Professional Mini-Master in Evaluation of Generative AI System Trustworthiness

Program facts

Study fee

Free

subjects

5

lessons

15

Estimated time

46 hours

Content last reviewed

23 October 2026

Access period

60 days

Study language

Arabic and English

About the program

This program builds an auditable evaluation plan for a defined generative system. It starts with a service claim and the boundaries of text, retrieval and tools, then designs withheld cases and reference judgments. It measures answer support, abstention, slices and uncertainty while testing injection, disclosure and disparity without confusing attempts with effects. Learners set prior acceptance gates, calibrate an evaluator and defend a bounded release with owners and reopening conditions. The final module covers production monitoring with valid denominators, controlled change tests, incident response and reauthorization. Regional examples and numbers are fictional teaching cases. The completion credential is internal, not an academic degree or external accreditation.

Practical relevance

You will produce an evaluation charter, withheld-set matrix, adjudication guide, denominator table, safe injection plan, acceptance rule, release memo, monitoring specification, incident timeline and reauthorization conditions. Exercises include worked results and explicit limits.

Why this program

Many product teams have score dashboards but lack a clear chain from use claim to test evidence and release decision. This program trains you to spot averages hiding exceptions, judges rewarding polished unsupported prose, and boundaries that quality gains cannot offset. It connects prerelease evaluation to what happens after a document or tool changes, so teams know when to retest and when to withhold a function.

Career relevance

Designed for model evaluation, quality assurance, application security, product and risk teams making operational decisions about generative assistants. The emphasis is on reviewable evidence memos rather than broad scores.

What you will learn

• Define the service promise, system under test, withheld cases and reference judgments.
• Measure claim support, uncertainty, slice gaps and version regression.
• Test safety, prompt injection and disclosure in isolation.
• Write acceptance gates and a conditional release decision with named owners.
• Operate reviewable monitoring, change, incident response and reauthorization.

Who it is for

• Professionals evaluating generative models and products who review answer quality and tool boundaries, plus quality, application security and risk teams proposing release decisions. It suits readers who can inspect case traces and simple rate tables, discuss error consequences with a service owner, and want to move from aggregate scores to actionable evidence for a clearly bounded deployment.

Curriculum

1. Subject 1 Open for free

   Define the Measurement Target and System Boundary

   A generative system trustworthiness plan starts with a bounded claim about an operating version, not an impression of an isolated model. Identify model, instructions, index, tools, server, interface, users and allowed actions, then connect each claim to evidence and a predeclared decision. Keep answer correctness separate from action safety: an unauthorized record disclosure triggers containment even when average answer accuracy is high. Build a sampling frame that distinguishes routine use, exceptions and adversarial challenges. Preserve numerator, denominator and weight for every slice without presenting assumed use weights as observed production. Protect the holdout from leaked case families and obsolete documents. Write dated reference judgments and decisive conditions before seeing responses; record reviewer disagreement and adjudication. Report unjudgeable cases as a coverage limitation rather than silently removing them from the evidence.

   3 lessons · 8.5 hours

   1. From a trustworthiness claim to a bounded test target 3 hours
   2. Build a test set for use and risk 3 hours
   3. Adjudicate reference answers and reviewer disagreement 3 hours
2. Subject 2

   Measure Accuracy and Reliability

   Evaluating generative-system accuracy separates correct outcome, claim-level grounding, completeness of a decisive condition and appropriate abstention. Name numerator, denominator and counting unit for every rate: task success among answerable questions differs from correctness among responses given and abstention on unsupported questions. Blanket refusal is not rewarded because it destroys useful coverage. Use an uncertainty interval such as Wilson's when observations are independent and reference judgments valid; the interval cannot repair selection bias or claims clustered within a session. Compare versions on paired cases and show fixes and regressions rather than only an overall difference. Test wording changes that should preserve a judgment and fact changes that should alter it. Another person's data disclosure or an unauthorized action remains a separate containment gate despite a higher average answer score.

   3 lessons · 8.5 hours

   1. Measure grounding, completeness and abstention 3 hours
   2. Estimate rates and uncertainty across slices 3 hours
   3. Compare versions and test robustness 3 hours
3. Subject 3

   Test Safety, Security and Fairness

   Safety evaluation ties each hazard to service scope, consequence and a preventive control. Contrast a permitted request with a nearby out-of-scope one so blanket refusal cannot masquerade as success; measure permitted coverage and safe out-of-scope handling with separate denominators. Retrieved text in a prompt-injection test supplies knowledge, not authority over tools. Observe the model request, tool, server decision, returned data and display. A server-denied attempt is not disclosure, while another account's synthetic canary displayed to an unauthorized user inside isolation requires tool containment. Use synthetic identities and records. Analyze answer quality by language and case type, retaining numerators, denominators and document equivalence; an observed gap alone has no proven cause. A quality-rate change never cancels the independent disclosure gate.

   3 lessons · 8.5 hours

   1. Classify safety hazards and appropriate refusal 3 hours
   2. Prompt injection and tool trust boundaries 3 hours
   3. Detect disclosure and performance differences across groups 3 hours
4. Subject 4

   Acceptance Decisions and Trade-offs

   Acceptance starts with rules declared before testing: effect gates that an average cannot offset, and performance thresholds with explicit denominators, weights and scope. Weighted answer quality cannot lift a cross-account reading restriction. An answer-only scope can be considered only after isolation and source validity are demonstrated. An automated judge needs calibration on supported and unsupported answers, independent human review, and adjudication of conflicting documents against their effective dates. Reviewer agreement alone does not validate a reference. A decision memo records the system version, service promise, each gate outcome, sample limits and sensitivity to assumptions. It names permitted and withheld functions, repair and monitoring owners, and reopening evidence. Unverified does not silently become passed; a conditional release must be enforceable and reviewable.

   3 lessons · 8.5 hours

   1. Hard acceptance gates and bounded trade-offs 3 hours
   2. Evaluator validity and conflicting evidence 3 hours
   3. Conditional release and residual-risk ownership 3 hours
5. Subject 5

   Monitoring and Continuous Evaluation

   Postrelease monitoring measures use, adjudicated quality, safety effects and measurement-pipeline health with separate denominators. Voluntary ratings describe respondents, targeted exception samples cannot be naively pooled with probability samples, and delayed labels remain pending. Record model, prompt, index, tool and judge versions to distinguish request-mix change from system or reference change. After a component update, select regression tests for its effect path and compare versions on pinned questions and sources while preserving individual regressions. Following observed disclosure, contain the path and verify blocking before causal investigation; preserve evidence and exposure limits. Restoration is a new decision requiring demonstrated repair, prevention, legitimate-use and alert tests on the same version, plus a named acceptor. If a decisive condition remains unverified, the affected function stays withheld or technically restricted.

   3 lessons · 8 hours

   1. Production monitoring with valid denominators 3 hours
   2. Change triggers and regression testing 3 hours
   3. Incident response and reauthorization 2.5 hours

Subjects unlock in sequence as you progress.

Strengths

• A complete decision chain from use claim to reauthorization.
• Calculations, slice diagnostics and effect gates with worked solutions.
• Parallel Arabic and English instruction with bounded fictional cases.

Evidence basis

Measurement and risk concepts draw on primary publications from NIST, MLCommons and OWASP and a published ACL paper. Example thresholds and fictional cases are teaching decisions, not official recommendations from those publishers.

Program questions

Certificate

Verification example: a real issued certificate