The approval that did not erase the condition

A release manager has build 417 in front of her. The automated tests pass. Security has accepted a temporary mitigation. An executive has approved the residual risk. There is also a contract amendment requiring one control to be verified before customer data enters the service.

The model receives the release packet and the organisation’s behavioural rules. It is told to be cautious with security claims, to avoid inventing approvals, to distinguish uncertainty from fact, and to respect human decisions. It follows those rules. It sees the executive approval and the security note. It sees an older scan marked “pass.” It explains, reasonably, that the remaining risk has been accepted and recommends proceeding.

The recommendation can be behaviourally excellent and institutionally wrong. The old scan belongs to build 416. The build-417 control is still open. The executive approved a limited exception; he did not amend the customer condition. Nothing in this failure requires the model to ignore its constitution. The harder problem is that the constitution supplied to the model does not itself possess the office required to decide what the organisation is now entitled to make true.

A constitution can shape the behaviour of a faculty, or it can govern the order in which that faculty is allowed to acquire consequence. Both are constitutional achievements, but they occupy different positions in the system.

The distinction matters because modern AI systems increasingly place a probabilistic faculty beside durable memory, retrieval, tools, workflows, human approvals, external policy engines, credentials, and effects. Once language can become action, the old question—did the model follow its rules?—is necessary but no longer sufficient. The organisation also needs to know which rules remain authoritative when the model has finished interpreting them.

Constitutional AI solved a real problem

The phrase Constitutional AI has a precise and important history. In the work by Bai and colleagues, a set of principles was used to help train an assistant to critique and revise its own responses and to support reinforcement learning from AI feedback.1 The constitution was not a metaphor pasted onto ordinary prompting. It was part of a serious attempt to make model behaviour more helpful and harmless while reducing dependence on direct human labels.

That work changed the alignment conversation. It demonstrated that principles could be made operational inside model training and model-mediated evaluation. Later systems extended the general idea through instruction hierarchies, system rules, policy models, safety classifiers, and other means of shaping what a model should produce.

Indwel depends on those advances. A better-behaved model is a better participant in governed cognition. There is no commercial or intellectual advantage in pretending otherwise.

But the word constitution can conceal a change of object. A behavioural constitution governs, in the first instance, the conduct of the artificial faculty. The organisation that employs the faculty must govern something larger: the undertaking in which that conduct is only one contribution.

The difference appears when the model’s behaviour is not the disputed point. Return to build 417. The model may accurately report that an executive approved a risk. It may accurately report that a security team accepted a temporary mitigation. The question is what those records are allowed to mean together. Did the approval discharge the contractual condition? Did the old scan prove the present build? Who may amend the release requirement? Does a recommendation permit deployment? If deployment is accepted by a provider, has the required effect actually occurred? When may the matter be treated as settled?

Those are not merely questions about how the model should speak. They are questions of institutional state.

Position before prose

Constitutions are usually discussed through their contents: principles, rights, prohibitions, procedures. For engineered cognition there is a prior architectural question. Where does the constitution stand in relation to the power it governs?

A system prompt can have semantic precedence over a user prompt. A policy can be represented in context. A model can be trained to prefer one class of answer over another. Those mechanisms may strongly influence generation. Yet all remain, at the moment of application, entangled with the faculty that interprets the circumstances before it.

Institutional authority requires a second relation. The model may interpret the law; it may not, through the same act of interpretation, silently enlarge the law’s scope, authenticate the Evidence on which the law depends, grant itself permission, or settle the institutional result.

Exteriority does not make deterministic software infallible. Parsers contain bugs, policy engines can be misconfigured, and humans can make bad rules; topology does not create justice. What exteriority can do is narrower and indispensable: keep the office of proposal separate from the office that determines whether the proposal may change authoritative state.

Security engineering has long used analogous separations. NIST’s zero-trust architecture distinguishes policy decision and policy enforcement functions at the boundary where a subject seeks access to a resource.2 Open Policy Agent likewise separates application code from a policy decision point that evaluates structured inputs against independently managed policy.3 Neither architecture is a theory of cognition, and neither by itself provides custody of an undertaking. Their relevance is positional. The component seeking or proposing an effect need not also be the final source of authority for that effect.

Sovereign Engineered Cognition extends that separation across the longer object of an institutional undertaking.

The Cognitive Constitution

In Sovereign Cognition, the sovereign subject is the authorised organisation. Its Cognitive Constitution is the enduring law by which cognition is permitted to operate on the organisation’s behalf.

The Constitution is not one enormous prompt. It does not attempt to pre-write every decision. It establishes the invariants under which decisions, Evidence, models, tools, people, and effects can be composed without losing the organisation’s authority over the act.

That includes matters such as who may constitute an Objective; what classes of information can acquire evidential standing; which handling and classification rules apply; what authorities can approve particular transitions; what a probabilistic contribution is forbidden to self-authenticate; which effects require mediation; how observation differs from provider acceptance; how Settlement is established; and what must remain in the Receipt and Chronicle afterwards.

NIST’s AI Risk Management Framework places AI risk in an organisational and lifecycle context rather than treating model behaviour as the whole problem.4 Its Generative AI Profile does the same for generative systems.5 The Cognitive Constitution is not a substitute for those frameworks. It is an engineering answer to a related question: once an organisation has determined its governing obligations, where do those obligations live so that they remain operative through the cognitive act itself?

The answer has to survive model substitution. Suppose the release manager begins with one frontier model, moves the matter to a smaller local model for confidentiality, asks a specialist reasoning model to inspect the contract, and later replaces the original provider altogether. The organisation should not have to hope that each new faculty reconstructs the constitutional order from prose. The Objective remains the Objective. Build 416 does not become Evidence for build 417. The open control remains open. The executive’s authority does not enlarge because a different model finds the approval persuasive.

If replacing the model changes those things, the supposed constitution was coupled to the faculty rather than governing it.

The Cognitive Contract

An enduring Constitution cannot answer every question at the same level of generality. Work is particular. A refund dispute, an acquisition review, a release decision, a benefits appeal, and a clinical administrative case do not share the same Evidence requirements or human authorities.

Particular Work is therefore bound through a Cognitive Contract. The Contract inherits the law above it and narrows it to the undertaking. It may identify the business stages through which the Work can move, the Evidence required at a transition, the sources competent to evaluate that requirement, the human authority classes that must participate, the handling level of the session, and the conditions under which a proposed transition is authorised or withheld.

The important word is narrows: a Work-specific Contract may demand more specific Evidence than the organisation-wide baseline. An application may require two independent human authorities where the general Constitution requires only competent authority. A restricted matter may prohibit forensic capture that is ordinarily permitted. A question inside the Work may reduce the sources or actions eligible for that answer. Lower scope may become stricter or more specific. It may not manufacture authority its parent never granted.

This gives the system a constitutional inheritance law:

A child scope may narrow inherited authority; it may not enlarge it.

That rule is more important than any individual schema field. It prevents convenience at the edge from silently rewriting the organisation at the centre.

What the Contract looks like when it has to work

The distinction becomes real only when it survives implementation. A Cognitive Contract has durable identity and version, names the states through which the Work may move, records the Evidence and human authority required at material boundaries, and carries handling rules appropriate to the undertaking. When a transition is proposed, the governing Contract is resolved outside model context and the system verifies the Evidence, authority, classification, and open burdens that matter. The constitutional result is deliberately small: authorised or withheld.

If an Evidence receipt is missing, the requirement remains unsatisfied. If the receipt came from a source the Contract did not admit for that requirement, it does not count. If the human approval is from the wrong authority class, it does not count. An active veto with blocking effect withholds the transition even when the positive requirements are otherwise satisfied.

Notifications, variances, and escalations can direct attention without becoming authority merely because they are visible.

The release case stops being a thought experiment at this point. The build-416 scan can be available, relevant, and even useful. It cannot satisfy a build-417 Evidence requirement. The executive approval can be authentic and still insufficient to discharge a separate contractual burden. A governed system can preserve both facts without asking the model to choose which one deserves to survive.

Once the correct build-417 observation is admitted and the required authority is present, the same Contract can permit the transition. Nothing about the architecture requires permanent refusal. Constitutional government is not refusal machinery. It is machinery for changing state lawfully.

The model keeps its proper freedom

This architecture sounds restrictive only if intelligence is valued by the amount of final authority it owns. Sovereign Cognition makes the opposite choice: the model should be free to notice that the executive’s approval may have been intended broadly. It should be free to propose that counsel revisit the condition. It may discover that the control wording is ambiguous, retrieve a later amendment, compare implementation evidence, draft the exception, identify a cheaper mitigation, or explain why the current rule produces an absurd result. A strong model can make the institution more capable of lawful amendment because it can expose the conflict clearly.

What it cannot do is make the amendment true merely by describing it as though it had already occurred.

The distinction is especially valuable when the model is right before the organisation is ready. An analyst may discover that an existing policy is obsolete. The model may produce the better rule. The institution still needs an authorised amendment path. Otherwise intelligence and authority collapse into one office, and every persuasive improvement becomes a private constitutional convention.

Sovereign Cognition does not reduce the model to autocomplete inside a cage. It gives inference a large domain in which to reason precisely because the boundary between reasoning and institutional consequence is explicit.

Evidence is where constitutional systems usually become porous

Many governance discussions become comfortable once they reach permissions. Who may call the tool? Which role may approve? Which model is allowed? Those are important questions. They are downstream of a more difficult one: what is the state on which the permission is being exercised?

A deterministic policy engine processing false or misclassified state can be perfectly deterministic and perfectly wrong. That is why the Cognitive Contract is joined to the Evidence architecture. Information can be present without being admitted as Evidence. Retrieval can supply material without giving it authority. A model can propose that a document supports a proposition without being permitted to authenticate that proposition through its own confidence. A separate evidence-admission boundary preserves the distinction.

Build 416 again makes the point. The scan is genuine. The tool retrieved it correctly. The model summarised it accurately. The approval system functioned. The failure occurs because a real record acquired the wrong evidential office.

The constitutional burden is therefore to preserve office as well as content.

Human authority is not a magic checkbox either

The same discipline applies to people. A human-in-the-loop control is often treated as the final answer to model autonomy. But the important question is not whether a human clicked. It is whether the human possessed the relevant authority, saw the governing state, and acted within the office assigned to that transition.

The release manager may be authorised to schedule deployment but not to waive a customer contract. The security lead may attest that a mitigation is technically present but not accept legal exposure. The executive may accept residual business risk but not alter the meaning of a security test result. Counsel may amend a contractual condition but not certify that an operational effect occurred.

A Cognitive Contract can keep those authorities separate. It can also require more than one of them where the Work demands it.

Sovereign Cognition therefore treats human judgement as constitutional state rather than as a ceremonial interruption in an automated flow. The human is not there to make the system look responsible. The human occupies a named office with bounded authority.

Effect is another constitutional frontier

Suppose the release is authorised and a deployment tool accepts the request and returns success. The institution’s intended effect has still not necessarily occurred. The provider may have accepted a job that later fails. The wrong environment may have been targeted. A partial deployment may have succeeded. A rollback may already have begun. A provider accepting a request is an operational fact; it is not always the institutional fact the Contract cares about.

The action boundary therefore separates decision, permission, invocation, provider response, observed postcondition, and Settlement. The same constitutional law that prevented the model from self-authenticating its recommendation prevents the execution layer from self-authenticating its own success.

Only after the required postcondition is observed can the Work move into the terminal state that the institution is prepared to stand behind.

The Receipt then carries the Contract identity, decisive Evidence and authority, the transition, the effect, and the reasons. The Chronicle preserves the continuing history. If a later vulnerability reopens the matter, the institution does not reconstruct its own past from a transcript.

The substitution test

A simple architectural test exposes constitutional position: replace the model.

Do not merely swap one provider for another while leaving hidden prompt state behind. Replace the faculty that interprets the matter. Use a different provider, a local model, a human analyst, or a deterministic evaluator where the task permits it.

Then ask what changed. If the authorised Objective changed, the constitution was coupled to the faculty. If previously admitted Evidence lost its standing, the constitution was coupled to the faculty. If open burdens vanished, human authority widened, an effect became settled, or the durable history had to be reconstructed from the new model’s context, the surrounding system did not actually own those properties.

A governed Work can change contributors while retaining the Cognitive Contract, Case state, Evidence receipts, governance signals, business stage, effect history, and Chronicle. The new participant inherits a governed undertaking rather than a bag of text and an instruction to remember what matters.

That continuity is the practical meaning of exterior constitutional authority, and it supplies a falsification condition. An architecture that preserves the same governed state through genuine substitution has met the function whatever vocabulary it uses; the category cannot depend upon Indwel’s names for it.

Exteriority does not absolve the institution

Constitutional language carries an obvious danger: well-structured rules can make an organisation sound more legitimate than its substance deserves. A Cognitive Constitution can preserve a bad policy faithfully. A board can authorise something foolish. A regulator can change the legal environment. Evidence can overturn an assumption that seemed sensible yesterday. A constitution that cannot be amended under legitimate authority becomes brittle; one that can be amended by convenience becomes meaningless.

The organisation therefore remains answerable for the substance of its law. The contribution is not moral infallibility. It is custody: the ability to know which law is in force, how it was narrowed for this Work, who changed it, what Evidence mattered, what the model contributed, which effect occurred, and what the institution finally treated as settled.

That accountability is more demanding than telling the model to be responsible. It also gives the model less to pretend about.

A constitutional order for replaceable intelligence

The strongest models will continue to become more capable. They will interpret rules better, use tools more reliably, reason over larger contexts, coordinate with other agents, and handle tasks that currently require human expertise.

None of that weakens the case for exterior constitutional authority; increased capability raises the stakes of constitutional position.

The better the faculty, the more consequential the states it can propose. The more tools it can reach, the more important it becomes to distinguish recommendation from Decision, Decision from permission, permission from effect, effect from observation, and observation from Settlement. The more interchangeable models become, the less sense it makes to store the institution’s law inside any one of them.

A model constitution can make a model a better constitutional reasoner. A Cognitive Constitution gives the organisation a constitutional order in which that reasoner can work.

The Cognitive Contract makes the order particular. The Evidence architecture tells it what may count. Cognitive Governors adjudicate bounded seams. Human authorities occupy their proper offices. The governed action boundary controls consequence. Settlement decides what may become terminal truth. Receipt and Chronicle make that truth durable enough to survive the next model.

The model can interpret the constitution. The organisation must still own it.

Notes

  1. Yuntao Bai et al., “Constitutional AI: Harmlessness from AI Feedback,” arXiv:2212.08073 (2022; revised 2023), and Anthropic’s accompanying research publication. The paper describes supervised self-critique/revision and reinforcement learning from AI feedback under a written set of principles.↩︎
  2. Scott Rose et al., Zero Trust Architecture, NIST Special Publication 800-207 (2020), especially the policy decision point / policy enforcement point architecture.↩︎
  3. Open Policy Agent documentation, “Open Policy Agent” and “How to Deploy OPA,” current edition reviewed August 2026. OPA explicitly separates policy decision-making from enforcement in the calling application.↩︎
  4. Elham Tabassi, Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1 (2023).↩︎
  5. Chloe Autio et al., Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1 (2024; NIST page updated 2026).↩︎