Intelligence isnot authority.
Consequential Work rarely offers certainty. Models can estimate risk, compare possibilities, and recommend a course with extraordinary skill. None of that confers jurisdiction.
Ordinary institutional distinctions still apply to machine intelligence.
Organizations already separate advice from office. Indwel preserves the distinction when the adviser is a model. The question is simple: does the model inform the institution, or stand in for it?
A single model can interpret a request, reason across the record, choose a plan, act through tools, and write the account of what happened. One interface can compress those acts. It does not merge their offices. Without an external authority boundary, the answer can quietly choose the question.
Probability and authority remain separate states.
Illustrative confidence
P(correct | E)=0.97
An illustrative high-confidence estimate can support judgment without creating jurisdiction.
Both can be true. Correctness and jurisdiction are independent.
Indwel represents them as separate states. A proposition about the Work can carry high confidence and remain non-binding. Inferred intent remains a candidate until an authorized source or office gives it standing.
A prompt can express or clarify the Work’s intended outcome; it cannot authenticate its own power to amend the Work Definition. Every material amendment has an author, an authority, a reason, a scope, and an effective time. Conflicting instructions are resolved by authority, not averaged into an inferred preference.
Probability receives its proper office after the Work’s purpose and jurisdiction exist. Otherwise, the system can optimize toward a purpose it was never entitled to choose. Objective before probability.
Model failure can be contained before institutional state changes.
The same separation contains failure. A model can be wrong—or compromised—without rewriting institutional state.
“The customer approved the refund.”
A generated proposition exists.
Approval unchanged.
No admitted Evidence or authorized decision has changed institutional state.
Behavioral alignment shapes what a model tends to say and do. Constitutional authority determines what the institution recognizes as binding. Inference may propose; authority decides whether institutional state changes.
The error remains a cognitive contribution. It has not appointed itself to an office.
Better inference improves cognition without manufacturing authority.
Better cognitive faculty✓
More authority×
A stronger model may deserve more trust as an adviser. It still does not inherit the right to waive an obligation or declare the Work complete.
Because authority lives outside the model, providers, models, tools, and execution venues can change without moving the mandate. The Work Definition and open burdens remain durable; Evidence retains its provenance and admission status; Memory may inform context without becoming authority; and decision rights remain with the responsible institution. Effect history, Settlement conditions, continuity, and reopening remain attributable to the Work. The model is a replaceable faculty; the institution can change its intelligence without changing what the Work means.
Work Definition and Cognitive Contracts make the authority boundary executable.
For the enterprise, the consequence is practical: better models can be adopted without renegotiating what the Work means. Indwel places probabilistic intelligence inside deterministic authority. The Work Definition states what the undertaking is for and what completion means. Cognitive Contracts govern what may count, who may act, and which Effects may cross into the world. Verification, Settlement, continuity, and reopening remain explicit.
Prediction can inform the institution without becoming the institution.