Resources / Method
From Plausibility to Proof
Constructive falsification, operational representation, and the grounds of institutional reliance.
In this article
- 1. Plausibility leaves the grounds unexamined
- 2. Constructive proof and constructive falsification
- 3. What a finite procedure can and cannot settle
- 4. Operational language preserves the causal account
- 5. EOSA: a discipline for reconstructing the claim
- 6. Work through the compensation case
- 7. Correspondence across domains must be constructed
- 8. Retain the result and the means of correction
- 9. The method must face its own falsification
To make an AI claim useful for institutional judgment, we must recover what its language leaves implicit: the objects it refers to, the operations it asserts, the evidence it requires, the interests it affects, and the reliance it asks someone to undertake.
Runcible applies constructive falsification to that task. Construction exposes how a claim could fail; examination determines which conditions survive, which are defeated, and which remain open. “Proof” in this context must name the kind of evidence supplied. A formal derivation, an empirical observation, and a record of institutional authority carry different burdens.
1. Plausibility leaves the grounds unexamined
A model can generate “the revised policy will improve cooperation” because the sentence fits a familiar managerial argument. That does not identify the mechanism by which behavior changes, the interests the policy alters, the evidence for the expected result, or the authority to impose the change.
The problem becomes harder when several reasonable statements surround one unsupported step. Unequal territories may exist. Teamwork may be desirable. Compensation may influence behavior. None of those propositions establishes that redistributing already earned commissions is permitted or will achieve the desired result.
The required examination decomposes the bundle. Each claim must carry its own grounds; rhetorical proximity cannot transfer support from one to another.
2. Constructive proof and constructive falsification
In constructive mathematics, establishing existence generally requires a construction or witness of the relevant kind. That discipline is useful: asking what must be constructed can reveal the difference between asserting that a solution exists and supplying it.
Runcible extends the operational question to claims about the world. Identify the actors, objects, actions, resources, dependencies, and sequence required for the asserted state. Examine whether each necessary relation can hold and whether the supplied evidence establishes that it does.
This use of construction is falsificationary. A contradiction, impossible operation, or missing prerequisite can defeat the proposed account. A construction that survives the required challenges supports possibility under stated conditions. It does not establish that a historical event occurred, that a party consented, or that an actor possesses authority.
Failure to discover a construction is also not proof that none exists. The record must distinguish an established impossibility from an incomplete search or an unresolved condition.
3. What a finite procedure can and cannot settle
A procedure that tests a particular construction is not automatically a decision procedure for all claims in its domain. General algorithmic decidability requires correct termination for every admissible input. Turing’s results remain applicable; semantic interpretation does not bypass them. Primary source: Turing, 1936.
The institutional objective is contextual decidability. For this claim, this use, these parties, and these limits, what examination would justify the requested reliance? What has actually been resolved? What remains material?
A finite examination may end with a decisive contradiction. It may establish the necessary conditions for a bounded conclusion. It may also end with missing evidence, competing interpretations, or a requirement for an authorized human judgment. The last result is useful when it identifies precisely what prevents closure.
4. Operational language preserves the causal account
Terms such as “fair,” “effective,” “safe,” “authorized,” and “true” can conceal several different claims. Operational representation replaces that compression with identifiable relations:
- Entities and identities: Who or what is involved, and how are references resolved?
- Operations and sequence: What changes, by whose action, from which prior state?
- Conditions and resources: What must exist for the operation to be possible?
- Evidence and correspondence: What observations or records support the asserted event or relation?
- Interests and obligations: Who gains, who bears costs, and which commitments already bind the parties?
- Reliance and responsibility: What conclusion is being used, for what purpose, under which limits and remedies?
RDL develops this representation into a stronger compiler boundary. A typed representation makes omissions inspectable; it does not itself certify that its fields are accurate or complete.
5. EOSA: a discipline for reconstructing the claim
The normalization sequence is enumeration, operationalization, serialization, and adversarial disambiguation. These operations recur as the examination discovers new dependencies.
- Enumerate. Separate assertions, meanings, parties, and required findings. Identify the intended reliance before choosing the standard of examination.
- Operationalize. Replace an abstract predicate with the operations, observations, interests, and constraints that would make it testable.
- Serialize. Establish the temporal and dependency order. An authorization required before an action cannot be supplied by a favorable result afterward.
- Adversarially disambiguate. Challenge substitutions, hidden premises, scale changes, causal leaps, omitted costs, and apparent contradictions. Repair the representation and repeat where necessary.
Formal logic, mathematical models, code, retrieval, and empirical tests can participate wherever they preserve the distinctions the question needs. The method is not a ban on abstraction. It is a requirement to account for what an abstraction leaves out.
6. Work through the compensation case
Candidate: Pool sales commissions to improve teamwork. First separate earned compensation from proposed future terms. Identify the relevant agreements, earning conditions, affected employees, proposed transfer, decision-maker, and jurisdiction. If those materials are absent, do not invent their contents.
Possibility: The payroll system may be able to perform the transfer. That establishes a technical operation, not a contractual entitlement. Test the proposed behavior mechanism separately: which incentives change, which conflicts remain, and what observations would support or defeat the teamwork hypothesis?
Reciprocity and obligations: Identify who performed the work, what was promised, who would bear the changed cost, whether consent is required, and which interests a remedy must preserve. A favorable managerial goal does not erase those prior relations.
Repair: Examine prospective agreements, voluntary arrangements, changed future assignments, or other means of addressing the coordination problem. A candidate repair must satisfy its own resource, evidence, consent, and authority conditions. This is an illustrative examination method, not a legal determination about an actual compensation dispute.
Disposition: Preserve the defeated proposal, any surviving alternatives, and the information or decision still needed. Oversing then governs authorization and any subsequent execution. The analysis cannot write itself into payroll.
7. Correspondence across domains must be constructed
The same operational questions recur in science, law, engineering, and institutional work, but the evidentiary procedures differ. A physical measurement, a signed agreement, a source-code test, and a witness statement establish different kinds of facts.
Commensurability comes from mapping the relevant objects, operations, measures, conditions, and consequences. Reusing the word “proof” does not create that mapping. Nor does a lower-level physical possibility establish a higher-level institutional right.
The first-principles hierarchy is therefore a constraint discipline: a proposed institutional operation cannot depend on an impossible physical prerequisite, but compatibility with physics does not settle the institutional question.
8. Retain the result and the means of correction
The Decidability Record should identify the claim and version, evidence, actual examinations, surviving and defeated conditions, open dependencies, repairs, and permitted reliance. It must make the difference between a performed test and a proposed test inspectable.
Oversing connects that record to the responsible decision, delegated mandate, execution, and observed consequences. Later evidence can reopen a conclusion. A repaired proposal can be compared with the original. Qualified cases can improve protocols, compiler behavior, and evaluation sets.
9. The method must face its own falsification
Test whether the system discovers material dependencies that strong baselines miss; whether it actually performs the required examinations; whether it distinguishes contradiction from missing evidence; and whether its repairs preserve the legitimate objective without creating new defects.
Measure false closure and unnecessary rejection, repeated-run stability, review effort, latency, and total cost. Include adverse cases and unauthorized-action attempts. The operational claim is earned by those results, not by naming the procedure a compiler or formatting its output as a proof.
The practical transition is from a plausible statement to an inspectable account of what survives—and then, through a separate authority boundary, to work the institution can responsibly perform.
