Triangulation

§01 The idea

One capability, seen from several sides

Fawns and colleagues call the kinds of evidence we assess proxies, because each one stands in for a capability you can’t observe directly (p. 4). Triangulation, in their account, means testing an inference about a student against different perspectives on the same capability and looking for coherence and gaps (p. 12). Programmatic assessment calls the same principle saturation.

One demonstration is rarely enough to judge a complex capability. Students also tackle the same task in different ways, so what you get to see depends on the task as each student carried it out. Reliable judgements rest on samples from different situations and from more than one kind of evidence (p. 11), and you can vary the evidence in three directions.

…it is the interaction of proxies, rather than their aggregation, that generates a clearer picture of capability.

§02 What can go wrong

When a source isn’t what you think it is

Triangulation only works if each source is the kind of evidence you take it to be. Of the misalignments Fawns and colleagues describe (pp. 9–11), two bear directly on this.

Sources can also disagree simply because they are different kinds of evidence. In a physics study the authors cite, video of students doing a lab task and the students’ own documentation of their process differed significantly and earned different scores, although the students knew they were being filmed (p. 10).

Across a programme the risk is over-saturation. One kind of evidence, most often submitted work or knowledge exams, is used so often that it tells students, without anyone saying so, what academic success is (pp. 10–11). If every assessment yields the same kind of evidence, triangulation has little to compare.

Our addition. In the Swiss cheese picture from Step 5, slices of the same kind tend to have their holes in the same places.

§03 Your course

A second kind of evidence beside the product

Products dominate everyday assessment (p. 9), so within a course the usual job is to give a product some backing. Fawns and colleagues suggest a performance, such as an oral defence or a discussion of how the work was made (p. 11). The assessment twins in Step 4 work this out in detail.

TEQSA’s 2023 resource has examples where the second source is an account of the process. In one, physics students document their decisions and dead ends. In another, a journalism group that drafts with AI explains its prompts and how the group shaped what the tool produced (Lodge et al., 2023, p. 5). Fawns and colleagues would call such accounts products about process. Written afterwards, they can turn into polished stories that fit what the assessor expects (p. 7), so they count for more when a live performance can test them.

Even the conversation is a sample. One observed performance gives a partial and unstable view (p. 6), so decide in advance what it should show and keep the criteria to that.

Tell students what role AI has in each part. Limits on AI use can only be enforced where someone supervises. Elsewhere they are discursive changes in the sense of Corbin, Dawson and Liu (2025, p. 1087), instructions that students can follow or ignore, as Step 1 explained.

§04 Your programme

Triangulation across a programme

Sampling every aspect of every outcome in every context is neither possible nor desirable. Fawns and colleagues separate outcomes that are core to the qualification, or required by an accrediting body, from the rest. The core ones must be evidenced for each student, and the others can be assured in more personal ways. Across a programme, evidence can build up over different tasks, contexts, assessors and stages, which is sometimes easier and more credible than sampling within one course (p. 11). Some of it can stay formative and still add to the picture (p. 14).

The authors also raise two questions worth taking to colleagues (p. 14). Must every course assess every outcome on its own? Should a student fail a course for missing a single outcome? Nor will securing every course assure the programme’s outcomes if the programme is poorly aligned (Lodge et al., 2025, p. 9).

Then there is the question of who puts the evidence together. Often an algorithm adds up the marks or averages them into a GPA. Students can take part through self- and peer assessment, and in programmatic assessment a team reviews evidence from many points before deciding. When several people judge together they read the evidence more broadly, but unless they calibrate, their decisions can drift apart (p. 12).

When sources disagree, the disagreement shows where the evidence of a capability is incomplete or inconsistent (p. 12). It doesn’t say which source is right. The authors mention an interview at a school that revealed far more of a student’s understanding than the student’s product did, and note that the reverse also happens (p. 9).

Our addition. A mismatch is a reason to collect more evidence, such as a second conversation or another sample, before anyone draws a conclusion. On its own it is never a verdict.

§05 Design

Five steps for your own design

Our addition. The five steps are our summary of the sources on this page. They also assume that students’ use of AI is hybrid. In one assignment a student may work alone, ask AI for help and hand a subtask over, in no fixed order, so a list of tasks to keep and tasks to hand over won’t fit. What stays constant is the student’s own understanding, so look for evidence of it wherever AI comes in, and ask under which conditions you could see it.

§06 Read it for what it is

The limits of these sources

Fawns, Boud and Dawson’s article is conceptual. It proposes a framework and reports no test of it. The authors call their critique “not radical” and anchor it in Australian standards, which already require assessment methods that can confirm that all specified learning outcomes are achieved (p. 14).

Wide sampling takes resources (p. 11), and the cost falls unevenly. Exam administration is mostly paid for by the institution, while observing lab work tends to cost the course itself (p. 13).

More sources can also mean more watching. TEQSA’s 2026 resource asks that process data put students’ agency and privacy first and avoid surveillance, and it says such data are no foolproof way of assuring learning (Lodge et al., 2026, p. 18).