Foundations of Research and Technical Inquiry

Introduction

Research is a systematic process for producing, evaluating, and communicating knowledge. In artificial intelligence and machine learning (AI/ML), research may involve developing a new learning algorithm, testing whether a model generalizes to unseen data, studying how users interact with an intelligent system, or evaluating the social effects of automated decisions. Although AI/ML research often uses mathematics, programming, and statistical analysis, technical competence alone does not make an investigation research. Research requires a defensible question, an appropriate method, credible evidence, transparent reasoning, and responsible communication.

This chapter introduces the intellectual foundations that support later work in research design, data collection, statistical analysis, experimentation, and technical writing. It distinguishes major purposes and types of research, explains how scientific inquiry relates to theory and evidence, and considers why claims of causality require stronger designs than simple descriptions or predictions.

Learning Objectives

By the end of this chapter, students should be able to:

1. Define research and distinguish important research terminology.

2. Explain the purposes of research and the role of paradigms.

3. Describe the scientific method as a flexible cycle of inquiry.

4. Distinguish concepts, constructs, variables, hypotheses, theories, and models.

5. Evaluate the quality and relevance of evidence.

6. Explain correlation, causation, confounding, and mechanisms.

7. Differentiate basic, applied, exploratory, descriptive, explanatory, and evaluative research.

8. Identify ethical, governance, and reproducibility issues in AI/ML research.

9. Apply these ideas to the analysis of a practical research problem.

1. What Research Is—and Is Not

Research is a systematic, transparent, and critical investigation undertaken to develop or test knowledge. Four features are central.

Systematic means that activities follow a reasoned plan rather than relying only on intuition or convenience. Empirical means that claims are connected to observations, measurements, experiences, documents, or other forms of evidence. Critical means that alternative explanations, limitations, and possible errors are considered. Communicable means that methods and conclusions are reported so that others can understand, assess, and, where feasible, reproduce or extend the work.

Research is not simply searching the internet, collecting a large dataset, implementing a complex model, or confirming an existing opinion. A software project can be technically excellent without answering a research question. Conversely, a modest study can be valuable if it addresses a meaningful question with a well-justified design.

A research problem identifies a situation in which knowledge is incomplete, uncertain, contested, or insufficient for a practical decision. A research question converts that problem into an answerable inquiry. For example:

• Problem: A hospital’s diagnostic model performs differently across demographic groups.
• Research question: How does model error vary across patient groups, and which data or modeling factors explain the difference?

A good research question is sufficiently specific to investigate, sufficiently important to justify effort, and sufficiently open that its answer is not already known.

Important Terminology

A research objective states what an investigation intends to accomplish. A research question states what the investigation seeks to answer. A hypothesis is a testable proposition about a relationship or expected outcome. A methodology is the overall logic connecting research assumptions, design, methods, and interpretation. A method is a particular procedure, such as a survey, interview, experiment, observation, or benchmark evaluation. Data are recorded representations of observations or information; evidence is data interpreted in relation to a claim and supported by an appropriate reasoning process.

A population is the complete set of units relevant to a study, while a sample is the subset actually studied. A parameter describes a population, whereas a statistic describes a sample. A variable is a characteristic that can take different values. An independent variable is a presumed influence or input, and a dependent variable is an outcome of interest. In machine learning, these labels must be used carefully: a predictive feature is not automatically a causal independent variable.

Reliability concerns consistency of measurement or procedure. Validity concerns whether a measure or design supports the intended interpretation. A model may produce highly consistent predictions and still be invalid for the intended application if the target variable is poorly defined or the data do not represent real deployment conditions.

2. Purposes and Types of Research

Research categories often overlap. A project can be both applied and explanatory, or exploratory and qualitative. The categories below describe dominant purposes rather than rigid boxes.

Basic and Applied Research

Basic research seeks to improve general knowledge, concepts, or theory without requiring an immediate practical use. For example, studying the mathematical conditions under which a learning algorithm converges is primarily basic research. Its value may become apparent only later.

Applied research addresses a practical problem, often in a particular organizational, technical, or social context. Developing a reliable system for detecting manufacturing defects, or evaluating a fraud-detection model for a bank, is applied research. Applied work still requires generalizable reasoning; a functioning prototype alone does not establish that the approach is effective under relevant conditions.

The distinction is about immediate purpose, not quality. Basic research can have practical consequences, and applied research can contribute to theory.

Exploratory Research

Exploratory research investigates a topic that is poorly understood or insufficiently specified. It may identify relevant variables, generate hypotheses, clarify user needs, or reveal unexpected patterns. Interviews with data scientists about barriers to reproducible experimentation, or an initial analysis of an unfamiliar dataset, may be exploratory.

Exploration is especially useful at the beginning of a project. However, exploratory findings should not automatically be presented as confirmed conclusions. A pattern discovered while inspecting data may be due to chance, selection bias, or hidden preprocessing decisions and should normally be tested in later work.

Descriptive Research

Descriptive research documents characteristics, distributions, behavior, or conditions. Examples include estimating the proportion of online transactions flagged by a system, describing the demographic composition of a dataset, or reporting the latency and accuracy of a deployed model.

Description answers questions such as “What is happening?”, “Who is affected?”, “How often does it occur?”, and “Under what conditions is it observed?” It does not, by itself, establish why an outcome occurs.

Explanatory Research

Explanatory research examines relationships and mechanisms to answer “why” or “how” questions. It may test whether a training intervention improves programmer performance, or whether a change in feature representation causes a reduction in classification error.

Explanation generally requires stronger control of alternative explanations than description. Researchers must consider temporal order, confounding variables, measurement quality, and plausible mechanisms.

Evaluative Research

Evaluative research judges the value, effectiveness, efficiency, usability, fairness, or consequences of a program, intervention, product, or policy against explicit criteria. An evaluation of an AI tutoring system might assess learning gains, accessibility, cost, user satisfaction, and unintended effects.

Evaluation should define success before examining results where possible. It should also consider comparison conditions, implementation fidelity, stakeholder perspectives, and whether outcomes justify adoption in the intended context.

3. Research Paradigms and Worldviews

A research paradigm is a broad set of assumptions about reality, knowledge, values, and appropriate methods. Paradigms influence what researchers treat as a legitimate question and what counts as convincing evidence. They are not merely labels for techniques.

Positivist and Post-Positivist Orientations

A positivist orientation emphasizes observable phenomena, measurement, regularities, and systematic testing. In contemporary social and applied research, a post-positivist position is more common: it accepts empirical testing but recognizes that observations and measurements are fallible, theories are provisional, and complete objectivity is difficult.

Quantitative experiments, benchmark comparisons, and statistical estimation are often compatible with this orientation. The strength of a result depends not only on numerical significance but also on measurement validity, design quality, practical importance, and generalizability.

Interpretivist and Constructivist Orientations

Interpretivist and constructivist approaches emphasize how people create meaning within social and cultural contexts. Interviews, ethnography, document analysis, and participatory observation may be appropriate when the goal is to understand experiences, practices, or interpretations that cannot be reduced adequately to predefined variables.

For example, a survey may show that users distrust an automated decision system, but interviews may reveal whether distrust results from poor explanations, previous institutional experiences, or fear of surveillance.

Pragmatism and Mixed Methods

Pragmatism emphasizes the research problem and the usefulness of knowledge for addressing it. It permits quantitative, qualitative, or mixed methods when their combination provides a stronger answer. A study of an AI hiring tool might combine statistical fairness analysis with interviews of applicants and recruiters.

Mixed methods are not simply the addition of two datasets. The researcher must explain how the strands connect, whether one informs the other, and how apparently conflicting findings will be interpreted.

Critical and Participatory Perspectives

Critical approaches examine power, inequality, exclusion, and the consequences of apparently neutral systems. They ask who defines the problem, whose interests are served, who bears risks, and whose knowledge is ignored. Participatory research involves affected stakeholders in defining questions, interpreting evidence, or designing interventions. These perspectives are particularly relevant to AI systems used in employment, education, health, policing, and public services.

4. The Scientific Method as a Cycle of Inquiry

The scientific method is not a single rigid sequence. It is a family of disciplined practices built around question formulation, observation, reasoning, testing, criticism, and revision. A typical cycle includes:

1. Problem identification: recognize a gap, anomaly, need, or decision problem.

2. Background investigation: examine prior research, concepts, theories, and relevant data.

3. Question formulation: state precise research questions and, where appropriate, hypotheses.

4. Design: select participants, data sources, measures, comparison conditions, and procedures.

5. Data collection or generation: obtain evidence while documenting procedures and deviations.

6. Analysis: use appropriate qualitative, statistical, computational, or comparative techniques.

7. Interpretation: connect results to the question, theory, context, and limitations.

8. Communication and scrutiny: report methods and findings for review, replication, criticism, or practical use.

9. Revision: refine explanations and identify further questions.

The cycle is iterative. Literature review may change the question; pilot testing may reveal that a measure is unsuitable; unexpected findings may motivate a new hypothesis. Such changes should be documented rather than concealed.

A hypothesis is not required for every project. Exploratory and interpretive studies may begin with sensitizing concepts or open questions. What is required is a clear relationship between the question, the evidence, and the reasoning used to reach conclusions.

5. From Concepts to Theory and Models

Concepts and Constructs

A concept is an abstract idea, such as fairness, trust, intelligence, usability, or reproducibility. A construct is a concept deliberately defined for research. Constructs are not directly observed; they are inferred through indicators or measures.

Operationalization specifies how a construct will be represented in a study. “Model fairness” might be operationalized as equality of a selected error rate across groups, but that choice captures only one aspect of fairness. Different operationalizations can produce different conclusions. Researchers must therefore justify definitions rather than treating a measurement choice as self-evident.

Variables, Relationships, and Hypotheses

A hypothesis may state that an intervention is associated with an outcome, that one condition produces a different result from another, or that a relationship varies across groups. A useful hypothesis is clear, conceptually grounded, and testable with available evidence.

For example: “Providing uncertainty explanations will improve users’ calibration of trust in an AI recommendation system compared with providing recommendations without explanations.” This statement identifies an intervention, an outcome, and a comparison. The researcher must still define “uncertainty explanations,” “calibration of trust,” the user population, and the conditions under which the claim is expected to hold.

Theories and Models

A theory is an organized set of concepts and propositions that explains patterns and supports predictions. A theory is not merely a guess; in academic usage it is a structured explanatory account that has received support, although it remains open to criticism and revision.

A model is a simplified representation of a system, process, or relationship. Models may be mathematical, computational, graphical, or conceptual. A predictive model can estimate outcomes without explaining the mechanisms that produce them. A causal model represents assumptions about how variables influence one another. A conceptual framework maps important ideas and their proposed relationships.

Theory building may be deductive, beginning with theory and deriving hypotheses for testing, or inductive, developing concepts and explanations from observations. Abductive reasoning seeks the most plausible explanation for surprising evidence and often moves iteratively between data and theory. Strong research makes clear which form of reasoning is being used.

6. Evidence, Inference, and Quality

Evidence supports a claim only when it is relevant, sufficiently accurate, and interpreted through an appropriate design. A large dataset does not automatically provide strong evidence. Quality depends on several dimensions:

• Validity: whether the study supports the intended conclusion.
• Reliability: whether procedures and measurements are consistent.
• Credibility: whether the account is convincing given the methods and context.
• Precision: how narrowly an estimate is determined.
• Transparency: whether methods, data decisions, code, and limitations are documented.
• Relevance: whether the evidence addresses the actual research question.
• Generalizability or transferability: whether findings apply beyond the studied setting, and under what conditions.

Researchers should distinguish a finding from an interpretation and an interpretation from a recommendation. If a model has lower error on one benchmark, the finding concerns that benchmark and procedure. It does not automatically prove superiority in deployment, fairness, or social benefit.

Quantitative and Qualitative Evidence

Quantitative evidence uses numerical measurement and may support estimation, comparison, or statistical testing. Qualitative evidence includes language, images, observations, documents, and participant accounts analyzed for meaning, themes, processes, and context. Neither form is inherently superior. The appropriate form depends on the question.

Statistical significance, when used, addresses compatibility of data with a specified model or null hypothesis under assumptions. It does not by itself establish practical importance, truth, or causality. Researchers should report effect sizes, uncertainty, design limitations, and relevant subgroup results where appropriate.

In AI/ML, evaluation should normally separate training, validation, and test information. Test data must not influence model selection or tuning if it is to provide an unbiased estimate of final performance under the stated assumptions. Data leakage, duplicate records, inappropriate preprocessing, and distribution shift can make reported performance misleading.

7. Causality and Explanation

Causality concerns what would happen under a change or intervention. A causal claim is stronger than a statement that two variables are associated. If students who use an AI tutor obtain higher scores, several explanations are possible: the tutor may help; motivated students may be more likely to use it; prior ability may influence both use and scores; or the groups may differ in instructor support.

Three ideas are especially important:

1. Association: variables vary together.

2. Temporal order: the presumed cause precedes the effect.

3. Non-spuriousness: the relationship is not adequately explained by confounding or bias.

A confounder is a variable related to both the presumed cause and outcome that can create or distort an association. Randomized experiments reduce confounding by assigning conditions using a random process, subject to compliance and other limitations. When randomization is impossible, researchers may use matching, statistical adjustment, natural experiments, longitudinal designs, or causal diagrams. These methods rely on assumptions that must be stated and assessed.

A causal mechanism describes how an influence produces an outcome. For example, an explanation interface might change user behavior by improving understanding of uncertainty, which then changes reliance decisions. Mechanistic reasoning strengthens explanation but does not eliminate the need for evidence.

A common error in AI research is to interpret predictive importance as causal influence. A feature can improve prediction because it is a proxy for a cause, an effect, or an artifact of data collection. Prediction and intervention answer different questions.

8. A Detailed Illustrative Case: Evaluating an AI Triage Assistant

The following is a fictional illustrative case.

Background and Problem

A regional hospital introduces an AI assistant that assigns an urgency score to incoming patient records. The hospital reports that the system reduces average triage time. Administrators want to know whether it should be deployed more widely and whether it performs equitably.

Research Questions

The evaluation team formulates four questions:

1. Does the assistant reduce triage time compared with the existing process?

2. Does it improve the accuracy of urgency classification?

3. Does performance differ across patient groups?

4. What factors affect clinicians’ use and trust of the system?

These questions combine evaluative, explanatory, descriptive, and qualitative purposes.

Analysis and Approach

The team defines triage time precisely, specifies the clinical reference standard, and identifies relevant patient groups before analysis. It conducts a controlled rollout in which some clinics use the assistant while others continue with the existing procedure. Assignment is made at the clinic-week level to reduce contamination between clinicians. The team records patient mix, staffing, workload, system recommendations, clinician decisions, and outcomes.

For accuracy, it reports sensitivity, specificity, and calibration rather than a single accuracy value. These metrics answer different questions and may vary with the threshold used to classify urgency. The team examines confidence intervals or other uncertainty estimates and performs subgroup analyses. It also audits whether missing data, language differences, or unequal documentation affect the results.

Interviews with clinicians explore whether the tool is ignored, over-trusted, or used as a prompt for further review. Researchers maintain an audit trail for changes to the protocol and prevent outcome information from being used improperly during model evaluation.

Possible Outcome

Suppose the assistant reduces median triage time but does not improve overall classification accuracy. It improves recognition of some urgent cases but produces more false alarms for another group. Interviews show that clinicians interpret a high score as authoritative because the interface does not communicate uncertainty.

The appropriate conclusion is not simply that the system “works” or “fails.” It may provide operational benefits while creating safety and equity risks. Deployment would require threshold review, subgroup monitoring, clinician training, an appeals or override process, and further investigation of the unequal error pattern.

Learning Points

The case demonstrates that evaluation requires explicit criteria, suitable comparisons, multiple outcome measures, attention to context, and stakeholder perspectives. A technically accurate model can still be unsuitable if its errors are distributed unfairly or if users misunderstand its outputs. It also demonstrates why research questions should be separated: time savings, predictive performance, equity, and user experience are related but distinct outcomes.

9. Ethics, Governance, and Responsible Inquiry

Research involving people, personal data, or consequential automated decisions requires ethical review appropriate to the institution and jurisdiction. Core principles include respect for persons, informed consent where applicable, minimization of harm, fair participant selection, privacy protection, and accountability.

AI/ML research raises additional concerns. Datasets may contain sensitive information, historical discrimination, copyrighted material, or records collected without clear permission for the proposed use. De-identification reduces some risks but does not guarantee anonymity, especially when datasets can be linked. Researchers should apply data minimization, access controls, retention limits, security safeguards, and documented provenance.

A responsible study should consider:

• whether participants understand the purpose and risks;
• whether vulnerable groups require additional protection;
• whether the system could cause discriminatory or unsafe outcomes;
• whether researchers have conflicts of interest;
• whether negative results and protocol deviations will be reported;
• who is accountable for decisions made using the research output.

Governance is not limited to compliance. Ethical reflection should influence the research question, data selection, evaluation criteria, deployment decision, and communication strategy. Claims should be proportionate to evidence, especially when results may affect public policy or individual opportunities.

10. Reproducibility and Contemporary Research Practice

Reproducibility means that an analysis can be repeated using the same materials and procedures, while replicability often refers to obtaining consistent findings in a new study. Exact terminology varies across disciplines, but both ideas support trustworthy research.

Best practices include preregistering hypotheses when suitable, keeping version-controlled code, documenting data transformations, recording software and hardware environments, separating exploratory from confirmatory analyses, using fixed evaluation protocols, and sharing data or synthetic alternatives when legally and ethically possible. Reproducibility does not mean that every study must release sensitive data. It means that limitations to verification should be clearly explained and that sufficient information should be provided for meaningful scrutiny.

Contemporary AI research also requires attention to benchmark limitations, foundation-model evaluation, distribution shift, energy and computational cost, human-AI interaction, and the difference between laboratory performance and real-world impact. A benchmark is an instrument, not a complete definition of intelligence or usefulness. Researchers should ask what the benchmark measures, what it omits, and whether optimization against it creates undesirable behavior.

Best Practices for Beginning a Research Project

A student or professional beginning an investigation should:

1. State the practical or scholarly problem in precise terms.

2. Conduct a focused literature review and identify what remains uncertain.

3. Define key constructs and operational measures.

4. Match the research design to the question rather than choosing a fashionable method first.

5. Identify plausible alternative explanations and sources of bias.

6. Plan data management, ethics, and security before collection.

7. Pilot instruments, code, or procedures where feasible.

8. Separate training, development, and evaluation information in ML studies.

9. Document decisions, exclusions, preprocessing, and deviations.

10. Report limitations and avoid claims broader than the evidence permits.

Chapter Summary

Research is a systematic, empirical, critical, and communicable process for developing or evaluating knowledge. Its quality depends on alignment among the problem, research question, paradigm, design, evidence, analysis, and conclusion. Basic research advances general understanding, whereas applied research addresses practical problems. Exploratory research clarifies unfamiliar topics; descriptive research documents conditions; explanatory research investigates relationships and mechanisms; and evaluative research judges interventions against explicit criteria.

Paradigms shape assumptions about reality, knowledge, values, and suitable methods. Quantitative, qualitative, mixed, critical, and participatory approaches can all produce valuable knowledge when appropriately designed. Concepts and constructs must be operationalized carefully, because measurements represent only selected aspects of abstract ideas. Theories organize explanations, models simplify systems, and hypotheses connect concepts to testable expectations.

Evidence must be relevant, valid, reliable, transparent, and interpreted within the limits of the design. Association does not establish causation. Causal reasoning requires attention to temporal order, confounding, alternative explanations, and mechanisms. In AI/ML, predictive performance is not equivalent to causal influence, fairness, safety, or social value. Ethical governance, privacy, reproducibility, stakeholder involvement, and responsible communication are integral parts of technical inquiry.

Key Terms

Research: A systematic, critical, and communicable investigation that develops or evaluates knowledge.

Research problem: A situation involving an important gap, uncertainty, need, or decision.

Research question: A focused question that an investigation is designed to answer.

Methodology: The logic connecting assumptions, research design, methods, and interpretation.

Paradigm: A set of assumptions about reality, knowledge, values, and inquiry.

Construct: An abstract concept defined for research and represented through indicators or measures.

Variable: A characteristic that can take different values.

Hypothesis: A testable proposition about an expected relationship or outcome.

Theory: An organized set of concepts and propositions that explains patterns and supports predictions.

Model: A simplified representation of a system, process, or relationship.

Evidence: Data interpreted in relation to a claim through an appropriate reasoning process.

Validity: The extent to which a measure or design supports the intended interpretation.

Reliability: The consistency of a measurement or procedure.

Confounder: A variable related to both a presumed cause and an outcome that can distort their association.

Causality: A relationship in which a change or intervention produces a difference in an outcome under stated conditions.

Basic research: Research primarily intended to advance general knowledge or theory.

Applied research: Research intended to address a practical problem.

Exploratory research: Research that investigates an insufficiently understood topic and often generates hypotheses.

Descriptive research: Research that documents characteristics, distributions, or conditions.

Evaluative research: Research that judges the effectiveness, value, or consequences of an intervention or program.

Review Questions

Short Answer Questions

1. What features distinguish research from ordinary information gathering?

2. Define a construct and explain why operationalization is necessary.

3. What is the difference between descriptive and explanatory research?

4. Why does association alone not establish causality?

5. State two ethical issues relevant to AI/ML research involving personal data.

Long Answer Questions

1. Compare basic, applied, exploratory, descriptive, explanatory, and evaluative research, using an AI/ML example for each.

2. Explain how positivist, interpretivist, pragmatic, and critical perspectives may lead researchers to study the same AI system differently.

3. Describe the scientific method as an iterative cycle and explain why research questions may change during a study.

4. Discuss the relationship among concepts, constructs, variables, hypotheses, theories, and models.

5. Explain how validity, reliability, transparency, and generalizability influence the credibility of research evidence.

Analytical / Application Questions

1. A model predicts employee turnover with high accuracy using an employee’s number of training sessions. Design a study to determine whether additional training reduces turnover, and identify at least two possible confounders.

2. An AI education tool improves average examination scores but produces larger benefits for already high-performing students. What further evidence would you collect before recommending adoption?

3. A benchmark shows that a new language model outperforms an older model. Analyze at least four reasons why this result may not demonstrate superior real-world usefulness.

Multiple Choice Questions with Answers

Question 1

Which feature most clearly distinguishes research from ordinary information gathering?

A. It uses a large amount of information

B. It follows a systematic and critical process

C. It is conducted using computers

D. It produces a numerical result

Correct Answer: B. It follows a systematic and critical process

Question 2

What is the primary purpose of exploratory research?

A. To estimate a population parameter with maximum precision

B. To establish a final causal law

C. To investigate an insufficiently understood topic and generate ideas

D. To assess compliance with a regulation

Correct Answer: C. To investigate an insufficiently understood topic and generate ideas

Question 3

Which statement best describes a construct?

A. A directly observed raw value

B. An abstract idea defined for research and represented through measures

C. A statistical test result

D. A complete population of interest

Correct Answer: B. An abstract idea defined for research and represented through measures

Question 4

Which statement about a theory is most accurate?

A. A theory is an organized explanatory account that remains open to testing and revision

B. A theory is an untested personal opinion

C. A theory is identical to a dataset

D. A theory is always a mathematical equation

Correct Answer: A. A theory is an organized explanatory account that remains open to testing and revision

Question 5

Why does a correlation between AI-tool use and student performance not by itself establish causality?

A. Correlation can never be calculated for educational data

B. The variables must always be measured qualitatively

C. Confounding and selection effects may explain the association

D. Causality requires a larger vocabulary

Correct Answer: C. Confounding and selection effects may explain the association

Question 6

Which research type primarily judges the effectiveness or consequences of an intervention?

A. Evaluative research

B. Exploratory research

C. Descriptive research

D. Basic research

Correct Answer: A. Evaluative research

Question 7

In an ML study, why should test data generally not be used for model tuning?

A. It makes the dataset qualitative

B. It can produce an optimistically biased estimate of final performance

C. It prevents any model from being trained

D. It eliminates the need for validation data

Correct Answer: B. It can produce an optimistically biased estimate of final performance

Question 8

Which example is primarily descriptive research?

A. Testing whether an interface causes greater trust

B. Developing a theory of algorithmic accountability

C. Reporting error rates across demographic groups

D. Determining whether a policy should be adopted

Correct Answer: C. Reporting error rates across demographic groups

Question 9

What is a confounder?

A. A variable related to both a presumed cause and an outcome that can distort their association

B. A measurement with perfect reliability

C. A participant who refuses consent

D. A model that has no input variables

Correct Answer: A. A variable related to both a presumed cause and an outcome that can distort their association

Question 10

Which practice most directly supports reproducible AI research?

A. Changing evaluation criteria after seeing results without documentation

B. Reporting only the best experimental run

C. Documenting code, preprocessing, versions, and evaluation procedures

D. Removing all negative findings from the report

Correct Answer: C. Documenting code, preprocessing, versions, and evaluation procedures