The 3-Value BVA Question Every AI Model Gets Wrong

3-Value BVA with 4 AI Models

Vitalijs P.

9/8/20264 min read

A few weeks ago I ended up in a genuinely useful rabbit hole with an AI assistant that I think is worth sharing with this community — partly as a fun story, partly as a caution about how we all (myself included) lean on AI for test design questions these days.

How it started

I was working through some boundary value analysis material with Claude, and I fed it a real question from the official ISTQB CTFL v4.0 Sample Exam, Set B (Question 21):

You are testing a form that verifies the correctness of the length of the password given as input. The form accepts a password with the correct length and rejects a password that is too short or too long. The password length is correct if it has between 6 and 12 characters inclusive. Otherwise, it is considered incorrect. At first, the form is empty (password length = 0). You apply boundary value analysis to the "password length" variable. Your set of test cases achieves 100% 2-value boundary value coverage. The team decided that due to the high risk of this component, test cases should be added to ensure 100% 3-value boundary value coverage. Which additional password lengths should be tested to achieve this? a) 4, 5, 13, 14 b) 7, 11 c) 1, 5, 13 d) 1, 4, 7, 11, 14

Claude answered b) 7, 11, and explained it confidently — boundary at 6, boundary at 12, one extra neighbour on each side of the valid partition. Curious whether this was just one model having a bad day, I ran the identical question through ChatGPT, Gemini, and DeepSeek. All three gave the same wrong answer, with the same reasoning. It's also the explanation you'll find in a lot of tutorials, so it's not surprising this is what they converged on.

The official answer is d.

Why the "obvious" answer is wrong

The shortcut most of us use — including, apparently, AI models — treats each transition between two partitions as a single boundary: "the boundary is between 5 and 6, so test 5, 6, and one more on the valid side." That gives you 3 values per transition, and for two transitions (low side and high side), it lands on option b.

The problem is that's not what the syllabus actually says. Here's the definition, word for word, from the ISTQB CTFL v4.0 syllabus:

"The minimum and maximum values of a partition are its boundary values... In 3-value BVA, for each boundary value there are three coverage items: this boundary value and both its neighbors."

That's defined per partition, not per transition. When two partitions sit next to each other, that's two boundary values (each partition's own edge), not one shared point. Applied properly to the password question, there are three partitions: too-short {0–5}, valid {6–12}, too-long {13, 14, …}. Each partition's own min and max is a boundary value in its own right:

  • Boundary 0 (min of "too short," and also the natural floor since length can't be negative) → neighbours: only 1 exists (there's no −1)

  • Boundary 5 (max of "too short") → neighbours 4 and 6

  • Boundary 6 (min of "valid") → neighbours 5 and 7

  • Boundary 12 (max of "valid") → neighbours 11 and 13

  • Boundary 13 (min of "too long") → neighbours 12 and 14

Union all of that and 100% 3-value coverage requires {0, 1, 4, 5, 6, 7, 11, 12, 13, 14}. The 2-value set already covers {0, 5, 6, 12, 13}, so the additional values needed are 1, 4, 7, 11, 14 — exactly option d.

The takeaway: at an interior transition between two partitions, 3-value BVA needs 4 distinct test values, not 3, because each side of the transition is its own independent boundary value with its own pair of neighbours.

I wasn't the first to notice this

Credit where it's due — this isn't a new discovery on my part. Once I had the discrepancy in front of me, I went looking for proves, and found:

  • A second, independently produced ISTQB sample exam (the GTB/German Testing Board edition of CTFL v4.0 Sample Exam Set B) has its own Q21 — a wine-storage temperature scenario instead of a password — and lands on the exact same pattern: 4 values per transition, not 3.

  • ISTQB's own Exam Working Group has published a dedicated white paper on exactly this — "Boundary Value Analysis According to the ISTQB® Foundation Level Syllabus" by Matthias Hamburg and Adam Roman (Roman co-authored the sample exam itself). It names the misconception outright: "It is a common misconception that 2-value BVA requires twice as many coverage items as borders, and 3-value BVA requires three times as many. Both are wrong."

  • Giorgos Valamatsas had already written this up clearly on Medium: "3-value Boundary Value Analysis: Misconception and Reality", stating the rule plainly — 3-value BVA always produces 4 coverage items per boundary pair, not 3 — with his own worked example and a note on how much detection rate you lose by getting it wrong.

So this is a known issue in the community, just apparently not one that's made it into how AI models explain the technique — at least not yet.

Why I'm posting this

Not to dunk on any single AI vendor — this isn't a Claude quirk, or a ChatGPT quirk, or a Gemini or DeepSeek quirk. All four converged on the same wrong answer with the same reasoning, which tells you it's a shared blind spot rather than one model's bug. The real point is this: if you're using AI to help write exam prep questions, sanity-check exercises, or explain a technique to students, and it involves anything with more than one boundary in play, check the answer against the actual syllabus and official sample exams before you trust it. The "obvious" shortcut here is wrong often enough, and confidently enough, that it's worth the extra minute.

Sources:

Contact us

info@dayconsulting.eu

+371 66-555-001

© 2026 DAY CONSULTING