Scientific and technical thesis · October 30, 2026 · Runcible research programme. This exposition distinguishes the method we propose, the software that embodies parts of it, and the experimental evidence still required. It complements—not replaces—the market-first account of the institutional AI opportunity.
Generation was the first half of AI. Constructive qualification is the complementary capability. A useful intelligence must produce possibilities, determine what survives the necessary examinations, preserve what remains unresolved, and retain independently checked results to improve later computation.
1. The computational problem is larger than generating persuasive language.
Foundation models have made the generation of hypotheses, explanations, documents, plans, code and alternative solutions inexpensive. That achievement is substantial. But a generated answer is a candidate, not automatically an account of what the world warrants. A sentence can be linguistically plausible while asserting a nonexistent authority, omitting a causal dependency, confusing a promise with an accomplished action, or treating an interested party’s preference as an obligation upon someone else.
Institutions require more than candidate language. They must establish which objects and claims are identifiable, what evidence supports them, which operations are physically possible, what costs and obligations follow, who has authority, what remedies remain available, and whether a proposed action produces its intended result.
The missing general function is therefore not another method for generating answers. It is a method for constructing and executing the examinations necessary to determine how far an answer may be relied upon.
2. Existing successes illuminate the missing function.
Mathematics, formal verification, scientific experimentation and software engineering differ profoundly in their subject matter and difficulty. What often enables computational success in these domains, however, is an identifiable way to check the result: an axiomatic derivation, an executable program, an empirical measurement, a reproducible experiment, or a verifier supplied by the environment.
This does not make scientific discovery, mathematical proof, or programming easy. It means that in many successful applications the conditions of a satisfactory result are more readily specified than the result itself. The verifier may remain costly or difficult to construct, but it can be separated from candidate generation.
A question about institutional behavior often adds another burden. The facts, definitions, causal relations, interests, authorities, liabilities, and criteria of acceptable resolution may themselves be contested or missing. The relevant examination must be constructed as part of solving the problem.
Runcible proposes to extend the discipline of external verification to semantic and behavioral matters in which the test is not supplied in advance.
The historical trajectory: from justification to falsification, construction and operational measurement.
Traditions of argument and justification long sought reasons to accept propositions; their achievements include logic, dialectic, evidence and criticism, not merely unsupported belief. Falsificationary science made adversarial testing especially explicit: a proposed explanation must face possible observations that would count against it. Constructive proof and algorithmic methods show how a conclusion can be established by constructing the required objects or operations rather than merely declaring them to exist.
These traditions overlap historically. Constructive reasoning and experimental test construction were not invented in the twentieth century, and no single date marks the completion of the sciences. What has become possible in recent decades is a more systematic attempt to express heterogeneous empirical and behavioral claims as operational constructions that can be examined computationally, with model-based semantic generation supplying much of the previously prohibitive labor.
The research diagnosis is that disciplinary specialization produced extraordinary verification methods within selected domains without producing one sufficiently general, inspectable account of semantic and behavioral examination across domains. This is a proposed explanation of the remaining gap, not a claim that contemporary AI researchers, linguists, neuroscientists, economists or philosophers lack relevant knowledge.
Easy and hard closure refer to the construction burden, not the intelligence required.
Where the relevant representation, constraints and verifier are specified, a candidate may be tested within a relatively closed problem. Where objects, causal dependencies, evidence requirements, affected interests and standards of resolution must first be reconstructed, closure is harder. These terms describe the practical burden of constructing sufficient examinations, not an assertion that mathematics is trivial or that every difficult institutional matter has an algorithmic solution.
3. Language is an instrument of measurement before it is an instrument of persuasion.
A language can identify objects, distinguish relations, describe events, specify actions, locate changes, reconstruct causes, and communicate limits. These are measurement functions: they permit a description to be compared with observations and operations outside the description itself.
Ordinary language is frequently used without maintaining those distinctions. We employ the same terms for intentions and outcomes, beliefs and facts, hypothetical contracts and performed commitments, plausible mechanisms and demonstrated causes. Such compression permits fluent accounts that cannot be examined without first recovering what has been conflated.
The Natural Law research programme therefore attempts to develop a constructive operational grammar: a way to express the relevant objects, differences, relations, actions, dependencies, evidence and limits so that proposed accounts can be examined. Mathematics and formal logic remain indispensable special-purpose grammars, but the target is not merely a more elaborate mathematics. The target is language with sufficient operational and causal correspondence to represent the irregularities of the subject matter.
The constructive mechanism: from first principles to the necessary tests.
The underlying claim is stronger than fact-checking. Runcible does not merely ask whether a source agrees with a sentence or which prewritten tests resemble the question. It attempts to reconstruct the conditions under which the claimed phenomenon could exist or occur, derive the necessary examinations from those conditions, and identify exactly what has and has not been established.
The geometry analogy.
In geometry, a diagram can look plausible without being constructible from the specified primitives and permitted operations. A straightedge-and-compass problem succeeds only if a valid construction exists under those rules. Other constructions are impossible under the same restrictions; a missing construction is not, by itself, an impossibility proof. Runcible asks the analogous question of language: what identities, objects, transformations, relations, constraints and evidence would be necessary for the claim to hold?
This is an analogy rather than a claim that every problem in reality can literally be built from triangles. Euclidean construction tests formal possibility under stipulated operations; claims about the world also need empirical correspondence, relevant obligations, authority and the limits of reliance.
The two ternary distinctions.
The Natural Law protocol corpus includes a ternary account of evolutionary transformations: positive production or increase of usable order; negative consumption, deterioration or decay; and equilibrium, restoration or stability. It also includes a separate epistemic classification: supported/true, refuted/false, and unresolved/undecidable under the tests and evidence applicable to the question.
A transformation’s evolutionary sign is not the truth value of an assertion about it. An economically productive event may be inaccurately reported; a correctly reported destructive event is nevertheless true as a statement of fact; an unobserved event may remain unresolved. Runcible’s examinations must retain both distinctions rather than collapse them into the same verdict.
The dependency hierarchy.
The first-principles protocol traces the relevant constraints from physics to chemistry, biology, cognition, cooperation, institutions, and narratives. Higher-order claims must not depend on contradictions of lower-level necessities. Its evolutionary construction sequence additionally identifies differences, charges and equilibria, assemblies, compartmentalization and memory, reproduction, cooperation, and differential evolutionary velocity across groups.
For a physical claim, the necessary examination can center on energy, materials, constraints, operations, and observations. For a cooperative claim, it may also require identifying persons, demonstrated interests, incentives, voluntary relations, exchanged value, imposed externalities, and available restitution. Institutional claims may additionally depend on actual permissions, applicable laws, procedures, liabilities, and authorized action. Not every test belongs to every question: the claim itself determines the necessary dependencies.
The model supplies candidate knowledge; first principles supply the constructive obligations.
Foundation-model knowledge expands the candidate search space: possible mechanisms, analogies, constraints, unexpected exceptions, causal explanations, and counterexamples. As models become more capable and knowledgeable, the range of relevant constructions Runcible can explore may increase without manually coding a test for every specific subject.
That is an opportunity, not an automatic guarantee. Additional learned data can contain errors, be recalled unreliably, or generate more plausible unsupported details. Model knowledge can propose a construction or examination; consequential factual claims still require independent evidence, provenance and demonstrated test execution. The relevant research question is whether richer models produce more necessary and valid tests, fewer material omissions, and more correct bounded judgments under the same runtime requirements.
Constructive epistemic execution and contextual decidability.
The compiler’s intended sequence is to enumerate the distinct claims and identities; operationalize the necessary objects and actions; serialize their causal and temporal dependencies; adversarially disambiguate conflations, omissions and substitutions; construct plausible mechanisms and alternatives; derive and execute relevant first-principles tests; then account for evidence, obligations and liabilities within stated limits.
Its valid outputs include qualified closure, demonstrated contradiction or failure, and explicitly unresolved conditions. A possible construction does not prove that it occurred; a verified fact does not automatically warrant an action; and a failure to close does not establish falsehood. The contextual standard is the satisfaction of the demand for infallibility in the context in question without the necessity of discretion.
Runcible tests whether a claim can actually be constructed from the relevant first principles, whether the required evidence and conditions hold, and how far the resulting conclusion may legitimately be relied upon.
The scientific claim is falsifiable. On held-out institutional problems, compare ordinary prompting, evidence-checking agents and the same model governed by Runcible. Test whether the runtime derives necessary examinations absent from its prior protocol list, actually executes those examinations, distinguishes falsity from missing evidence, constructs valid repairs, and delivers enough benefit to justify its cost. This is the gap between a promising research method and independently demonstrated general execution.
A complete internal statement of the technical dependencies, three audience explanations and experimental requirements is maintained separately. The public Runcible AI page explains what is available to users; the institutional opportunity page explains the commercial problem to which the method is addressed.
4. Constructing the right examination is often harder than executing it.
Suppose an executive proposes pooling sales commissions already earned by employees, because unequal territories make the existing allocation appear unfair. A fluent model may argue convincingly either for or against that plan. The necessary examination is more particular. Which commissions were earned under which agreements? Which territories differed, and why? Which obligations already exist? Are revised terms prospective or retroactive? Which proposals are voluntary? Who would bear which costs, and what alternatives preserve the legitimate objective of improving fairness?
No single similarity score, sentiment classification, or citation lookup settles that matter. The investigator must construct the distinct questions whose answers determine what can be warranted, and must not pretend they have been answered when necessary evidence is absent.
A corresponding problem appears in administrative law, medical necessity, insurance coverage, contractual disputes, contested policies, and organizational planning. The required tests vary with the identity of the matter, its actual operations, the relevant participants, and the consequences at stake.
5. Evolutionary computation supplies the general structure.
The Darwinian analogy distinguishes four operations: generation of candidate variations; adversarial examination against constraints and evidence; retention of what survives under stated conditions; and adaptation of subsequent generation and examination using the retained results.
Modern foundation models perform candidate generation extremely well. They also have some capacity to criticize, test and improve their outputs. But when the same model supplies the hypothesis, chooses its own evidence, declares the examination complete, and assigns the final verdict, no independent epistemic warrant follows merely from that fluent procedure.
The proposed advance is to make the second operation an explicit and inspectable computation. The runtime must determine which tests are required, whether they were actually executed, what independent evidence was used, what survived, what failed, and what remains undecidable.
Survival is not itself a synonym for truth: a bad hypothesis may survive inadequate tests, and a true claim may remain unestablished for lack of evidence. The adequacy of the tests and the provenance of the evidence are therefore part of the examination.
6. Construction proceeds in both directions.
Bottom-up construction begins with identifiable objects, observations, possible operations, causal relations, constraints, and demonstrated consequences. It reconstructs how the claimed state of affairs could occur.
Top-down deconstruction begins with the proposed whole: an assertion, policy, plan, agreement, decision, or remedy. It enumerates the claims and proof obligations that would have to hold for that whole to be warranted.
The two directions meet under adversarial examination. A claim that cannot be connected to a possible operation, sufficiently established evidence, a coherent relation, or an accounted-for liability remains unqualified. A contradiction or omission is a finding to record, not something to conceal in an agreeable conclusion.
The associated normalization discipline is enumeration, operationalization, serialization, and adversarial disambiguation (EOSA): identify the distinct claims and meanings; reconstruct what each would require in operations and evidence; establish dependency and temporal order; and test for conflation, substitution, ambiguity, and other failure modes.
7. Different questions impose different epistemic obligations.
Physical questions require possibility and empirical correspondence. Inferential questions require coherence and valid entailment. Cooperative questions additionally require examination of relevant participants’ demonstrated interests, voluntary conditions, reciprocal obligations, imposed externalities, and available restitution. Administrative and legal questions add applicable rules, authority, jurisdiction, procedure, and review. Consequential actions add warrantability, foreseeable liability, and defined limits.
These are not universal tests to be ritualistically applied to every statement. Reciprocity is relevant when the question concerns cooperation among agents or actions intersecting their interests—not to whether a physical particle follows an equation. The context determines which obligations are necessary; the runtime must show which were applicable and which were discharged.
The result may be an established finding, an adverse finding, an identified contradiction, a constructive repair, a conditional judgment, or a specifically bounded unresolved state. Not all useful answers are positive verdicts.
8. Contextual decidability is not universal algorithmic decidability.
Decidability: satisfaction of the demand for infallibility in the context in question without the necessity of discretion.
That definition concerns the sufficiency of an examination for a specified purpose, given stated limits, evidence, procedures and standards. Its point is not that every proposition can be decided unconditionally. The Church–Turing results establish that no general algorithm decides every instance of certain formal problems. Constructive epistemic examination does not repeal those results [4].
Instead, the method seeks to enumerate what is required for judgment, recursively test what can be closed, preserve what cannot, and decide whether the remaining uncertainty exceeds the demand for infallibility appropriate to the use. A bounded medical billing exception, a disputed contract interpretation, and a scientific theory do not demand exactly the same evidence or tolerate the same unresolved dependencies.
A negative or undecidable result can itself be operationally important. It may identify the missing observation, required authority, conflicting obligation, or alternative course needed for a later resolution. The claim of generality applies to the examination procedure, not to guaranteed closure of every subject matter.
9. A compiler and runtime for warranted semantic claims.
The proposed architecture assigns different operations to different components. The foundation model constructs meanings, hypotheses, comparisons and candidate repairs. The Runcible runtime specifies necessary protocols, applies admissibility and examination gates, requires evidence and execution artifacts, preserves partial closure or non-closure, and authorizes merits adjudication only within the permitted conditions. The Decidability Record preserves what was examined, the evidence and its provenance, the tests that held or failed, dependencies, liability and authority limits, and the remaining work.
In this arrangement a protocol is not merely a passage telling a model to be rigorous. It is a required operation whose execution can be checked independently of the model’s assertion that it complied. A valid output schema and a polished explanation cannot substitute for actual execution of required epistemic work.
Oversing supplies the complementary institutional state: persons, roles, purposes, organizations, resources, responsibilities, permissions, workflows, actions, and observed consequences. It need not replace its relational and deterministic workflow architecture with an LLM. It supplies the continuing world against which semantic proposals can be evaluated, authorized, executed, remembered and revised.
Runcible AI and Oversing are distinct products. Their combination connects qualified knowledge to responsible cooperation and institutional action.
10. The distinction from adjacent research must be experimentally established.
Existing AI research already demonstrates important versions of verification. OpenAI has shown the value of process supervision in mathematical reasoning [1]. DeepMind’s AlphaProof couples candidate search with formal proof checking, where the verifier is supplied by Lean [2]. NIST is developing integrated probes and machine-readable audit trails for the factual grounding of agentic outputs [3]. These are relevant advances, not evidence that competitors do no verification.
Our stronger proposed distinction is the construction of contextually necessary semantic and behavioral tests, the independent qualification of their execution, and the retention of bounded institutional conclusions, alternatives and consequences. A system does not become different merely because it names more protocols. It becomes different if it detects missing obligations, constructs material tests that baselines omit, executes them, avoids unwarranted closure, and resolves more problems at acceptable cost.
11. A qualified examination is both an output and a potential learning asset.
Ordinary answers provide candidate content. A qualified examination can additionally supply the operations necessary to warrant it, evidence supporting or contradicting it, the boundaries of its validity, alternatives that repair failure, and subsequent observations about what actually happened.
The hypothesized learning loop is: real institutional matters reveal missing tests; independently checked outcomes identify which examinations succeeded or failed; the protocol library and runtime incorporate validated corrections; later examinations require less reconstruction, omit fewer obligations, and preserve more correct distinctions. Only rights-compatible, sufficiently qualified records become reusable assets.
This improvement does not require changing a model’s weights. A separate, longer-term research programme may use qualified examinations to improve model evaluation, training data, attention or learned representations, with the aim of making causal distinctions and warrantable judgments more reliable and efficient. That remains to be demonstrated rather than asserted.
12. The scientific claim must survive controlled comparison.
Experiment A — Test construction
Present previously unseen semantic and behavioral matters with incomplete facts, hidden material dependencies and competing interpretations. Compare an ordinary foundation model, prompting with self-review, conventional tool-using or evidence-checking agents, and the same model under the Runcible runtime. Independently score whether each system discovers necessary tests that were not provided in advance, and whether the tests are actually performed.
Experiment B — Closure and non-closure
Supply cases with sufficient evidence, contradictory evidence, insufficient evidence and deliberately altered premises. Measure false positive closure, false rejection, correctly preserved unresolved conditions, stability under irrelevant changes, sensitivity to material changes, provenance, and independent reproducibility. A persuasive explanation is not an adequate substitute for a verified result.
Experiment C — Constructive repair and institutional value
Present defective agreements, policy proposals, compliance exceptions or operational decisions with identifiable legitimate objectives. Test whether the system can preserve those objectives while eliminating invalid conditions, develop feasible alternatives, and connect the resulting matter to authorized work. Measure human review, time to completion, quality of resolution, model cost, institutional implementation effort and observed consequences.
For each experiment, pre-register the test corpus and adjudication rules, maintain independent ground truth or qualified expert review, include held-out problems, report adverse cases and inconclusive outcomes, and account for total cost. Where no unique correct institutional settlement exists, grade the stated objective, constraints, grounds, affected interests and admissible alternatives—not arbitrary preference agreement.
13. What has been implemented, and what remains to be proved.
Current company records describe a working early-stage Runcible protocol runtime, compiler and gateway; retained development examinations; a substantial institutional operating application; and one reported integrated deployment. These artifacts support an engineering foundation. They do not by themselves establish complete non-bypassable execution of all necessary semantic protocols, independent superiority, fully general contextual decidability, repeatable enterprise implementation, or improved foundation-model weights.
The near-term engineering obligation is to show actual execution of required epistemic operations, durable and independently inspectable records, correct treatment of non-closure, and successful adversarial tests. The near-term commercial obligation is to demonstrate improved resolution of a consequential, recurring institutional matter at a measured total cost. The long-term scientific opportunity follows from those results, rather than replacing them.
14. Why this matters beyond AI research.
Human societies now coordinate through institutions whose scope, interdependence and rate of change exceed unaided individual comprehension. Their recurring failures concern unreliable accounts, incomplete evidence, conflicting purposes and obligations, hidden costs, insufficient coordination, unaccountable authority, and forgotten consequences.
Reducing the cost of qualified collective judgment and cooperation could improve enterprise performance, legal and contractual settlement, insurance administration, healthcare operations, public decisions, and the adaptation of institutions to new knowledge and constraints.
The ambition is therefore not merely a better conversational assistant. It is to make warranted semantic computation a repeatable institutional capability: one that generates possibilities, examines what survives, records what failed and remains unresolved, supports authorized action, and learns from observed consequences.
Foundation models generate possibilities. Runcible examines what survives. Oversing connects qualified knowledge to responsible action and learning.
References and research context
[1] OpenAI, Improving Mathematical Reasoning with Process Supervision (2023).
[2] Google DeepMind, AI Achieves Silver-Medal Standard Solving International Mathematical Olympiad Problems (2024); see also the AlphaProof research article.
[3] NIST, Building Evaluation Probes into Agentic AI (2026).
[4] Stanford Encyclopedia of Philosophy, The Rise and Fall of the Entscheidungsproblem.
Related Runcible material: The Institutional AI Opportunity (market argument) · From Plausibility to Proof (earlier research essay) · Current capabilities and development boundaries. This essay refines earlier claims rather than silently changing their historical text.