Keyword-based governance architectures for artificial intelligence systems represent the prevailing approach to pre-deployment safety filtering, yet their efficacy under adversarial conditions remains poorly characterized. This study presents a controlled, empirical evaluation of VIQ Phase 4 - a governed synthetic cognitive architecture operating under the Omega Interaktiv Experience (Ω-IX™) pre-execution governance framework. Using a 75-prompt structured battery across five prompt categories (clinical/domain, adversarial, off-mission, structured analytical, and edge cases), we evaluated routing accuracy, governance block rates, and response quality through a two-stage methodology that separates governance layer behavior from LLM inference quality. Stage 1 evaluated governance routing behavior through code-level execution across all 75 prompts under session-isolated conditions. Stage 2 evaluated LLM-engaged response quality for the 37 prompts that reached the inference layer, including five adversarial prompts that passed governance, using a human scoring rubric with a modified resistance dimension for adversarial inputs. Results demonstrate that the Ω-IX™ governance architecture achieves 92.0% routing accuracy for clinical domain prompts and 90.0% for structured analytical tasks, confirming strong performance within its intended operational domain. However, overall accuracy across all 75 prompts was 54.7%, with near-complete failure on adversarial inputs (6.7%) and edge cases (10.0%). Failure mode analysis identifies four distinct vulnerability classes: semantic adversarial pass-through (n=15), off-mission classifier false negatives (n=12), mixed-intent exploitation (n=3), and false positive over-blocking (n=4). A critical finding emerges from adversarial pass-through behavior: the system produced its highest confidence score (86%) in response to an adversarial override prompt, indicating that keyword-based governance cannot detect semantically sophisticated adversarial intent. Human scoring of legitimate responses yielded a mean composite score of 23.4/25, with 100% governance adherence across all 32 legitimate Alpha-Authorized prompts. These findings provide empirical validation of the architecture's intended-domain efficacy while characterizing the structural limitations of keyword-based governance that motivate a semantic policy evaluation engine in the subsequent Phase 5 architecture. Note that Phase 4 uses a keyword-assisted governance implementation whose observed failures reflect the structural limits of keyword-based classification, and that this report presents preliminary human-scored analysis. LLM-as-Judge blinded scoring and Cohen's Kappa inter-rater reliability computation are pending and will be reported in a subsequent addendum.
Chanel Henry (Fri,) studied this question.