This paper examines the gap between the naming of cognitive‑architecture components and their actual implementation in recent agent systems built on local large language models. The authors introduce a four‑question static audit protocol – 1) does the module appear in the source code; 2) is it ever called; 3) does the call chain propagate meaningful data; 4) does the result reach the production output. The audit requires line‑number evidence and call‑chain tracing, making it applicable to any code base.
The target is a system that claims 25 cognitive mechanisms. Falsifiable predictions were pre‑registered for 11 modules and the protocol was replicated on two independent systems. Results show that 18 of the 25 modules never produce any output; three novel dead‑code patterns emerge – connected‑but‑never‑called, called‑but‑inert, called‑but‑content‑blind. Across 61 frozen criteria and 12 interventions, 53 were confirmed, 2 partially adjudicated, 5 falsified, and 1 remained ambiguous. Two repair experiments on critical failures raised planner success from 0.028 to 1.000 and cut state‑controller error by 52%.
The authors argue that module names and their computations must be audited as separate claims. The proposed protocol and five‑class failure taxonomy are reusable across systems, enabling case‑level depth assessment of the claim‑implementation gap. Review