Contract-Gated Self-Modification: A Deterministic Symbolic Substrate for Bounded Structural Adaptation
Year: 2026
Type: preprint
Abstract: Autonomous artificial intelligence is dominated by the autoregressive transformer, a fixed matrix of weights that generates output by statistical continuation. Scaling that model does not by itself supply a formal certificate of the properties autonomous operation requires: an internal check on truth, a check on an action before it is taken, retention of earlier skills, and grounding in mechanism rather than correlation. We present an imperative, self-modifying, symbolic execution substrate with a deterministic core, proposed as an alternative for safety-critical autonomous execution. In place of a fixed weight matrix, it holds an executable graph that rewrites its own structure, and it admits each change only after a formal check. The substrate is deterministic and contains no stochastic generator, so the check applies to the computation that will actually run. Under sound checking and faithful interpretation, this gives an admission guarantee that does not depend on a generative model behaving as required or on a runtime fallback for noncompliant model outputs. We report a reproducible evaluation at controlled toy scale in which execution, structural contract checking, structural repair, and memory run together without observed interference. The self-modification we evaluate is a synthesis-time search. Scale and autonomous problem decomposition remain open. This record includes the paper and a self-contained reproduction bundle that regenerates the reported tables and benchmark results.
Author: Vandenberg, Ibrahim
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7213774570
Cites:
Cited by:
Related to: