Turquoise International Journal of Educational Research and Social Studies

Controllability Tradeoffs in High-Dimensional Generative AI

Adrian Whitford, Eleanor Markham — Wed, 15 Apr 2020 00:00:00 +0300

High-dimensional generative AI models offer exceptional expressive capacity, yet controlling their output
behavior remains a central challenge in practical deployment settings. As latent spaces expand in complexity,
semantic representations often become non-linear and entangled, making precise directional steering difficult
without compromising generative richness. This study examines how prompt conditioning, latent-vector
manipulation, and external constraint mechanisms influence model controllability, stability, and expressive
diversity. Through iterative generation analysis, workflow-driven integration testing, and robustness evaluation
under input perturbations, the results reveal inherent tradeoffs between creativity and predictability. The findings
underscore the need for context-aware controllability design, where control intensity is adapted to application
domain, user intent, and operational constraints. Such adaptive balancing strategies enable generative models to
achieve both expressive variability and reliable task-aligned behavior.

Dynamic Memory Grant Calculation Thresholds in Oracle PGA Allocation Systems

Amelia Wexford, Daniel Armitage — Sun, 19 Apr 2020 00:00:00 +0300

Efficient Program Global Area (PGA) memory allocation is essential for maintaining stable query execution
performance in Oracle database environments, particularly under fluctuating concurrency and mixed workload
conditions. Static memory grant thresholds often fail to adapt to real-time load variations, resulting in work
area spills, temporary I/O overhead, and degraded response times. This study introduces a dynamic threshold
adjustment approach that recalibrates PGA memory grants based on runtime workload behavior and memory
pressure signals rather than static configuration or optimizer estimates. Experimental evaluation across
analytical, transactional, and mixed workloads demonstrates that dynamic thresholding reduces spill
frequency, improves latency stability, enhances throughput fairness among sessions, and accelerates recovery
following overload events. The results highlight the role of adaptive memory tuning in sustaining predictable
performance in enterprise systems, especially those driven by interactive, user-variable application layers.

Ethical Constraint Encoding in High-Autonomy AI Simulation Scenarios

Adrian Falkner, Rowan Merrick — Sun, 26 Apr 2020 00:00:00 +0300

High-autonomy AI systems trained in simulation environments must be guided by ethical constraints that
influence not only the outcomes they produce but also the internal reasoning processes through which
decisions are formed. Traditional constraint strategies based on rule enforcement or reward shaping often fail
under complex or adversarial conditions, leading to behavior that superficially meets ethical requirements
while violating deeper normative expectations. This study introduces a framework for encoding ethical
constraints directly into the representational and policy layers of autonomous agents, combined with dynamic
context-based modulation that adjusts ethical priorities according to situational demands. Simulation results
across cooperative, competitive, and mixed-motivation environments show that agents with embedded ethical
priors exhibit consistent value-aligned behavior, maintain strategic adaptability, and resist exploitation
attempts that circumvent rule-based controls. The findings highlight the importance of treating ethical
alignment as a structural learning principle rather than a post-hoc regulatory mechanism.

Explainability Fidelity Metrics for Post-Hoc Model Interpretation

Gregory Ashford, Celeste Rivenhall — Thu, 30 Apr 2020 00:00:00 +0300

Post-hoc explanation methods are widely used to interpret complex machine learning models, yet the fidelity of
these explanations how accurately they reflect the model’s true reasoning remains difficult to assess. Explanations
that are easy to understand may oversimplify or distort the decision logic, while highly detailed explanations may
be accurate but unusable in practice. This study presents a structured evaluation framework for measuring
explainability fidelity through local sensitivity testing, global attribution coherence, representation-space
alignment, and causal influence validation. Experimental results show that many commonly used attribution
techniques generate persuasive but mechanistically incorrect explanations, particularly in deep models with
distributed internal representations. Methods that incorporate causal perturbation and representation-level
reasoning exhibit significantly higher fidelity. Additionally, deployment tests in cloud-integrated Oracle APEX
environments reveal that explanation stability depends on system execution context, reinforcing that fidelity is
both a modeling and operational concern. The findings provide a foundation for selecting and validating post-hoc
interpretability techniques in high-stakes enterprise applications.