes uncertain.
In benefits administration, for example, a plan document, amendment, employee election, carrier file, and administrative policy may all contain relevant facts without having equal authority. The Work can establish which source governs which proposition as of which date, what a model may infer from those facts, which operation may be attempted, which person owns an exception, and what must be observed before the result can be settled.
What undertaking is being performed, for whom, under which organization, role, obligation, and Cognitive Contract.
What may be reached, handled, retained, exported, and relied upon for the propositions the Work must establish.
Which cognitive roles, models, providers, venues, context, independence requirements, and fallback paths are permitted.
Which tools, agents, connectors, operations, targets, and credential references may be exposed to this Work.
Which consequential actions may be attempted and what post-condition must be observed before their result can be accepted.
What the institution is finally prepared to carry forward as the terminal truth of the undertaking, with the receipts that support it.
Across all six stages, principal, Cognitive Contract, Governors, time and budget, custody, and receipts continue to govern. A clock can make Work eligible to resume but cannot mint new authority. A cheaper model can become available but cannot become admissible by price alone. A successful call can report transport success without establishing that the intended Effect occurred.
Enterprise security
Indwel works with existing identity and access management, DLP, content inspection, malware protection, key and secret management, SIEM, network controls, vulnerability management, retention, backup, and incident response. Rather than displacing these investments, the cognitive architecture gives them a place to govern the undertaking: DLP and classification may restrict provider egress or model venue, tenant identity may constrain available tools, and quarantined material can be stopped before Evidence admission. Relevant security events can reach the security operations center without turning the SIEM into an accidental repository for prompts or sensitive Evidence.
The AWS reference implementation applies that model with Cognito/JWT identity, IAM-scoped API and Lambda execution, KMS and Secrets Manager custody, S3 landing/quarantine/vault boundaries, GuardDuty Malware Protection for S3, OpenSearch Serverless and Aurora PostgreSQL/pgvector retrieval, governed Bedrock inference routes, EventBridge clocks, encrypted SQS dead-letter handling, and CloudWatch, CloudTrail, GuardDuty, Security Hub, and Config telemetry. These services implement the boundary; they do not define its authority.
Indwel architectureAuthority stays enforceable through the cognitive act.
Security is not simply an inspection before an answer is returned. Organizational conditions determine what information may become evidence and what the cognitive runtime may do with it.
One governed undertaking
Indwel begins with the institutional undertaking in which models, agents, gateways, processes, and systems all participate. That perspective preserves distinctions easy to lose in conventional application stacks: information must be admitted before it can serve as Evidence, model inference cannot grant institutional authority, technical access is not permission to act, and an attempted operation is not yet an observed business Effect. Even a confirmed Effect must satisfy the Work’s completion conditions before Settlement can follow.
Settlement is therefore a security primitive. It governs what the institution is finally allowed to carry forward as having happened, preventing a model output from becoming a decision, an attempted action from becoming an accomplished Effect, or an indeterminate result from becoming institutional truth.
Runtime enforcement
The business does not have to trust that the model remembered the policy, that a developer reproduced it in every agent, or that a reviewer reconstructs the act from logs afterwards. Authority is resolved into execution and remains inspectable across people, models, tools, systems, failures, and time.
That is Indwel’s technological distinction: powerful probabilistic cognition can participate in consequential business work without becoming sovereign over the institution that employs it.
Assurance
Architecture is only the claim. Compliance exposes the control universe, framework mappings, evidence, diligence material, and assurance state by which the claim can be examined. Design, implementation, testing, operating effectiveness, and independent assurance remain separate facts.
Deployment
Indwel Cloud is the managed offering, now in private beta. Dedicated and Licensed deployments provide separately scoped alternatives for different operating requirements. Each preserves the institution’s authority over its Work as infrastructure, model providers, and integration boundaries change.