Scott FaustAI, learning & software

Research

From prediction to evidence—and from evidence to responsible action.

I study how AI systems represent learners and changing state, how evaluation supports particular claims, and where human judgment must remain explicit.

Current work

The foundation of my research path.

One examines the technical auditability of learner models. The other examines the evidence required before model performance can support educational decisions.

M.S. thesis · final semester

Auditable Knowledge Tracing with a Temporal Heterogeneous Student Knowledge Map

The thesis develops a knowledge-tracing representation that separates shared structure, mutable learner state, and immutable evidence. It uses train-only provenance and prediction-before-update replay so the state available to a prediction can be reconstructed and challenged.

The final evaluation was prospectively locked, and a separate verification workflow reconstructed the frozen release artifacts. The result supports a small, bounded graph contribution and technical auditability; it does not claim universal graph superiority, mastery validity, fairness, teacher usability, or improved learning.

Research area
Knowledge tracing, temporal learning systems, graph-aware modeling
My contribution
Research design, implementation, governed evaluation, analysis, visualization, and manuscript
Submitted manuscript · 2026

From Prediction to Educational Consequence: An Evidence-to-Action Framework for Adaptive Assessment

This structured critical narrative review synthesizes psychometrics, computerized adaptive testing, learner modeling, knowledge tracing, sequential decision policies, foundation-model-assisted assessment, human factors, and governance.

I am the first and corresponding author, working with Xin Zhong. The manuscript has been submitted to Studies in Educational Evaluation.

Contribution 1
An integrative six-layer evidence-to-action framework
Contribution 2
An explicit boundary between learner-state inference and adaptive policy
Contribution 3
Claim-specific evidence mapping and recurring invalid inference jumps
Status
Submitted manuscript; not yet a peer-reviewed publication

Research questions

What I am trying to understand.

The questions are technical and human at the same time: what the model represents, what its output means, and what action the evidence can actually support.

Representation

What state should a learner model retain?

Preserve temporal order, concept relationships, and provenance without collapsing history into an opaque score.

Evaluation

Which claim does a metric support?

Distinguish prediction, calibration, validity, fairness, instructional effect, and operational usefulness.

Action

Who decides what happens next?

Keep inference, policy, reporting, human authority, and governance visible as separate functions.

Public artifacts

Related prototypes and applied work.

Public repositories provide implementation context for the broader research direction. Their individual README files remain the source for current technical status.

PythonResearch prototype

Temporal Heterogeneous Graph Knowledge Tracing

An earlier public prototype exploring temporal heterogeneous graphs for student interactions and concept dependencies.

View repository
PythonPersistent AI

Holistic User Map Starter

A FastAPI reference implementation for persistent assistant personalization using compact, continuously updated user models.

View repository
AccessibilityHuman-computer interaction

Click-to-Talk

A voice-driven mouse interaction prototype intended to support people with fine-motor challenges.

View repository

Research practice

Evidence before narrative.

Validity

Do not substitute accuracy for meaning.

Predictive performance alone does not establish mastery, fairness, validity, or improved learning.

Reproducibility

Preserve how the decision was reached.

Keep data boundaries, state transitions, comparisons, and verification artifacts available for inspection.

Boundaries

State what the work does not prove.

Strong research is clearer—not weaker—when its contribution and limits are both explicit.