Real Computation or Merely Naming? A Pre-Registered Mechanism Audit of 25 Cognitive-Architecture Modules in an LLM-Based Agent System
Organizations: Shenyang Institute of Technology, Shenyang, China
Abstract
Agent systems built on local large language models are increasingly marketed through the vocabulary of cognitive architecture: modules named global workspace, theory of mind, causal reasoning, or active inference are listed as features, and the presence of the name is taken as the presence of the mechanism. We ask the prior question -- does the named computation actually happen -- and contribute a transferable verification method for the claim-implementation gap: a four-question static audit protocol with line-number evidence and call-chain tracing, demonstrated on two systems. We audit a system claiming 25 cognitive mechanisms, pre-register falsifiable predictions for 11 modules and execute the interventions, replicate the protocol on the IIT-phi core of a second, independent system with its own pre-committed falsification criteria, and run two repair experiments as positive controls. In 18 of 25 modules the computed value never reaches any production output; three idle-wiring shapes new to the dead-code and code-smell literatures emerge (connected-but-never-called, called-but-inert, called-but-content-blind). Of 61 frozen criteria across 12 interventions, 53 were confirmed, two partially adjudicated, five falsified, one adjudication-ambiguous. Repairs raised planner task success from 0.028 to 1.000 and cut state-controller error by 52%. Module names and module computations must be audited as independent claims; the protocol and the five-class failure taxonomy are reusable for the claim-implementation gap at two-system, case-level depth.
Figures & tables
| # | Module (claimed mechanism) | Class | Reach | Key evidence (line numbers in repository) |
| 1 | causal_reasoning (causal reasoning) | name def | ✗ | hardcoded strength table (L269–275); simulate_intervention returns {} (L482–485); direction set by correlation sign alone (L530–534) |
| 2 | theory_of_mind | rule table | ✗ | beliefs keep last write only (L106); mentalize returns (L291) |
| 3 | world_model | rule table | ✗ | “state similarity” compares entity counts (L637–642); dynamics reduce to (L726) |
| 4 | self_awareness | rule table | ✗ | hallucination = keyword count (L238–269); confirmation bias computes then return 0.0 (L322); uncertainty = random.choice (L514) |
| 5 | active_inference (FEP) | name def | ✗ | real binary Shannon entropy (L207) but no argmin/gradient; see suture23 |
| 6 | negative_entropy | rule table | ✓ a | keyword match + fixed coefficients; sign of “negentropy” decided by a coin flip ( random.random() , L332) |
| Suture | Module | Criteria | Headline result |
| 19 | nine-state controller | 5: 3✓, 1✗, 1 ∗ | Transition/gating perturbation leaves output bit-identical (dead code confirmed); recognizer learns its task (acc 1.000 vs. chance 0.111); the broken residual is information-free ( broken random , identical MSE to 4 decimals). P3 falsified in the system’s favor : once wired with multiplicative gating, error drops 52% (Section 7 ). ∗ |
| 20 | dopamine system | 5: 3✓, 1 , 1✗ | TD learner diverges ( up to on the 7-state MDP, on the 5-state; production ): the update applies to the arrival state with coefficient (the product of the TD learner’s rate, trace, and discount terms). After repairing credit assignment the learner converges (corr 0.99) but still underperforms a textbook TD( ) by 0.32–0.70; and on the production state space converges to the constant – the state is a result label, uninformative about the next step. |
| 21 | causal reasoning | 6: 6✓ | Direction recovery accuracy 0.5000 on balanced synthetic graphs (chance); direction is a pure function of the correlation sign; confounded structures produce direct-edge false positives 3/3 with confidence 0.91–0.93; interventions propagate nothing ( downstream_effects ); counterfactual consequents are graph-independent and constant ( ). All three defining operations of causality – do, counter- factual, direction – are absent. No minimal repair exists. |
| 22 | holographic memory | 5: 5✓ | Global RNG pollution in 100% of runs (victim output shifts 0.3238 0.5356); synonymous vs. unrelated cosine differs by (no discrimination); synonym retrieval top-1 rate 0.05 = chance; mean pairwise cosine 0.7538 (all-positive orthant): retrieval is arbitrary ranking. No HRR structure exists. |
| 23 | active inference | 5: 5✓ | “minimize” occurs only in class names; across 200 self-timed inference cycles free energy changes by (while three goals are generated); FE changes only through external belief updates ( ). The computation is real; the minimization – the F in FEP – does not exist. |
| 24 | intuition engine | 4: 4✓ | On 200,000 random points and the production grid, the reachable count for patterns PAT_003/004 is 0 (key-space intersection is empty); satisfaction text hits the correct answer 0% of the time and returns None 90% of the time; confidence ratchets monotonically to 1.0 because the production caller always passes {"success": True} . |
| Class | Definition | Instances | Repair prospect |
| 1. Pipeline defect | Real algorithm; a plumbing defect severs it from the search/output | nine-state controller (19); least-action planner (29); aura significance hook (38) | Good: repairs verified effective (52%; 1.000); aura repair is a one-entry dictionary fix plus compute budget |
| 2. Algorithm defect | Implemented algorithm is mathematically broken | dopamine system (20) | Moderate: repair exposed a second, representational defect |
| 3. Definitional absence | The named concept’s defining operation is not implemented at all | causal reasoning (21); active inference (23) | None by patching: requires new implementation |
| 4. Silent side effect | The module corrupts unrelated state | holographic memory (22) | Contained: isolate the RNG |
| 5. Unreachable / dead-wired / constant-fed | Real or fake computation that is never invoked on real inputs | intuition engine (24); global workspace (25); thalamic filter (26); hemisphere coordination (27); attention competition (28) | Mixed: wiring repair is easy; the content-blind variants are class 3 in disguise |
AI Agent Auditing
AI agent auditing examines agents’ actions, tool use, evidence, and outcomes to assess safety, compliance, reliability, and accountability.
See all 312 papers