pymetrics Cards Game: Complete Practice Guide | Game Assessment Prep

What is the pymetrics Cards game?

Cards is an implicit-learning game based on the Iowa Gambling Task. Four face-down decks sit in fixed positions. You begin with a running balance, select one position, and reveal a gain plus, on some draws, a simultaneous penalty. The balance updates and you choose again. Nothing labels a deck as safe, risky, good, or bad while the game is running.

The challenge is not arithmetic. It is learning from a noisy sequence. A deck can produce an appealing gain now and still be costly over repeated draws. Another can feel less exciting while producing a better long-run balance. A single penalty does not identify a schedule, so you must integrate evidence across choices rather than reacting to the latest card.

Our reconstruction starts at $2,000 and lasts 80 draws. The four underlying schedules are shuffled across the four visible positions from a seeded session randomizer. You see only the selected card's outcome and the overall balance during play. Per-deck totals, quality labels, coaching hints, and “best deck” advice stay off-screen because any of them would damage the learning task.

What does Cards measure?

Cards can produce several related observations. Exploration is how broadly you sample before committing. Learning is whether choices shift toward the net-positive schedules as evidence accumulates. Persistence describes whether an attractive high-gain schedule continues to hold your attention after its larger penalties become apparent. These behaviors can also reflect tolerance for uncertain gains and losses.

The most useful result is therefore not merely the final balance. Money depends partly on which penalty happened to arrive within your chosen sequence. Our practice percentile uses the percentage of net-positive-deck draws in the final 40 choices. That later window asks whether learning affected the end of the session, after you had an opportunity to sample outcomes.

This is still only a transparent practice metric. Pymetrics does not publish its production formula, feature weights, or employer thresholds. We cannot tell you that one deck percentage guarantees a role match, and an employer may interpret decision patterns alongside the rest of the battery.

The four schedule types

The canonical structure comes from Bechara, Damasio, Damasio, and Anderson's 1994 Iowa Gambling Task. Two schedules are disadvantageous over repeated blocks: they offer $100-scale gains but impose penalties large enough to create a net loss. Two are advantageous: their $50-scale gains are smaller, but their penalties leave a positive long-run return.

Loss frequency is the load-bearing distinction inside each pair. One disadvantageous schedule uses frequent, smaller penalties; the other uses an infrequent, very large penalty. The advantageous pair has the same split: frequent small penalties versus a rarer larger one. A simulation with only “high payout bad” and “low payout good” misses an important part of the paradigm because frequency changes how a deck feels even when its long-run value matches its partner.

Our default schedules repeat static ten-card blocks modeled on the canonical totals. Position-to-schedule assignment changes each seeded session, so memorizing “the second deck is safe” does not work. A configurable drift flag exists for future evidence, but it is off because non-stationary pymetrics payouts are plausible rather than verified.

What is known—and what remains uncertain?

Cards is the least settled game in the core battery at the parameter level. Eighty draws is the recommended default because it is the only specific pymetrics count reported alongside the $2,000 balance in an internally consistent account. Claims of 39 and 135 are refuted cross-attributions: 39 belongs to Balloons and 135 belongs to Arrows. The academic Iowa task often uses 100 trials, but that does not prove pymetrics does.

The starting $2,000 has low-to-medium support and matches the classic task. Exact gains and penalties remain unresolved across conflicting preparation sources. We use canonical Bechara-style blocks because they preserve net value and loss-frequency structure, not because we have recovered pymetrics' private card table.

Candidate-facing deck labels are also unconfirmed. Academic papers call them A, B, C, and D, but that notation may never appear in the product. Our interface is position-only. All these values live in one configuration object so credible production evidence can correct them without changing the learning engine.

Six practical strategies

1. Explore deliberately at the beginning

Give every position enough attention to generate evidence. Drawing repeatedly from the first exciting deck can leave the alternatives unknown. Exploration has a cost, but no learning is possible when two or three positions remain untouched.

2. Compare net outcomes, not headline gains

Mentally pair each gain with any penalty on the same card. A +$100 headline with a −$250 penalty is a −$150 outcome. Large gains are intentionally salient; the balance tells you what actually happened.

3. Track positions with a simple impression

Use a lightweight mental label such as “steady,” “volatile,” or “still unknown.” Trying to memorize every card can overload working memory. What matters is the emerging pattern of net effect and penalty frequency.

4. Require repeated evidence

One large loss can come from a generally positive but infrequent-loss schedule, and several clean gains can precede a very large disadvantageous penalty. Update your belief after each draw, but do not reverse it completely from one event.

5. Notice when exploration should become use

By the middle of the task, ask which positions have produced the most convincing repeated evidence. Continuing to sample every deck equally can prevent your learning from influencing the final window. Commitment should follow evidence, not an arbitrary draw number.

6. Do not chase losses

The next card is not obligated to recover a penalty. Choosing the same position to “win it back” replaces outcome learning with a sunk-cost response. Treat the current balance as history and make the next choice from the schedule evidence you have.

How to read your practice result

Final balance reports the literal outcome of the 80 draws. Good-deck draws shows the share of all choices from the two net-positive schedules, and Decks explored counts how many positions you sampled. The percentile score uses only the last 40 good-deck percentage so it emphasizes learned behavior rather than early exploration.

The insight looks for a point after which you found the net-positive pair and stayed, or describes the concentration of positive choices late in the task. It is computed from your actual log. The log also stores schedule assignment, reward, penalty, net value, reaction time, and balance after every draw under the session seed.

Cards FAQ

Why are the decks not called A, B, C, and D?

Those letters are an academic convention, not a confirmed pymetrics screen detail. Position-only cards avoid implying evidence we do not have and prevent a learned letter rule from carrying between simulations.

Does a penalty mean I chose a bad deck?

No. Both net-positive and net-negative schedules include penalties. Their size, frequency, and long-run total distinguish them across repeated draws.

Why do I not see deck statistics during play?

The task measures whether you learn from experienced outcomes. A per-deck dashboard would perform that integration for you and change the construct. Detailed interpretation belongs after completion.

Is final balance the score?

It is a useful outcome but includes schedule luck. We rank the last-40 net-positive choice percentage because it more directly reflects whether later decisions used the available evidence.

Is there a pass mark?

No public universal pass mark exists. Use the result to improve deliberate exploration and evidence integration, not to reverse-engineer an employer's undisclosed role model.