Applied Frameworks and Studies
This section presents the Center's peer-oriented research: focused investigations, empirical findings, and theoretical contributions that advance AI governance as a discipline. Work here examines adaptive lifecycle control, bias propagation and amplification in large language models, and epistemic decay through model autophagy. These are shorter, argument-driven pieces intended to test frameworks against evidence, surface failure modes, and establish measurable standards for reliability, transparency, and trust.
Adaptive AI Governance as a Lifecycle Control System: From Static Compliance to Continuous Oversight
ABSTRACT
This white paper introduces Adaptive AI Governance, an approach that reframes AI governance as a closed-loop lifecycle control system rather than a static compliance exercise. The framework integrates upstream Boundary Governance for dataset certification, supplier integrity, epistemic traceability, and representational diversity enforcement with the Adaptive Lifecycle AI Governance Framework (ALAGF), which orchestrates continuous monitoring through MIDCOT for drift detection, BME Auditor Pro for behavioral evaluation, SymPrompt+ for runtime alignment enforcement, and LGSB for structured stress testing. The closed-loop feedback architecture ensures that operational signals from all subsystems propagate bidirectionally, enabling real-time corrective action and upstream revalidation when governance thresholds are breached. The approach is explicitly aligned with the NIST AI Risk Management Framework and ISO/IEC 42001, demonstrating that integrated instrumentation and continuous evidence generation are prerequisites for substantive standards compliance.
PUBLISHED DOI PAGES ORCID
January 10, 2026 10.67002/AIGR-2026-001 155 0009-0004-7950-024X
ABSTRACT
Model autophagy (the recursive ingestion of synthetic or contaminated data through iterative retraining cycles) threatens the epistemic integrity of large language models at scale. This paper presents a four-layer causal architecture integrating session-level confirmation bias, corpus-level data contamination, generation-level parameter drift, and civilizational-scale epistemic consequences. Formalized through 21 recurrence equations governing corpus integrity I(n), bias B(n), and quality Q(n), the model predicts exponential decay trajectories with phase transitions around I(t) = 0.5. Validated through controlled GPT-2 recursive retraining experiments (10 generations, S(t) 0.10 to 0.80), Phase 3a (5 tracks, 55 observations) confirms exponential decay (alpha = 1.93, R-squared = 0.98, floor = 0.468). Phase 3b (5 tracks, 55 observations) demonstrates that G2+G3 governance (w_p = 0.60) partially reverses decay: I(10) = 0.894 versus baseline 0.489 (U = 0.0, p = 0.004). Calibrated BRF = 0.115 yields FIF*BRF = 0.179, satisfying the stability condition. These results provide the first empirical demonstration that provenance-based governance can arrest and partially reverse epistemic decay in recursive AI training.
PUBLISHED DOI PAGES ORCID
April 7, 2026 10.67002/AIGR-2026-002 155 0009-0004-7950-024X