Daily Puzzle

8 hours left • solved by 14 people (14%)

How to Play

Diagnose the patient. Each guess reveals features that overlap between your guess and the answer across four tiles: Signs & Symptoms, Diagnostics, Medications, and Therapies. Your goal is to reveal enough information in all the medical linkages so that you can identify the unifying disorder.

Features are rankedby how specific they are the disorder, not necessarily the most common. Cough versus hemoptysis, for example.

After each guess you'll proximity percentage. The closer your guess is to the answer in feature space, the hotter you are. HOWEVER, a strong feature-disorder association doesn't always mean the disorder causes the feature. Consider partial hearing loss as a strong signal for Contact Dermatitis. Two words: hearing aids.

Stuck? GET HINT 🔎 reveals the disorder that sits closest to the answer in feature space (excluding any you've already tried). The hint doesn't take a guess, but it scores as a 0% and removes your unused guess bonus. One hint per puzzle.

Disordle is a hard game. Solving just one puzzle is an achievement.

About Disordle

Disordle is built and maintained by Ben Turley, a student at the Georgetown University School of Medicine. It was first released in 2023, inspired by the popular Wordle and not wanting to study for the Step 1 board exam. Version 2 was released in 2026 during the dedicated study period for Step 2 CK. The disorder pool currently holds 2000+ disorders. Daily puzzles draw a fresh one each day, and random puzzles are effectively infinite. See below for details on how puzzles are generated.

The dataset built using the NIH's MeSH descriptors. For each descriptor, PubMed is crawled to retrieve every paper indexed under it. The result is a sparse disease x feature co-occurrence matrix.

Association is scored as the naive-Bayes log likelihood ratio of a feature on a disease:

score(d, f) = log P(f | d) - log P(f | ¬d)
P(f | d) = (c(d, f) + α) / (c(d) + α·F)
P(f | ¬d) = (c(¬d, f) + α) / (c(¬d) + α·F)

A feature uniformly distributed across disorders has P(f | d) ≈ P(f | ¬d), so its log LR sits at zero. Laplace smoothing damps low-count pairs so a single co-occurring paper can't dominate. A feature is eligible to display when its log LR is positive (real above-chance association)Tile rows are rescaled 0-100 within the disorder for visual binning.

Proximity between two disorders is the cosine distance of their (clipped, non-negative) shrunk-PMI vectors over all clinical features:

shrunk-PMI(d, f) = log( (c(d, f) + α) / (E[c(d, f)] + α) ) − σ / √(c(d, f) + 1)
where E[c(d, f)] = c(d) · c(f) / N.

Tile ranking and disorder similarity are different problems. Log LR is sharper for "which finding discriminates this disorder?", but its cells stay positive across most low-count pairs once smoothed, so cosine over log-LR vectors comes out too dense to read. Shrunk PMI naturally drives noisy low-count cells negative; the resulting sparse vectors put unrelated disorders at distance ≈ 1, so a guess only earns a high proximity % when there's real overlap. Per-guess feedback is log-rescaled to the answer's nearest-vs-farthest distance range, so equal % steps reflect equal log-ratios in distance.

Per-puzzle scoring tracks the highest proximity you reach across your guesses (a solve is 100%). The end-of-game curve fits a normal distribution to every solver's best closeness % for that day, with your guess marked. The average % in the header is the mean of your best closeness across every daily you've played. Your Differential Level for a puzzle is the mean closeness across all 8 guesses, with unused guesses credited at 100%. Solving in fewer guesses pushes you up the ladder.

The dataset is closed-source, proprietary, unvalidated, and should not be reproduced.

Victory!

Medical Student Syndrome