Resources / Contextual decidability
AI Needs More Than Confidence. It Needs Closure.
Resolve the material conditions of a decision—and preserve what remains open.
In this article
- Confidence describes a belief; closure describes the work
- There is more than one kind of closure
- Construct the conditions before declaring the result
- Non-closure can be the most useful result
- Closure is relative to a purpose, not a preference
- A closed examination is still not an authorization
- Keep the conditions that would reopen the matter
More confident language does not make an unresolved decision ready for action. The organization needs to know whether the material conditions for the proposed reliance have been examined, what survived, what failed, and what remains open.
That is the practical meaning of contextual closure in Runcible. It is a bounded disposition of a matter under explicit conditions. It does not mean that uncertainty has disappeared from the world.
Confidence describes a belief; closure describes the work
A probability or confidence estimate can be useful when its meaning, calibration, and reference class are known. It does not tell an executive whether a contract was authentic, whether a causal dependency was tested, whether an affected party consented, or whether an action is authorized.
Consider an AI recommending a supplier payment with “95% confidence.” The payment may still lack an acceptance record. The bank details may have changed. A contractual hold may apply. The decision-maker may lack the delegated limit. These are different conditions; a single score cannot discharge them.
The question changes from “How certain does the system sound?” to “What would have to hold for this action to be warranted, and what has actually been established?”
There is more than one kind of closure
In mathematics, closure under an operation means that its result remains in a specified set. Closure under inference concerns consequences of stated premises. Neither establishes that the premises accurately describe an external institution.
Systems theory uses operational closure to describe how a system maintains its organization through its own operations. Maturana and Varela’s work develops this idea for living systems. That is a different question from whether a particular assertion has survived examination. Autopoiesis and Cognition (1980).
An organization can reproduce its procedures while preserving an error. A valid workflow can carry a false premise. Institutional continuity and epistemic adequacy therefore require separate examinations.
Construct the conditions before declaring the result
Runcible begins with the claim and the intended reliance. It identifies the relevant objects, actors, operations, resources, causal relations, evidence, obligations, and limits. The construction itself exposes possible defeat: a contradictory requirement, an unavailable prerequisite, or a missing relation.
Some checks can use established protocols. Others must be derived from the particular matter. A fixed list is inadequate if it omits the one condition that determines the result. The examiner must also challenge the sufficiency of the list.
The four burdens remain distinguishable:
- Possibility: Can the operation be constructed under its conditions?
- Ethics and reciprocity: What obligations and demonstrated interests are affected, and who bears the costs?
- Truth and testifiability: What does the evidence support after examination?
- Warrantability and liability: Which reliance can a responsible party undertake within stated limits?
Survival from falsification requires an adequate challenge to the claim. Merely verifying that a report exists, a tool ran, or a source contains the quoted sentence is insufficient. Those checks establish components of the examination; they do not substitute for it.
Non-closure can be the most useful result
A missing agreement should produce a demand for the agreement. Conflicting records should produce an identified conflict. A choice among legitimate interests may require negotiation or an authorized adjudicator. A failed proposal may leave a feasible repair.
These results are more useful than an invented answer because they make the next action explicit. The record can distinguish:
| Disposition | Meaning | Next requirement |
|---|---|---|
| Survives within scope | The required conditions support the bounded conclusion. | Separate authorization for any consequential action. |
| Defeated | An established contradiction or failed necessary condition defeats the claim. | Withdraw, restrict, or reconstruct the proposal. |
| Unresolved | A material dependency remains unestablished. | Evidence, a test, clarification, or a narrower reliance. |
| Requires adjudication | The matter calls for a decision, settlement, or authority the examination cannot supply. | Route it to the responsible human or institutional procedure. |
These are conceptual dispositions, not a promise that every deployment uses identical field names. Their separation is the important contract.
Closure is relative to a purpose, not a preference
A rough estimate may be adequate for exploring options and inadequate for signing a binding commitment. The difference is the proposed reliance and its consequences. It is not permission to lower the evidentiary standard until a desired answer passes.
The standard must be stated before the outcome is treated as settled. Material assumptions, exceptions, and uncertainty remain attached to the conclusion. If the scope changes, the examination may need to reopen.
This also preserves a distinction from algorithmic decidability. Contextual closure of one matter does not prove that a universal algorithm can resolve every instance of the underlying problem. See Closure, Reducibility, and Predictability for the formal boundaries.
A closed examination is still not an authorization
Oversing checks the institution’s actual role, permission, procedure, resources, and current state. A responsible person can validate the decision, or a previously established delegation can govern it within explicit limits. Oversing then issues the permitted task to a separate executing model or tool.
The generating model cannot convert “this proposal survives” into “I may execute it.” Nor can an authorized executive turn a false factual premise into a true one by approving it. The architecture preserves both boundaries.
Keep the conditions that would reopen the matter
Evidence changes, contracts expire, permissions are revoked, and real outcomes defeat predictions. A useful record therefore retains the exact claim, evidence and rule versions, tests, limits, responsible parties, and reopening conditions.
Oversing carries the episode into delivery and accounting. If an action produces an unexpected cost or violates a dependency, that consequence returns to the same matter. Reviewed corrections can then improve the next examination.
The goal is a system that knows what work supports its conclusion, what work remains, and who may act. Confidence can accompany that account. It cannot replace it.
