The LearnClash Misconception Index is a curated cut from a read-only capture on 2026-09-09, between 22:30:26 and 22:30:44 UTC. The export read 968 public topic records, then stopped at 40,000 questions: 21 of 33 topic-query batches completed. This cap produces a partial, nonrandom sample. Of 495 automatically eligible rows, we reviewed all 16 where one wrong option led the correct answer and retained 8. The linked CSV supplies every table on this page. Kaggle retains the separate July edition.
What the counts include
The current counter contract records eligible first SRS encounters through the duel pipeline, excluding timeouts, information-assisted joker answers (50:50, Peek and Crowd) and web-play replays. Double Down remains eligible. Under the current write contract, human-side answers in AI-rival duels can contribute; standalone Solo does not write this question-level counter. Answer keys are content-version scoped, while first-encounter eligibility follows the separate meaning-change history. A content-version change does not necessarily reset that history. This source review describes today's contract, not verified enforcement for every event accumulated in these counters.
The export date is a capture date for cumulative counts. Per-answer dates, confidence, language-specific groups, option impressions and distinct people across rows are not available. The data cannot isolate answers from that day or support a July-to-September trend. These are recorded LearnClash app answers, not a representative population sample; distinct people cannot be counted.
What the diagnostic labels mean
Each exposure shows the correct answer plus 3 distractors sampled uniformly from a pool of M wrong options (M is 5 on 8 rows here, 7 on 0). Under pure guessing, one specific distractor expects a 0.75/M share of all answers: 15% at M=5. The exploratory p-value is the one-sided binomial tail, the chance of seeing at least the observed count on that option if players were guessing.
- Strict: p below 0.05/M, a Bonferroni correction for the M wrong options that could each have been the top one. This correction covers options within one question, not selection across many questions or repeated players.
- Nominal: p below 0.05 but not below the corrected bar.
- Descriptive: everything else. The observed share and n remain descriptive; the label does not establish absence of a pattern.
How the questions were selected
- Floor: 30 first-encounter answers per row. No row on this page or in the file sits under it; the median row has 42, the largest 71.
- The current cut selects only source-reviewed examples from the 16 wrong-option-leading rows. Original export ranks are retained, so the displayed rank numbers have gaps. It excludes ambiguous or unsupported wording. It is not the complete n of 30+ set, a representative topic mix or an estimate of misconception prevalence.
- "Beats correct" describes the counts. Small samples and selecting the most extreme rows limit interpretation; neither a fixed margin rule nor the signal label establishes a population-level effect.
Earlier and related data
Prepared by David Moosmann, founder of LearnClash. No individual user records are included.
Questions, corrections, or a cut this page doesn't show: admin@learnclash.com. Press facts and logos are on the press page.
