Research Articles — PROTEX
RESEARCH ARTICLES — AI EVALUATION, KNOWLEDGE ARCHITECTURE AND CONTROLLED AI SYSTEMS

This page collects selected research papers developed within the PROTEX programme.

The publications form the methodological foundation for PROTEX as a research, teaching, and AI evaluation environment. They examine how knowledge can be structured, retrieved, evaluated, and used within AI systems operating under explicit constraints and defined epistemic boundaries.

The work supports a broader investigation into trustworthy AI systems, structured retrieval architectures, Human-AI Collaboration, knowledge-based decision support, and new methods for evaluating AI-generated answers beyond simple correctness.

Selected publications
Adaptive Benchmark Generation for Knowledge-Based AI Evaluation

A methodological framework for generating adaptive benchmarks for knowledge-based AI systems, supporting systematic evaluation across changing knowledge environments, task complexity, and retrieval conditions while extending AI assessment beyond fixed question sets and static benchmarks.

Beyond Procedures: A Scenario-Based Framework for Assessing Operational Decision-Making

A scenario-based assessment framework for examining professional judgement and operational decision-making under ambiguity, incomplete information, competing priorities, and contextual constraints. The framework shifts assessment beyond procedural recall towards the reasoning processes underlying real-world operational decisions.

Quantitative Evaluation of Native Microsoft Copilot Studio on the PROTEX Behavioural Homicide Corpus

A 200-question benchmark evaluating factual retrieval, comparative reasoning, false-premise handling, uncertainty preservation, and semantic contamination resistance within a structured behavioural knowledge corpus. This study also serves as a teaching example for demonstrating how AI evaluation can move beyond simple question-answer accuracy.

Epistemic Corpus Design and Retrieval Stability in Enterprise AI: A Case Study of PROTEX Migration to Microsoft Copilot Studio

A study examining retrieval stability, knowledge structure, uncertainty preservation, and behavioural consistency following migration from a custom retrieval architecture to native Microsoft Copilot Studio.

Beyond Retrieval Accuracy: Evaluating Enterprise AI Knowledge Retrieval Systems Across Organisational Knowledge Environments

A comparative evaluation of enterprise AI knowledge retrieval across organisational knowledge environments, examining retrieval fidelity, evidence adherence, omission risk, consistency, and knowledge architecture as determinants of trustworthy AI performance.

RAG System Design

A methodological exploration of retrieval-augmented generation systems based on structured knowledge access, controlled context, evidentiary grounding, and transparent response generation.

Hybrid Retrieval Architecture

A study of combined semantic and deterministic retrieval approaches, focusing on control, filtering, contextual precision, metadata-based routing, and epistemic constraints in AI-supported analytical systems.

Case Description Framework

A framework for structuring complex case material into distinct analytical layers, preserving the separation between factual observation, behavioural description, interpretation, uncertainty, and narrative framing.

Procedural AI Assistants

A study of AI systems designed for procedural and operational environments, focusing on bounded system roles, structured organisational knowledge, decision support, and responsible human oversight.

Research, teaching and evaluation context

These publications support the development of PROTEX as a complete research and teaching environment for knowledge-based AI systems.

Together, they show how an AI project can move through a full methodological cycle: knowledge architecture design, structured corpus development, system implementation, retrieval control, benchmark construction, AI evaluation, and educational demonstration.

This makes the research useful not only for system design, but also for lectures, workshops, seminars, and practical demonstrations on AI evaluation, RAG systems, knowledge architecture, AI governance, and human-centred decision support.

The central idea is that AI systems should not be treated only as open-ended generators. They can also be designed, tested, and taught as structured systems grounded in knowledge quality, source traceability, uncertainty modelling, explicit constraints, and measurable reliability.

Przewijanie do góry