Every revision forum hosts the same anxious arithmetic: "I'm on 68% in bank A, people say you need 70%, but I was on 74% in bank B, am I ready?" The honest answer is that the question is unanswerable as posed, because raw percentages from different banks are different instruments measuring different things, and none of them converts into an examination score. This page explains why, mechanism by mechanism, and replaces the magic-number habit with a readiness dashboard that actually predicts.
Five mechanisms that break comparability
Question difficulty and selection: banks differ in average difficulty and in distribution, one bank's 70% is another's 60%, and neither is anchored to the live examination's difficulty, which itself varies by paper and is corrected by equating on the examination's side only. Candidate selection: the user base whose averages you compare against differs by bank, a bank favoured by resitters or by early-preparation students drags its percentile norms in opposite directions. Repetition and memory: second passes inflate scores through recognition rather than ability, and banks differ in how much repetition their default modes serve, so a percentage without a first-pass/unseen split is uninterpretable. Adaptive serving: banks that deliberately target your weak edge, ours explicitly among them, suppress raw percentages by design, because you are always practising where you fail, which is optimal for learning and terrible for vanity metrics; comparing an adaptive percentage against a random-serve percentage punishes exactly the better preparation. And mode mixing: timed mock questions and untimed learning-mode questions measure different performances, and blended percentages average them into noise.
Why no bank score converts into an exam score
The examination's own reporting makes the conversion impossible in principle for the best of reasons: modern examinations report equated scaled scores precisely so that standards stay constant while paper difficulty varies, MRCP Part 1's 450 being the worked example, an equated scaled score that is expressly not a raw percentage and not a fixed number of correct answers: /blog/mrcp-part-1-pass-mark-450-explained. A bank percentage sits on none of these scales: it is unequated, un-anchored, and generated from items the examination never audited. Score-prediction features deserve matching scepticism, credible only where the provider can show a genuine statistical bridge, unseen holdout items, calibration against real candidate outcomes, published error bars, and honest ones say "estimate" loudly; anything presenting a bank percentage as an examination score, with or without decimals, is decorating a guess.
The readiness dashboard that actually works
Replace the single number with five signals. Unseen-item performance: your percentage on genuinely first-exposure questions, the only percentage worth tracking weekly, because it is the closest proxy for examination day. Blueprint coverage: proportion of the official syllabus met in practice, since a strong average over half the map is a weak position. Timed-conditions performance: full blocks at true length and cadence, which for current USMLE candidates now means the 30-minute unit: /blog/usmle-step-2-ck-sixteen-blocks-2026. Stability across mocks: two or three full rehearsals landing in the same range beats one flattering outlier, and the trend line matters more than any point on it. And weak-domain floors: no blueprint domain below a threshold you set deliberately, because examinations sample everywhere and a single hollow domain is where passes leak. Five numbers, one dashboard, and the forum question dissolves: readiness is a profile, not a percentage.
How to use bank percentages anyway
They retain three honest uses. Within one bank, one mode, first-pass only, your trend over time is a real learning signal. Domain-level percentages localise weakness better than global ones, and are the correct input to adaptive practice. And percentile context, where a bank publishes cohort norms, tells you something about position, with the candidate-selection caveat above attached. What they never do is licence the sentence "people who score X% pass", which survives on forums because it is comforting, not because it is true; the number that deserves that sentence does not exist, and the dashboard above is what replaces it.
Frequently asked questions
My bank shows a predicted score; should I trust it?
Ask what it is calibrated against: unseen holdout performance bridged to real outcomes with published uncertainty deserves cautious use; anything else is a percentage in costume.
Is a low percentage in an adaptive bank a bad sign?
Usually the opposite: adaptive serving keeps you at the productive edge, and the informative number inside an adaptive system is progress on targeted weaknesses, not the raw average the targeting suppresses.
How many full mocks are enough?
Enough to demonstrate stability, typically two or three at true length and timing in the final month, spaced to allow correction between them; ten mocks that replace learning are worse than three that confirm it.