Author: Curt Doolittle, Founder NLI and Runcible.
Date: April 12, 2026
Abstract
Existing AI systems generate candidate speech. They do not close claims. They do not decide whether a statement is true enough, possible enough, authorized enough, reciprocal enough, and warrantable enough for institutional action. My work extends closure from systems theory into epistemology, law, and AI governance. Runcible implements that extension.
My work serves as an expansion in the meaning of closure:
- Classical systems theory: closure preserves the identity of a living, cognitive, or communicative system.
- Luhmannian legal theory: closure preserves law as law through the legal/illegal code.
- Doolittle’s work: closure becomes the condition of decidable action.
- Runcible: closure becomes an institutional AI process: converting unclosed model output into tested, scoped, warrantable, auditable records.
AI cannot become institutionally useful until closure moves from communication to action.
From Generated Text to Closed Action: Why AI Needs a Theory of Closure
Artificial intelligence has produced a revolution in linguistic competence. It can summarize, draft, translate, classify, compare, infer, simulate, and advise. But institutional action requires more than competent speech. Institutions cannot act on fluency. They act on claims, evidence, authority, procedure, liability, and warrant.
This is the unsolved problem in enterprise AI: the model can produce an answer, but the institution must know whether that answer is fit for action.
That distinction is the difference between generation and closure.
In systems theory, closure has a precise history. Maturana and Varela used operational closure to describe living and cognitive systems whose operations are produced by their own internal organization. Luhmann extended the concept to social systems, including law. In his account, the legal system is operationally closed because only legal communication can produce further legal communication. Facts, morality, politics, economics, and social pressure may perturb the system, but they do not directly become law. They must first be translated into the legal code and legal procedure.
This insight is useful because it explains how domains preserve their identity. Law remains law by processing the world legally. Science remains science by processing claims scientifically. Markets remain markets by processing exchange economically. Each system survives by allowing only system-compatible operations to produce further operations.
But this is only the first step.
Operational closure preserves the identity of the system. It does not guarantee the truth of the claim, the reciprocity of the action, the possibility of the prescription, the legitimacy of the authority, or the warranty of the consequence.
- A legal system can be operationally closed and still produce legal fictions.
- A bureaucracy can be operationally closed and still produce procedural nonsense.
- A scientific institution can be operationally closed and still produce fashionable error.
- An AI system can be operationally closed and still produce fluent hallucination.
Closure of the system is therefore necessary, but not sufficient.
The next step is closure of the claim.
In my work, closure is not merely the self-reproduction of operations inside a system. Closure is the satisfaction of the demand for decidability in the context in question. A claim is not closed because it has been stated. It is not closed because it sounds plausible. It is not closed because an expert endorses it, a model predicts it, or an institution accepts it. A claim is closed only when the relevant tests have been satisfied, and when the remaining uncertainties have either been eliminated or explicitly identified.
This requires a broader grammar of closure.
- A claim must close over identity: what exactly is being asserted?
- It must close over operation: what actors, actions, objects, sequences, and conditions are involved?
- It must close over evidence: what observations, records, or testimony support it?
- It must close over causality: what mechanism connects premise to consequence?
- It must close over possibility: can the proposed action occur under real constraints?
- It must close over reciprocity: does the action preserve sovereignty in demonstrated interests, or does it impose an uncompensated externality?
- It must close over authority: who has standing to decide, bind, command, transfer, certify, or warrant?
- It must close over liability: who bears responsibility if the claim fails?
- It must close over context: what degree of infallibility is required given the severity, reversibility, population affected, and institutional consequence?
This is the difference between communication and decidability.
- Luhmann explains how law remains law.
- Natural Law explains how a claim becomes decidable.
- Runcible implements that decidability as an institutional process.
This matters because the AI industry is currently trapped between two incomplete solutions.
On one side, foundation models generate increasingly capable candidate speech. They are powerful hypothesis engines. They can produce language that is often useful, often insightful, and often directionally correct. But they do not, by themselves, produce closure. They produce possibilities.
On the other side, the industry tries to manage this incompleteness with confidence scores, benchmarks, reinforcement learning, human review, policy filters, retrieval systems, and alignment constraints. These are useful instruments. But they do not solve the closure problem. They improve generation, constrain generation, or review generation. They do not transform candidate speech into institutionally warrantable action.
- A confidence score is not closure.
- A citation is not closure.
- A benchmark is not closure.
- A policy filter is not closure.
- A human-in-the-loop is not closure unless the human is operating within a testable protocol of authority, evidence, liability, and decision.
This is why institutional AI remains blocked in high-liability domains. Insurance, medicine, law, finance, government, defense, compliance, procurement, and administration do not merely need better answers. They need outputs that can survive audit, challenge, appeal, correction, and liability.
They need closure.
Runcible begins from a different premise: the foundation model should not be treated as an oracle. It should be treated as a hypothesis generator. The model proposes. The closure layer adjudicates.
The Runcible process converts generated speech into operational form. It identifies the claim, actors, actions, objects, rules, evidence, missing conditions, authority chain, liability boundary, and applicable protocol. It then tests the claim against the relevant gates: truth, possibility, reciprocity, authority, liability, and contextual decidability.
The result is not merely an answer. The result is a Decidability Record.
A Decidability Record states what was claimed, what evidence was used, what rules were applied, what tests were passed, what tests failed, what remains undecidable, who has authority, where liability attaches, and whether the output is fit for institutional action.
This changes the role of AI.
- Without closure, AI remains a producer of plausible text.
With closure, AI becomes a participant in governed institutional workflow. - Without closure, the institution must absorb the risk of model output through human review, informal judgment, and discretionary override.
- With closure, the institution receives an auditable artifact showing why the output is actionable, non-actionable, prohibited, unsupported, unauthorized, impossible, irreciprocal, or undecidable.
The innovation is therefore not simply another AI application. It is an expansion in the scope of closure itself.
- Maturana and Varela gave us closure of the living system.
- Luhmann gave us closure of the communicative system.
- Natural Law gives us closure of the claim.
- Runcible gives us closure of the institutional act.
This is meaningful for the AI industry because the next market is not the assistant market. The next market is governed institutional action.
The assistant market rewards fluency, convenience, and productivity. The institutional market rewards authority, auditability, warrantability, and risk reduction. A chatbot can answer. An institution must decide. And a decision must be closed.
That is the missing layer between foundation models and high-liability adoption.
The largest opportunity in AI is not merely to make models more intelligent. It is to make their outputs testable, reviewable, certifiable, and actionable inside institutions that bear consequences.
AI does not need only more generation.
It needs closure.
Cites:
- Luhmann, N. (2004). Law as a Social System. Oxford University Press. (Originally published in German as Das Recht der Gesellschaft, 1993.)
- Luhmann, N. (2004). “The Operative Closure of the Legal System.” In Law as a Social System (English ed., translated by K. A. Ziegert). Oxford University Press.
- Luhmann, N. (2004). “Closure and Openness: On Reality in the World of Law.” In Law as a Social System (English ed.). Oxford University Press.
(Also circulated as: Luhmann, N. “Closure and Openness: On Reality in the World of Law,” downloadable as a standalone PDF in some law‑review style formats.) - Luhmann, N. (2012). Theory of Society, Volume 1. Trans. Rhodes Barrett. Stanford: Stanford University Press.
- Luhmann, N. (2024). “Closure and Structural Coupling.” Cardozo Law Review, 13(5). (Reprinted/translated essay, available in PDF from Cardozo Law repository.)
- Maturana, H. R., & Varela, F. J. (1980). Autopoiesis and Cognition: The Realization of the Living. D. Reidel.
- Capra, F. (2022). “The Organization of the Living: Maturana’s Key Insights.” Constructivist Foundations, 18(1), 5–13.
- Bednarz, J. (1988). “Autopoiesis: The Status of its System Logic.” BioSystems, 22(1), 29–41.