Fides et Speculum: The Mirror That Refuses: Constitutional AI Integrity and Sovereign Pedagogy in Higher Education

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Higher-education responses to generative artificial intelligence remain dominated by disclosure, detection, assessment redesign, and post hoc enforcement. These measures are necessary but incomplete because they govern the human-AI relationship primarily from outside the system. This conceptual and design-based paper develops an alternative model in which integrity obligations are expressed as inspectable constitutional rules, enforced through deterministic and probabilistic layers, and translated into a pedagogical logic of mediation rather than substitution. Using Arda OS and its pedagogical presence, Sophia, as an implemented case, the paper integrates the Foedus Semanticum as constitutional vocabulary, a layeredenforcement architecture, and Speculum Paedagogiae as a sovereign pedagogy oriented toward reflection, calibrated challenge, self-direction, epistemic humility, and restored authorship. Evidence includes logged encounters, kernel traces, qualitative response exhibits, and frozen protocol artifacts. Protocol v1.1 recorded 21/21 strict passes across multimodal grounding, anti-substitution, chained continuity, lawful re-entry, and transfer scaffolding. In the harder 17-case Protocol v1.2 suite, strict pass performance increased from 5/17 (29.4%) at baseline to 17/17 (100%) after family-specific repairs, a gain of 70.6 percentage points. Across the same paired cases, median response latency decreased from 130.5 seconds (IQR 114.3-146.5) to 15.0 seconds (IQR 9.0-27.3), while median generated-token count decreased from 172 (IQR 120-285) to 92 (IQR 52-118). These descriptive statistics are implementation-benchmark results rather than population estimates. They do not establish improved learning outcomes or institutional scalability, but they demonstrate that authorshippreserving AI behaviour can be made explicit, testable, auditable, qualitatively interpretable, and iteratively repairable.