A wrong answer is data, but only if you know what kind of wrong it was. Most candidates review a missed question by re-reading the explanation, which fixes only one of the twelve distinct reasons a question is actually missed. This is the error-analysis pillar of our framework series: a twelve-cause taxonomy that names each reason you lose a mark and pairs it with the one correct intervention, plus a medical exam question error log you can run on any bank, in any jurisdiction, this week. The practical outcome is that your revision time goes to the fix that matches the fault, instead of the fix that feels productive.
Why the same mistake gets the wrong fix on every exam
The default review habit treats every wrong answer as a knowledge gap, because re-reading the explanation is what a bank makes easy. But a misread lead-in, a pacing slip, and a genuine gap in what you know need three completely different responses, and applying the knowledge fix to a technique fault is one of the quietest ways to waste a preparation. The surface of the exam changes across jurisdictions; the error families do not. Whether you are sitting MRCP(UK) Part 1 — two three-hour papers of best-of-five items with no negative marking — USMLE Step 2 CK with its long next-step vignettes across eight one-hour blocks, or MCCQE Part I, which since April 2025 is multiple-choice throughout, the same twelve causes account for almost every mark you drop. That is the whole reason a taxonomy is worth building: it is portable in a way that any single exam's advice is not, and it turns "I got it wrong" into "I got it wrong for reason nine, so I do the reason-nine thing." We avoid pass-rate claims here deliberately; the point is not to predict outcomes but to route each error to its correct intervention.
The twelve-cause error taxonomy
The taxonomy sorts the reasons you miss questions into four families — knowledge, comprehension, reasoning, and execution — and gives each a single correct intervention. Copy this table into your notes; it is the embeddable core that every exam-specific guide in this series links back to.
| # | Error type | What it looks like | The correct intervention |
|---|---|---|---|
| 1 | Knowledge gap | You never encoded the fact at all | Encode it properly: a cited note plus spaced retrieval; add the topic to your coverage plan |
| 2 | Knowledge decay | You knew it once and forgot | Space it, do not relearn from scratch: schedule expanding-interval retrieval |
| 3 | Outdated knowledge | You applied a superseded guideline | Re-anchor to the current, dated source (NICE, CKS, SIGN, SmPC/eMC, or the relevant national guidance) |
| 4 | Lead-in misread | You answered a different question than the one asked | Stem-parsing drill: read the last line first, underline the exact task |
| 5 | Missed stem detail | You overlooked a value, age, timing or unit | Extract the data before reading the options; slow the first pass |
| 6 | Data-integration failure | You had the pieces but did not synthesise them | Force a one-line synthesis before answering; practise multi-data items |
| 7 | Premature closure | You anchored on the first plausible diagnosis | Generate a differential first; drill discriminating features |
| 8 | Wrong next step | Right diagnosis, wrong management or sequence | Practise "what is first / what changes management" framing |
| 9 | Distractor capture | You were pulled to a plausible wrong option | Run a why-each-option-is-wrong drill; name the trap |
| 10 | Pacing fault | You rushed, ran out of time, or stalled | Timed practice at exam pace; a triage-and-flag routine |
| 11 | Avoidable slip | Misclick, unit or calculation error, changed a right answer | A short checking routine; unit discipline; log answer changes |
| 12 | Flawed or out-of-scope item | The question is wrong or ambiguous, not you | Verify against the official source; if flawed, discard and do not over-learn |
The value of the four families is that they map to four different weeks of work. Knowledge faults (1–3) are content problems solved by encoding and spacing. Comprehension faults (4–6) are reading problems solved by stem discipline. Reasoning faults (7–9) are clinical-thinking problems solved by differential and management practice. Execution faults (10–12) are technique and hygiene problems solved by pacing, checking, and honest triage of bad items. A candidate whose misses cluster in one family has a very different next fortnight from one whose misses are spread evenly, and no completion percentage will ever tell you which you are.
The error log that makes the taxonomy usable
A taxonomy you do not record is just a nice idea. The instrument that operationalises it is a medical exam question error log: one row per missed item, filled in the moment you review it, so that by the end of a week you can tally your errors by type and read your intervention list straight off the page.
| Item ref and date | Blueprint domain | Error type (1–12) | My answer to correct answer | Root-cause note (one line) | Intervention chosen | Re-test date | Re-test result |
|---|---|---|---|---|---|---|---|
| e.g. Q4821, 12 Jul | Cardiology | 8 | Chose PCI to chose thrombolysis eligibility check | Knew STEMI, wrong on sequence | Management-sequence block | 19 Jul | correct, unseen |
| e.g. Q077, 12 Jul | Clin pharmacology | 3 | Old first-line to current first-line | Learned superseded guidance | Re-anchor to dated source | 26 Jul | correct, unseen |
The log's power is in the tally, not the individual rows. Twenty logged items usually reveal a lopsided distribution — say eight execution slips, six reasoning errors, and only two true knowledge gaps — and that distribution is the actionable finding. It tells the candidate who has been re-reading textbooks that their problem is not content at all, and it tells the candidate who feels "careless" that their misclicks are in fact premature-closure reasoning errors wearing a disguise. Crucially, the log carries a re-test column, because an error is not resolved when you understand the explanation; it is resolved when you get a differently-framed, unseen item on the same point correct under time. That re-test discipline is where this pillar hands off to the two-Q-bank rule and to the transfer-practice workflow, its sibling pillars.
Worked examples across three exam types
Written MCQ (USMLE Step 2 CK, United States). A long vignette describes an acute presentation and ends "which is the most appropriate next step in management?" You recognise the condition instantly and select the definitive treatment; the credited answer is the immediate stabilising step. This is not a knowledge gap — you knew the diagnosis — it is error type 8, a management-sequence fault, sometimes compounded by type 4 if you skimmed the lead-in. Logged correctly as type 8, the intervention is a block of next-step items across conditions you already understand, which takes an evening. Logged wrongly as a knowledge gap, you would re-read a pathophysiology chapter you did not need, feel studious, and miss the identical trap next week.
Structured response (RACGP AKT and KFP, Australia). The RACGP Fellowship written stage pairs a 150-item single-best-answer Applied Knowledge Test with a Key Feature Problem paper of 70 multiple-selection cases, each case independent. A partially credited KFP answer set almost never reflects a pure knowledge gap; it is usually type 6, an integration failure, or type 7, a differential that was too narrow. The distractor-capture and misclick error types matter far less here because the format does not present fixed distractors in the same way — which is exactly why blindly importing MCQ review habits to a structured-response paper misfires. The intervention is to force a complete, prioritised list of key features before committing, then compare it against the official model answer's weighting.
Clinical simulation (USMLE Step 3 CCS, United States, and MRCGP SCA, United Kingdom). In a Primum computer-based case simulation you order the correct investigations but advance the clock past the point where the intervention mattered, or you omit a monitoring order — that is type 8 or type 10, not a content fault. In an SCA consultation, one of twelve twelve-minute simulated encounters judged across data gathering, clinical management and relating to others, a missed safety-net is a type 8 error inside the management domain. The correct intervention for both is to rehearse the sequence in a fresh, unseen case, not to re-read the condition — which is why simulation errors flow naturally into the transfer-practice pillar rather than into more reading.
Where the taxonomy should not be applied
The taxonomy has limits worth stating plainly. It is a routing tool, not a pass predictor: a clean, well-distributed log means you are fixing the right things, not that you will pass, because pacing on the day, standard-setting and test conditions sit outside it. It should not become a cataloguing hobby — if you are spending longer classifying errors than practising, the log has stopped earning its place, and two or three well-chosen types per week beats a forensic twelve-way audit of every item. Some misses are genuine noise, and forcing every one into a category manufactures signal that is not there; a single fluke does not deserve an intervention. And error type 12 exists precisely so you do not over-learn from a broken question — when an item is flawed, ambiguous, or outside the published blueprint, the correct response is to verify against the official source and discard it, not to memorise its idiosyncratic key. Finally, the taxonomy classifies why you missed a question; it does not, on its own, tell you whether a whole domain is under-covered. For that you pair it with the blueprint coverage matrix, which works at the domain level where this works at the item level.
The evidence behind the taxonomy
The taxonomy is not a novel theory; it is a consolidation of well-established bodies of work applied to your own revision. The knowledge and reasoning families draw on the diagnostic-error literature, which distinguishes faults of knowledge from faults of reasoning — anchoring, premature closure, and the failure to integrate available data are documented cognitive dispositions, not character flaws, and each has a described countermeasure. The execution family reflects the plain observation that high-stakes exams are performances under time, so pacing and checking are trainable skills separate from knowing the medicine. And the whole method rests on the testing-effect evidence: retrieval followed by targeted feedback produces more durable learning than re-reading, which is why the log's re-test column, not its explanation, is where the learning is banked. Where a claim inside any explanation would change your behaviour — a dose, a threshold, a first-line agent — check it against a current, dated primary source rather than trusting the bank's prose, because a confidently written wrong explanation is itself a way to acquire an error type 3.
An iatroX workflow that demonstrates the taxonomy
iatroX is a question-bank and citation-first clinical-knowledge platform, and its role in this framework is deliberately narrow and honest: it is the verification and measurement layer, not the source of the log. Learn and drill in whatever bank you prefer, and keep the error log as you review. When an item is classified as a knowledge, decay or outdated-knowledge error (types 1–3), route the underlying clinical question through Ask iatroX for a current, cited answer, so you re-encode the correct version rather than the plausible one the explanation may have given you. Then, to close the loop, draw a fresh, unseen block from iatroX to serve as the re-test for that error — a differently-framed item your learning bank never showed you, so a correct answer means transfer rather than recognition. iatroX supplies the citation for the fix and the cold item for the proof; your main bank supplies the teaching. We make no proprietary-algorithm claims for this: the mechanism is simply that unseen questions measure honestly and cited answers verify honestly.
Implement it this week
You do not need a new subscription to start. This week, take the next twenty questions you get wrong and log each one against the twelve types the moment you review it, filling every column including a one-line root cause. At the end of the twenty, tally by type and by family. Pick your top two error types and do only their mapped interventions for the following week — if your top two are 8 and 10, that means a management-sequence block and timed pacing practice, and explicitly not another pass through your notes. Set a re-test date for each logged item using an unseen source, and record the result. One honest week of this will tell you more about where your marks are actually leaking than a month of undifferentiated review, because it replaces the comforting assumption that you simply need to know more with the specific, often surprising, truth about how you lose points.
A fully worked error log, start to finish
Take a candidate six weeks out from a broad written membership exam who feels their problem is knowledge and has been re-reading steadily. They log twenty misses. The tally comes back: three type 1 knowledge gaps, one type 2 decay, one type 3 outdated, two type 4 lead-in misreads, one type 5, two type 6 integration failures, three type 7 premature-closure errors, four type 8 wrong-next-step errors, one type 9, two type 10 pacing faults, and no type 11 or 12. Read as a distribution, this is decisive: only five of twenty are content faults, while nine are reasoning faults (7, 8, 9) and three are comprehension or pacing. The candidate's self-diagnosis was wrong. More reading would have addressed a quarter of the problem and left the largest cluster — right-diagnosis, wrong-management errors — completely untouched. The intervention list writes itself: a fortnight weighted heavily toward next-step and differential-discrimination practice, a stem-parsing habit for the misreads, and one timed block a week for pacing, with the five genuine knowledge points encoded and spaced rather than re-read. Every logged item gets an unseen re-test, and the re-test column, not the candidate's mood, decides what is resolved. In half an hour of logging, "I need to know more" became "I need to reason to management faster and read the lead-in properly", which is a plan.
The three cataloguing mistakes candidates make
First, defaulting everything to type 1. When in doubt, candidates label a miss a knowledge gap, because that fault feels legitimate and its fix — read more — is familiar; the result is that reasoning and technique faults hide inside an inflated content count and never get their proper intervention. Force yourself to justify a type 1 label by asking whether you truly never knew it, or whether you knew it and mishandled it. Second, logging the explanation instead of the error. Writing down what the correct answer was teaches you nothing new; writing down why your process produced the wrong one is the entire point, so the root-cause column must describe your fault, not the item's answer. Third, skipping the re-test. An error you understand is not an error you have fixed, and the only proof of a fix is a correct answer on a differently-framed, unseen item under time — which is why the re-test column exists and why a log without it quietly reproduces the same false confidence a rising Q-bank percentage does. Avoid those three and the log runs itself; fall into them and it becomes another comfort object.
Frequently asked questions
Does this framework work for every medical exam? Yes, because it classifies faults in the candidate rather than features of the exam. The twelve causes — knowledge, comprehension, reasoning and execution faults — appear whether the format is best-of-five MCQ, structured short-answer, key-feature problems or a clinical simulation, so the taxonomy transfers across UK, US, Canadian, Australian and European exams. What changes per exam is the mix: fixed-distractor MCQs generate more type 9 distractor-capture errors, while structured-response papers generate more type 6 integration failures, which is exactly why each exam-specific child article names the error types its format tends to produce and links back here rather than reproducing the taxonomy.
How often should the framework be updated? The taxonomy itself is stable and does not need updating; what you update is the log, continuously, and your intervention priorities, weekly. Re-tally your errors by type at least once a week during active preparation so your effort tracks your current fault distribution rather than the one you had a month ago. The only reason to revisit the taxonomy's content is a change in your exam's format — for example a paper moving from written response toward multiple choice, which shifts which error families dominate — in which case you re-weight your attention, not the categories.
Which metrics are valid across different Q-banks? The portable, valid metric is your error distribution measured on unseen, first-attempt items: the share of your misses falling in each type and family, tallied from questions you had not seen before. A bank's headline accuracy percentage and its per-topic scores are not comparable across banks, because they blend first attempts with recognition of repeated items and reflect each bank's authoring choices. Count your error types on cold questions and you have a metric that means the same thing in any bank; read a bank's own percentage and you have a number that means something different in every one, which is the point of the score-interpretation article this pillar sits beside.
How should AI-generated feedback be verified? Treat an AI explanation of why you missed a question as a hypothesis about your error type, not a verdict. If it names a knowledge gap, confirm that you genuinely did not know the point rather than mishandled it; if it asserts a clinical fact that would change your behaviour, check that fact against a current, dated primary source such as NICE, CKS, SIGN or the SmPC/eMC before you re-encode it, because a fluent explanation can deliver superseded guidance and manufacture a type 3 error. For any AI tutor you rely on to diagnose your mistakes, run the four-axis audit in the audit-an-AI-tutor pillar and, for automatically scored answers, the feedback-calibration protocol.
How does iatroX implement the framework? iatroX implements it as the verification-and-measurement layer, not the log itself. When your log flags a knowledge, decay or outdated-knowledge error, Ask iatroX returns a current, cited answer so you re-encode the correct version; when you need to prove a fix, iatroX supplies a fresh, unseen, timed block as the re-test, so a correct answer reflects transfer rather than recognition. iatroX operates a competing question bank, so we confine its role in this framework to the jobs your primary bank does not claim — citation-first verification and uncontaminated measurement — and make no proprietary-algorithm claims for how it selects those questions.
Why an error taxonomy outlives any single exam
The deeper reason to build this habit is that the skill it trains — knowing precisely why you were wrong and matching the fix to the fault — is the same skill clinical practice will demand of you for the rest of your career, at higher stakes and without an answer key. A doctor who can distinguish "I did not know" from "I knew but anchored" from "I was rushing" is doing, in revision, the metacognitive work that underpins safe practice and genuine learning from error. That is why the taxonomy is deliberately exam-agnostic: the specific banks and formats will change, candidates will sit exams that do not yet exist, and the twelve reasons a competent person gets a hard question wrong will not. A candidate who internalises the taxonomy now carries it into every future exam and every future clinical review, and the error log they keep for a membership paper is a rehearsal for the reflective practice they will keep for a licence. Reduce the whole pillar to one instruction and it is this: never fix a wrong answer until you know what kind of wrong it was — because the wrong fix, applied diligently, is how confident candidates run out of weeks. The exam-specific guides in this series each pre-populate this taxonomy with the error types their format tends to produce and the official sources to verify against, so build your first log here and every future exam becomes a matter of swapping in the new format's fault profile, which is exactly the portability a good framework should give you. Compare banks on how well they support this workflow at the comparison hub, and anchor your domain-level view with the coverage matrix so that item-level errors and domain-level gaps are read together.
Editorial notes and references
Written by Dr Kolawole Tytler, NHS GP and founder of iatroX. Last checked 19 July 2026. This is a cross-exam framework article; exam-specific guides in this series apply the taxonomy with named formats and official sources, and any vendor figures they cite are dated and labelled vendor-reported. Disclosure: iatroX operates question banks and a citation-first clinical AI and is not a consultation or case simulator; its role in this framework is confined to citation-first verification and unseen measurement — jobs a learning bank does not claim — and we make no proprietary-algorithm claims. Corrections via the feedback route on iatrox.com. References: diagnostic-error and dual-process reasoning literature (anchoring, premature closure, integration failure); the testing effect and feedback in durable learning; official exam-body formats for MRCP(UK) Part 1, USMLE Step 2 CK and Step 3 CCS, MCCQE Part I, RACGP AKT/KFP and MRCGP SCA as cited in the relevant child articles; related reading: why your Q-bank percentage is not your exam score and the sibling pillars on preserving unseen questions and transfer practice.
Copy the error log, run it on your current bank, then test the gap in iatroX →