The Drift Neural Network: A Neuro-Symbolic Cognitive Architecture for Autonomous Systems

Abstract

The Drift Neural Network is a neuro-symbolic cognitive architecture that joins four faculties in one closed loop, offered against the prevailing bet on scaling the autoregressive transformer. A transformer produces output by statistical continuation over a fixed matrix of weights, and four of its limits are structural, unmoved by more parameters: no internal test for truth, predictions that do not themselves certify that an action satisfies a rule, loss of earlier competence when a new task is learned, and grounding in correlation rather than mechanism. The architecture answers each with a faculty. It grounds continuous sensing into discrete terms and holds it within safe bounds. It derives causal mechanism by intervention rather than by fitting correlation. It retains earlier competence in a distributed associative memory. Its reasoning core is an explicit executable graph that rewrites its own structure, each change admitted only after a formal check of the exact computation that will run, so adaptation is verified construction rather than gradient descent on a fixed function. The campaign is audited and reproducible, at toy scale: each faculty was validated against an adversarial control, and self-modifying agents compose into a system that improves on isolated instances. Under a safety ablation the gate stays active in every arm: all arms generate the same 58.8% rate of structurally invalid proposals, none admits an invalid change, and repair drives the residual to zero, cutting search evaluations by 28.2%. The central obstacle, decomposing a task into parts, is met without a human: a compressing cut is found from a raw table, its sub-specifications derived, and a verified circuit synthesised where a monolithic solver times out, a filter pruning the search to a shortlist of one or two. The evidence supports viability, not a finished intelligence. Scaling to higher-order cuts and to functions without an exact decomposition remains open.