What happens after you answer a practice question

The grade never comes from the AI

When you submit, the server compares your choice with the answer stored on the question. That comparison is ordinary code. No model is called, nothing is generated, nothing is charged against your usage. The result comes back at once.

This is deliberate. A model asked to mark your work can mark it wrongly, and a study tool that is sometimes wrong about whether you were right is worse than no tool. Correctness is a fact already on the record, so it is read, not inferred.

A banner says which of three things happened: “Correct!”, “Incorrect — the correct answer is C.”, or, if the clock beat you, “Time’s up — the correct answer is C.”

The AI never decides whether you were right. It only explains the choice you made.

Every option comes back, annotated

The options do not disappear once you have answered. They are re-rendered, marked up with everything the bank knows about them:

• The correct option is framed green — whether or not it was yours.
• Your own pick is labelled “Your pick”, and framed red if it was wrong. Two frames, one glance: the answer, and the distance between it and where your reasoning landed.
• A bar on each option shows the percentage of the source bank’s test-takers who chose it. A wrong answer that most people also chose is a different problem from one almost nobody picked; the first is a hard discriminator, the second is usually a misread stem.
• A “Why not” toggle on each wrong option, where the bank supplied text for that specific choice. It stays collapsed, so you open the one distractor that tempted you instead of reading four rebuttals you did not need.

If the question belongs to a linked set, a This set strip sits above it — “Item 1 of 2”, and how each item in the set went — because a shared stem is one clinical picture, and the items are meant to be reviewed together.

The explanation, and the images that were held back

Under the options comes EXPLANATION: the bank’s full written explanation, along with any images placed inside it. Those images are deliberately not shown before you submit. A slide labelled with the diagnosis, or a graph with the inflection point ringed, answers the question on your behalf, so they are held back until the answer is settled.

Then EDUCATIONAL OBJECTIVE, in a box of its own: the single point the item was written to test. Read it even when you were correct. It tells you what to carry forward, which is rarely the fact you happened to remember.

Then the AI reviews the option you picked

AI ANALYSIS OF YOUR ANSWER streams in live, about a hundred and fifty words. It is not a second copy of the explanation. It is about your choice specifically: why that option was tempting, what it would have been the correct answer to, and the discriminator in the stem you read past.

Claims in it carry numbered markers. Tap one and the real textbook page opens, as an image of the page itself, so you can read the paragraph around the sentence. The citations are derived on the server from the passages that were retrieved: the model cannot invent one, and a marker that does not map to a retrieved passage is dropped before it reaches you. That is the same discipline the assistant uses everywhere else, described in how citations work.

The block carries a report flag, because an analysis that reads badly is a thing we want to see. And the analysis is stored against that attempt, so refreshing the page replays the one you already have rather than generating a new one — the review is the same review tomorrow, and it costs nothing to reread.

Asking the follow-up

Below the analysis are one to three suggested follow-up questions as chips, and a box that says Ask about this question… if none of them is what you wanted to know. Replies come from the same assistant that answers in chat, with the same citations, and they stay threaded under that attempt. Weeks later the thread still reads as a record of what confused you, attached to the question that confused you.

The order that works is dull: commit to an answer, submit, read the objective, then ask. Asking before you commit converts active recall into reading, which feels productive and teaches very little. More on that in trusting AI in medical study.

What this does not do

Timed sessions get none of it. No AI analysis, no follow-up chat. You still get the banner, the annotated options and the bank’s explanation — but the discussion is off, and it is the server that refuses, not the interface that hides a button. Under exam conditions the exercise is answering questions, not talking about them. The talking belongs to review afterwards.

It does not grade, it does not overrule the bank, and it does not claim to cover more than it does: the content here is Step 1. And like everything else on Med Guru, it is a study aid for exam preparation and revision, not medical advice — an explanation of why a board-style item has the answer it has, not guidance for a patient in front of you.