The Kernel–Solver Thesis: Adversarial Learning Toward Verified Mathematical Claims

Fuente: Zenodo
Saved in:
Bibliographic Details
Main Author: Figurelli, Rogério
Format: Recurso digital
Published: Zenodo 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866901068677382144
author Figurelli, Rogério
author_facet Figurelli, Rogério
contents <p>Math Machines established a seminal architectural separation between generative solvers and governed closure: generation explores, while machine-enforced contracts decide what may be promoted and what must remain held. Building on that foundation, this paper develops the Kernel–Solver Thesis as an operational and testable refinement and presents it in manifesto form: a two-agent loop in which a Solver proposes mathematical claims and proof attempts, while a governance Kernel enforces bounded promotability through explicit regimes, receipt requirements, monotonic promotion laws, and replayable verdicts. Rather than treating confidence as correctness, the Kernel formalizes a verdict algebra {HOLD, CONDITIONAL, PASS, INFEASIBLE} and prohibits upgrades under tightened admissibility or reduced verification budgets unless gaps are closed with admissible artifacts. We report adversarial refinement traces on linear-versus-cyclic threshold problems, where the Kernel detects regime slippage (interval reasoning incorrectly applied to ℤ_n), forces downgrades, and guides repair. The contribution is not the separation itself, but a concrete promotability protocol—regime ladder, receipt spine, a constitutional rule set for promotion, a proof promotability card schema, and convergence metrics—showing how verified mathematical claims can be produced as replayable state transitions under declared constraints.</p>
format Recurso digital
id zenodo_https___doi_org_10_5281_zenodo_18764454
institution Zenodo
language
publishDate 2026
publisher Zenodo
record_format zenodo
spellingShingle The Kernel–Solver Thesis: Adversarial Learning Toward Verified Mathematical Claims
Figurelli, Rogério
Kernel–Solver Thesis
bounded promotability
adversarial proof refinement
receipt-based verification
monotonic promotion laws
replayable verdicts
autonomous mathematics
governance computation
manifesto
<p>Math Machines established a seminal architectural separation between generative solvers and governed closure: generation explores, while machine-enforced contracts decide what may be promoted and what must remain held. Building on that foundation, this paper develops the Kernel–Solver Thesis as an operational and testable refinement and presents it in manifesto form: a two-agent loop in which a Solver proposes mathematical claims and proof attempts, while a governance Kernel enforces bounded promotability through explicit regimes, receipt requirements, monotonic promotion laws, and replayable verdicts. Rather than treating confidence as correctness, the Kernel formalizes a verdict algebra {HOLD, CONDITIONAL, PASS, INFEASIBLE} and prohibits upgrades under tightened admissibility or reduced verification budgets unless gaps are closed with admissible artifacts. We report adversarial refinement traces on linear-versus-cyclic threshold problems, where the Kernel detects regime slippage (interval reasoning incorrectly applied to ℤ_n), forces downgrades, and guides repair. The contribution is not the separation itself, but a concrete promotability protocol—regime ladder, receipt spine, a constitutional rule set for promotion, a proof promotability card schema, and convergence metrics—showing how verified mathematical claims can be produced as replayable state transitions under declared constraints.</p>
title The Kernel–Solver Thesis: Adversarial Learning Toward Verified Mathematical Claims
topic Kernel–Solver Thesis
bounded promotability
adversarial proof refinement
receipt-based verification
monotonic promotion laws
replayable verdicts
autonomous mathematics
governance computation
manifesto
url https://doi.org/10.5281/zenodo.18764454