Nous: Adversarial Testing of a Dependency-Free Memory-and-Activation Substrate for Cognitive Architectures. An Engineering and Methodology Case Study

This report describes Nous, a dependency-free Python prototype of a memory-and-activation substrate for a cognitive architecture, and (more centrally) the software-engineering methodology used to build it. Nous comprises four layers: a graph substrate store of concept nodes and weighted roads; a deterministic spreading-activation engine; a multi-source activation field; and a thin end-to-end perceive-activate-form-coalitions-select-a-thought slice. The lower two layers are verified, by an adversarial property-style test suite, for three concrete properties: determinism (identical output across process runs, confirmed in fresh interpreters under randomized hash seeds), integrity (graph indices stay mutually consistent across every mutation), and purity (reads never mutate stored state). The upper two layers are exploratory: they run and are deterministic, but their cognitive meaningfulness is explicitly not claimed, and on clustered real input the end-to-end slice produces an incoherent "thought," which is recorded as an asserted negative result and known limitation, not a contribution. The central finding is a recurring failure mode: a component that passes on a friendly, hand-chosen input and fails on an adversarial or system-generated one, and the two practices that caught it: adversarial/property-based testing (including cross-process determinism and read-purity checks) and loop-closure (driving each stage with system-generated rather than hand-placed inputs). Together they exposed six distinct structural bugs that ordinary happy-path tests pass over. The artifact's verified totals: 130 assertions across 56 test functions, all passing, in five suites.