Answers are grounded in this paper and cite their sources.
Human-authored mathematical corpora are finite at any fixed time; the space of finite proof candidates is unbounded. We investigate whether exhaustive generation can convert this asymmetry into independently expandable verified training data. A frozen Lean 4 context specifies core syntax, universe parameters, declaration dependencies, a permitted axiom policy, and a partial checker. Fair enumeration discovers exactly its finite accepted certificates; declarative coverage requires checker adequacy, and semantic truth adds sound foundations. We preserve the established universal-library framework and develop safe scope-aware construction, transitive dependency restrictions, and statement-complete normal-form generation in an explicitly restricted implicational calculus. The distinction between exact proof coverage and theorem-level coverage determines which pruning is safe. Shared dependency-audit bounds, finite list-scheduling and work--span bounds, and explicit fair service of partial checks separate parallel throughput from eventual completeness. A computable feature model distinguishes proof-strategy diversity from syntactic proliferation and gives a standard finite-pool submodular selection guarantee with an exact version-space interface. Compositional fiber conditions, protected exhaustive service, faithful rendering, and persistent replay connect accepted proofs to a retained archive and learning exposure. An infinite decidable quotient yields unbounded corpus-relative extension without a new authored document per record. Conditional finite-class learning and a joint acquisition-to-learning result establish specialized improvement under explicit certificate and sampling assumptions; counterexamples rule out automatic improvement. Information and proof-search bounds explain why finite foundations can expose new consequences without creating an empirical evidence channel. The resulting theory supports selected mathematical-data independence while leaving useful theorem yield, affordable search, and broad neural-learning improvement unresolved. Research scope: This is a theoretical manuscript. It reports no implemented library, model-training experiment, or measured empirical performance result. Its guarantees are conditional on the mathematical, verification, selection, and learning assumptions stated in the manuscript. Contribution and AI assistance: Pranav Pandey reports contributing the project's core concept and main principles and revising the manuscript. He has confirmed that he accepts responsibility for its claims. OpenAI Codex was used extensively for mathematical development, including formulations, deductions, proof drafts, and examples; literature-review assistance; prose drafting and revision; LaTeX production; and publication preparation. Separate AI-agent analyses supplied internal analytical checks. These checks are not external peer review, independent human mathematical validation, or a machine-checked formalization of the manuscript. Provenance: This paper develops one of two complementary specializations of an earlier supplied 43-page unpublished foundation. The named public manuscript describes the foundation, companion work, shared results, and added developments. No external acceptance, endorsement, or historical-priority claim is implied by this repository description.
No takes yet. Share an insight, caveat, or question.
Pranav Pandey (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: