What happens to a wrong idea.

Two things share this word.

Worth separating before anything else, because you will meet both in the app.

Where they come from.

There is exactly one source: the AI grader, when it reads a free-form answer. Its response format has a slot for them next to the score and the feedback — a list of short phrases, each naming a gap or an error. Things like "treats average velocity as instantaneous". That list is saved with your attempt.

Which means the coverage is narrower than most people assume, and we would rather state it than let you infer it:

• Five of the fourteen exercise types can produce one

  Short answer, writing, fill in the blank, code bug spot, solve equation. These are the types where a model reads what you wrote.

• The other nine can never produce one

  Multiple choice, true or false, flashcard, numeric answer, matching, ordering, categorize, word order, code trace. They are graded by exact comparison, and a comparison cannot say why you picked the wrong option — only that you did.

• And two of the five often skip the grader

  Fill in the blank stops early when every blank matched, and solve equation stops early whenever your answer is a number it can compare. A wrong number is scored without a model ever reading it, so it records nothing.

• No AI provider, no misconceptions

  Running without a model configured, the graded types fall back to exact comparison where they can. Everything on this page then stays permanently empty.

Stored, it is a list of short phrases on the attempt itself — not on the concept, not on your profile. Nothing is aggregated at write time; every question below is answered by re-reading your attempt history.

The lifecycle.

There is no stored "open" or "resolved" column anywhere. The state is derived on every read by comparing two timestamps: when a wrong idea was last recorded on a concept, and when you last answered that concept correctly. That is the whole model, and it is deliberately that small — a flag that can be written can go stale, and a flag that is recomputed cannot.

The chance discount, and why a correct guess does not count.

Ignia already discounts answers you could have collected by luck. A perfect true-or-false is worth 0.75 towards durable memory rather than 1.0, because half of it is a coin flip; a four-option multiple choice is worth 0.875. The threshold for "this counts as knowing it" is 0.80.

Until recently the misconception flag ignored that discount and read your raw score. The result was a contradiction the app made about a single keystroke: a correct true-or-false was simultaneously not durable evidence of learning (0.75, below the bar) and proof that you had corrected your belief (1.0, above it). Since the misconception check is itself a true-or-false, that was not an edge case — it was the normal case. Roughly half the learners who still held the wrong idea had it marked resolved by guessing.

The closing test now reads the discounted score. Nothing else moved: no threshold was raised, no new constant invented. One consequence follows and is worth stating outright — the misconception check can no longer close its own flag. Clearing it takes a correct answer on an ordinary exercise for that concept, of a type where being right is not mostly luck.

The check itself.

While a concept carries an open misconception, the session planner may append one extra exercise to contest it. It is the only place in the app where true-or-false is used at all.

• A forced true-or-false, whose false statement encodes your error. The AI gets no choice of type here. True-or-false is otherwise banned outright — it is absent from every exercise-type group, so it can never be picked to fill a normal slot.
• Written from your recorded wording. The phrases the judge stored are quoted into the authoring instruction, with a demand that the false statement be phrased as you would phrase it rather than as a generic distractor. Where no usable wording exists, the instruction falls back to a generic one rather than inventing something.
• It must explain itself. An explanation is required on this exercise and re-checked when the AI's output comes back: a check that names a wrong idea and offers nothing to replace it is worse than silence, and is sent back to be rewritten. If the model still will not write one after the retries, the question is dropped from the session rather than served without it. The obvious route is the other one — ship it anyway rather than lose a question — and it is the wrong one here: this exercise states your own wrong idea back to you, in something close to your own words, and reading that with no correction attached is a documented way to make it stick. Dropping it costs nothing you would miss, because the check is appended on top of your session rather than replacing practice: what is left is still a complete session.
• At most two per session, chosen by struggle, deterministically. They are appended on top of your session budget rather than displacing real practice, which is why they are capped.
• Exempt from the cognitive floor. True-or-false normally cannot appear on an apply-level concept or above, because it measures below what the concept claims. The check is the one exception, and it is exempt because it is a diagnostic rather than a measurement.
• Never in an exam, a lesson session, or a recalibration. Exams are mastery-blind by design. Recalibration prompts forbid true-or-false explicitly, so the two rules cannot collide.
• Never dropped for a repeat. When the planner swaps fresh exercises for ones you have already seen, a misconception check is protected from that swap.

Who reads them.

Three parts of the app read the stored phrases, and they do not read them the same way. The differences are real, not incidental.

A fourth surface is not a reader of history at all: the exercise card shows the phrases from the answer you just submitted, immediately, next to the feedback.

The exam wrap-up follows one further rule, because it is the only place a wrong idea is named long after you wrote it: a wrong idea is shown only when the screen can also show what replaces it — the correct idea, or failing that a link back to the section that teaches the concept. With neither, the concept keeps its score and its "drill this" button, and the diagnosis is withheld rather than left standing on its own. The correct idea is written when the exam is generated, one sentence per question, and the wrap-up shows the one belonging to the question that went worst.

Note what the second row means in practice. Recorded misconceptions feed back into the course, not only into your practice: if a section keeps producing the same wrong idea, that is evidence about the section. Revising it re-injects those phrases into the rewrite.

Honest limits.

What this mechanism cannot currently do. Each of these is real today, and none of them is hidden behind a softer word elsewhere in these docs.

• A failed check records nothing. The misconception check is a true-or-false, which is graded by comparison, so getting it wrong writes no new phrase. The flag simply stays open at its original timestamp. That means how often you have relapsed and how badly are not representable at all — the only stored fact is "most recently recorded on this date". Answering severity would need data the app does not collect.
• Whole subjects can never register one. Because only free-form answers reach a judge, a course taught mostly through numeric answers, matching and multiple choice will produce no misconceptions at all, however wrong the learner is. This is a property of the exercise mix, not of the learner, and nothing currently tells you it is happening.
• The course-revision reader picks its 30 attempts by row identifier, not by date. Identifiers are random, so "the 30 most recent" is in fact an arbitrary 30 on a concept with a long history. It does not affect the flag or the session check, which scan the full history — only which phrases a revision suggestion happens to quote. Queued.
• You cannot see the list, contest it, or clear it by hand. The phrases are shown to you as they are recorded and in an exam wrap-up, but there is no screen listing what the app currently believes you believe wrongly, and no way to say "no, I never thought that". The only way to close one is to answer the concept correctly.
• The wording is the judge's, not yours. A phrase is the grader's description of your error, written for another model to read. It is deliberately never shown to you as a diagnosis of what you believe — only as feedback on the answer that produced it.
• Newly connected, not yet field-validated. The check fired for the first time only recently — for most of the app's life the flag was hardcoded off — and aiming it at your recorded wording is newer still. The mechanism is verified by tests; whether the questions it produces are good is a judgement no test makes.