Imagine one capital letter, then on each round pick the pattern that looks most like it. Afterwards the program combines your choices and ranks all 26 letters by how likely each one is to be your letter.
Don't click it or tell anyone. Each letter is drawn on the same 25 × 25 grid that the patterns use. When you're ready, press Start trials.
Each stimulus is a sum of these parts. About a third of them appear in any one stimulus, each at a random strength from 0.4 to 1. Every letter can be assembled from some of these parts.
For each run a simulated observer secretly picks a random letter. On each round it computes each pattern's correlation with that letter, adds Gaussian decision noise σ, and picks the highest. The chart shows how often the top-ranked guess is correct after each trial. It uses the trials, choices per round and σ from the settings above.
Reverse correlation. On each round the chosen image adds to each pixel's weight, and the average of the rejected images subtracts from it. Pixels that belong to your letter drift toward white. With 625 independent pixels it needs many rounds.
Gradual approximation. A hill-climb. Each round shows distortions of the last chosen image, made in ± pairs so that the options differ only in the distortion. The step size falls from σ = 0.35 to 0.08, and the final chosen image is what gets scored.
Letter mixtures. The 26 letters are dealt at random into one group per choice, and each group is shown as its average. Your letter is always in exactly one group, so every choice carries information about it. The estimate is the same chosen-minus-rest average as in method 1.
Adaptive splitting. The program keeps a probability for each letter. Each round it deals the letters into equal-size groups, like method 3, but arranges them so each group holds a similar share of the probability, then updates by Bayes' rule. Keeping the group sizes equal matters: a lone likely letter against a blurry average of 25 others makes a lookalike letter win every time. Within a group, letters are weighted by the square root of their probability, so the leading candidates become easier to see and compare. The model assumes you pick the group with your letter with probability 1 − ε, and each other group with probability ε / (choices − 1). The error rate ε isn't fixed: it's inferred along with the letter from how consistent your choices are. A perfect observer with 2 choices is found in about 5 rounds (log₂ 26 ≈ 4.7).
Letter-part noise. Reverse correlation in a 24-dimensional space of strokes instead of 625 pixels. It converges faster than pixel noise but can still produce shapes that aren't one of the 26 letters. The estimate is scored by correlation like methods 1–3.