Axiom Dynamics 8: Blind-Proposal–Machine-Check Acceptance Experiment on a Criterion-Acceptance Terminological Ontology for Generated Proposals

This paper reports a preregistered interlocking experiment: a generative modelproposes new is_a axioms to four real maintained OBO ontologies, and the axiomtransition-criterion machinery acts as the ledger-layer acceptor, examining, underconditions in which proposals are produced in batches and acceptance and generation belong to the same family, which recomputable readings the machine-checkgate can stably provide and which judgment dimensions it cannot cover. The experiment uses a two-arm design: the blind-proposal arm reads no corpus data, whilethe vocabulary-assisted arm may read only the name list and not the closure structure. The acceptance protocol is fixed before the proposals are frozen; verdicts areexecuted in the order of name matching, duplicate/reflexive, already derivable, cycleverification, and nonempty single gain; gain is based on direct pair-level closure simulation; batch structure is identified by full subset enumeration to find true-subsetdeadlocks, and zero-contribution members are removed from the batch accordingto the protocol while their single-proposal verdicts are retained. Readings are restricted to the Horn-visible fragment, and no semantic assertions are made aboutrelationship, two-parameter intersection, or union structures. Main resultsinclude: under strict name matching, the unresolved proportions of the two armsshow a describable name-level difference; already-characterized ledger defects aretriggered by corresponding verdicts on the proposal distribution of this arm, butthis does not constitute a proof of detector completeness; batch structure shows thatseveral batches have no true-subset deadlock and every proposal has irreducible contribution, while other batches have a unique true-subset deadlock and therefore nolegal sequential order. The separation of ledger legality and semantic correctnessis a constructive result of the protocol; its quantitative readings must be borne bya semantic-review channel independent of the verdict pipeline, and this paper doesnot upgrade the verdict pass rate into an empirical finding. All readings are existence evidence rather than statistical evidence; the delivery status of recomputationmaterials determines the effective conditions of the “publicly recomputable” claim.