Four-deck card game: method

What it measures

Decision making under uncertainty: learning from wins and losses to prefer the options with better long-run outcomes.

Procedure

Four decks are shown in random screen positions (shuffled per run). The player starts with 2,000 points and makes 100 picks, untimed. Each card shows its win and any loss at once. Deck kinds per 10 cards: A wins 100 per card and has five losses (150, 200, 250, 300, 350; net -250); B wins 100 per card and has one loss of 1,250 (net -250); C wins 50 per card and has five losses (25, 50, 50, 50, 75; net +250); D wins 50 per card and has one loss of 250 (net +250). Within each block of 10 cards of a deck, the loss positions are shuffled per block and per run. Each deck has 40 cards; an empty deck can no longer be chosen. The server regenerates every card from the seed and replays the picks.

Parameters

Scoring

Primary metric: net score = (picks from C + D) - (picks from A + B) over all 100 picks, range -100 to +100; higher means a stronger preference for the good decks. Also reported: net score per block of 20 picks (learning curve), final points, points change, picks per deck kind, and which screen deck was which. Once at least 30 eligible earlier runs exist, the result shows the percentage of players with a lower net score (ties count half). Reference values from open data (330 healthy adults, 4 studies with the same structure, first 100 picks; own calculation): mean net 2.5, median -2, 51% below 0; block means -2.5, -1.0, 0.4, 2.4, 3.1.

Valid-run rules

Exactly 100 picks are required, and a pick from an empty deck makes the run invalid. Every scored run is stored with its input device class (touch, mouse, keyboard), page language, scorer version and runner-config hash. Quality flags (for example many runs from one connection within an hour, a missing visitor id, an automated browser, impossibly fast answers) and earlier attempts by the same visitor, account or connection are recorded with the row. Such runs are kept, but they are left out of any comparison with other players. Flags never change the player's own result.

Differences from the classic task

The original task uses a fixed card sequence, play money or real rewards, and is sold as a licensed test. This version keeps the 1994 wins and losses per 10 cards but uses its own shuffled loss positions per block (the structure called 'payoff scheme 2' in the open data), random deck positions, points without money, and runs online.

Limitations

Healthy adults vary widely; about half score below zero, so a low score is common and not informative about health. Results depend on how much a player explores and whether they know the task. Reference values come from lab studies with other rewards and are only roughly comparable. Unsupervised setting. No clinical or diagnostic use.

Languages and review status

Numbers and card pictures only; instructions in all site languages.