AI-A03 · Module 6 — Assessment Preparation · Lesson 6.1 · Study time: 60 minutes · Language: English

Learning objectives

By the end of this lesson you will be able to:

• State exactly what each assessed component in this course requires and how it is marked.
• Read a rubric as an instruction for construction rather than as a description of quality.
• Diagnose and repair the weaknesses that most often cost marks in applied artificial intelligence.
• Plan and execute a submission under realistic time constraints.

Key terms

Rubric

A published statement of the criteria and standards used to mark. Read as an instruction, it tells you what to build; read as a description, it tells you nothing useful until it is too late.

Criterion

A dimension on which work is judged. This course uses four: accuracy of understanding, quality of evidence, soundness of reasoning and clarity of communication.

Threshold performance

Work that meets the pass standard on every criterion. Note that excellence on one criterion does not compensate for failure on another.

Constructive alignment

The link between outcomes, activities and assessment. Where you cannot see how a task connects to an outcome, ask; that is usually a sign that you have misread the task.

Use-of-AI statement

A short declaration of what generative tools you used, for what, and how you verified the output. Required on every assessed submission in this course.

Lesson

What is assessed and why

This lesson. Assessment Preparation belongs to Module 6 — Assessment Preparation of AI-A03 Multi-Agent and Agentic AI Workflows, taught in the School of Applied Artificial Intelligence. Work through the written lesson first, then the slide deck below, then the activity. Every element on this page is live HTML text: it can be enlarged, printed, searched, read aloud by assistive technology and machine-translated without loss.

The assessment in this course is designed backwards from the learning outcomes. Knowledge checks across the modules confirm that the vocabulary and core distinctions of multi-agent and agentic ai workflows are secure, because everything else depends on them. The analytical brief tests whether you can move from evidence to a stated judgement. The design artefact tests whether you can produce something usable by someone else, which is a different skill from being right. The evaluation report tests critical appraisal, the implementation plan tests realism, and the capstone integrates all of it.

Nothing is assessed for its own sake and nothing is assessed twice. If a component seems to duplicate another, look again at the criterion being emphasised: two tasks may use similar material to test different capabilities. Marks are issued by a human assessor against the published rubric. Automated tools may assist with checks, but no grade in this course is generated automatically and no appeal is decided by a machine.

Reading a rubric as an instruction

Most students read the top band of a rubric, note that it uses words like ‘insightful’ and ‘sophisticated’, and conclude that it describes a quality they either have or lack. That reading is useless. The productive approach is to translate every descriptor into an observable feature of the document you are going to hand in. If the rubric rewards critical engagement with sources, that translates into a countable feature: at least three places in the text where a source is questioned rather than reported. Build the feature and the descriptor follows.

Do this translation before you write, not afterwards. Then, in the final hour before submission, check each translated feature against the draft and mark where it appears. This is not gaming the assessment; it is the intended use of a published rubric, and it is exactly what constructive alignment is for. Students who do it reliably score above students of equal ability who do not.

Where marks are lost

Four failures account for most of the marks lost in applied artificial intelligence at this level. The first is description in place of analysis: an accurate account of what a source says, with no statement of what follows. The second is unsupported assertion: a strong claim about multi-agent and agentic ai workflows with nothing attached to it, usually in the conclusion, where writers relax. The third is definitional drift, where a term is used in one sense on page one and another on page three, quietly destroying the argument in between.

The fourth is presentational and is entirely avoidable: unreadable submissions. Missing references, inconsistent citation, no page numbers, tables without units, and text pasted as images. A marker who cannot read your evidence cannot credit it. This is the same principle that governs the course materials themselves, where every slide deck is delivered as live HTML text rather than as an image precisely so that nothing important can become unreadable.

Working to the deadline

Plan backwards from the submission date and put the finish line three days early. The last three days are not slack; they are for the checks that cannot be done while writing — verifying every citation resolves to the source you claim, confirming each rubric feature is present, reading the whole thing aloud once to catch the sentences that do not parse. Work submitted at the moment of completion is almost always worse than the same work submitted after a short gap.

If you are running late, protect the structure before the prose. A complete piece with a thin section scores better than an elegant fragment, because rubrics are applied criterion by criterion and an absent section scores zero on several at once. Write the judgement, the evidence and the limits first; polish only what remains. And if circumstances genuinely prevent submission, ask before the deadline rather than explaining after it.

Worked example: translating a rubric line

Take a rubric line reading ‘demonstrates critical engagement with the literature’. Translated into constructible features, that becomes four things: at least three sources are compared rather than listed; at least one source is questioned on grounds of method, scope or provenance; a disagreement between two sources is identified and adjudicated with a reason; and no claim central to the argument rests on a single source.

Every one of those can be checked in five minutes with a highlighter, and none requires talent. Now take a second line, ‘communicates with clarity and precision’. That becomes: every paragraph has an identifiable claim in its first or last sentence; no sentence exceeds roughly forty words; every technical term is defined at first use; and every table and figure is readable in black and white with units labelled. Between them, those eight features will move a mid-band submission up a band, and they are all mechanical.

Slide deck

A ten-slide presentation of this lesson, built in HTML so that it prints, translates, enlarges and reads aloud on any device.

12345678910

Click a number to jump to a slide.

Activity

Check your understanding

Q1. Why is reading a rubric as a description of quality less useful than reading it as an instruction?

Because descriptive language such as ‘insightful’ gives you nothing to build. Translating each descriptor into an observable, checkable feature of the submission tells you exactly what to construct and lets you verify it before you hand in.

Q2. Why does an incomplete but structurally complete submission usually beat a polished fragment?

Because rubrics are applied criterion by criterion, and a missing section scores zero on several criteria simultaneously. Thin coverage of every required element outperforms excellent coverage of some.

Q3. What must every assessed submission in this course include regarding AI use?

A use-of-AI statement identifying the tool, the purpose and the verification method. Use is permitted; concealment is a breach of integrity, and passing on an unverified AI claim is a substantive error in its own right.

References and further reading

• Russell, S. and Norvig, P. (2021) Artificial Intelligence: A Modern Approach. 4th edn. Harlow: Pearson.
• Lewis, P. et al. (2020) ‘Retrieval-augmented generation for knowledge-intensive NLP tasks’, Advances in Neural Information Processing Systems, 33.
• Sadler, D. R. (1989) ‘Formative assessment and the design of instructional systems’, Instructional Science, 18(2), pp. 119–144.
• Nicol, D. J. and Macfarlane-Dick, D. (2006) ‘Formative assessment and self-regulated learning’, Studies in Higher Education, 31(2), pp. 199–218.
• Biggs, J. and Tang, C. (2011) Teaching for Quality Learning at University. 4th edn. Maidenhead: Open University Press.
• Boud, D. and Molloy, E. (2013) ‘Rethinking models of feedback for learning’, Assessment and Evaluation in Higher Education, 38(6), pp. 698–712.
• Bearman, M. et al. (2016) ‘Support for assessment practice’, Teaching in Higher Education, 21(5), pp. 545–556.

Use of AI. You may use generative AI to summarise, translate or critique the material in this lesson, provided you verify every factual claim and citation against the sources listed above. Assessed work must carry a use-of-AI statement.