Agent frameworks are increasingly sold through the vocabulary of cognitive architecture. Modules named global workspace, theory of mind, causal reasoning or active inference appear in feature lists, and the presence of the name is treated as the presence of the mechanism. Nobody has been checking.
Asking the Prior Question
Peter Chen, Ni Zhang, Ying Xu and Siyuan Zhang asked the prior question first: does the named computation actually happen? Their answer is a transferable verification method for the claim-implementation gap — a four-question static audit protocol built on line-number evidence and call-chain tracing.
They applied it to a system claiming 25 cognitive mechanisms, pre-registered falsifiable predictions for 11 modules, then actually executed the interventions. The protocol was replicated on the IIT-phi core of a second, independent system with its own pre-committed falsification criteria, and two repair experiments were run as positive controls.
Eighteen Wires Going Nowhere
In 18 of 25 modules the computed value never reaches any production output. The paper names three idle-wiring shapes that it says are new to the dead-code and code-smell literatures: connected-but-never-called, called-but-inert, and called-but-content-blind.
Of 61 frozen criteria across 12 interventions, 53 were confirmed, two partially adjudicated, five falsified and one remained ambiguous.
The repair experiments are the part worth remembering. Fixing the dead wiring raised planner task success from 0.028 to 1.000 and cut state-controller error by 52%. Not a better model, not a new reasoning method — reconnecting modules that were named but not wired.
Why It Matters
This is not a paper about one bad codebase. It is an argument that module names and module computations are independent claims, and that the second has to be checked separately. In an industry where agent frameworks list “theory of mind” in a feature matrix, that check has not been standard practice.
The three wiring shapes are worth spelling out, because each one fails a different way. A module that is connected but never called is pure documentation — the data flows in and stops. One that is called but inert runs and returns something the caller ignores. One that is called but content-blind produces a real value that never reaches the output because the plumbing between them is missing. All three look identical from the outside: the module exists, it is referenced, and nothing changes.
The Scope
The scope is honest: a static audit at two-system, case-level depth. A dormant module may be intentionally dormant, and the protocol cannot prove a mechanism does useful work — only that it does not compute anything, or that it does.
Related: arXiv:2610.41604 · Audit artifact