elements.csv

element,display_name,benchmark,paper_id,paper_name,setting,module,summary,description,grading,grading_rule,consistency_types,types,domains,perspectives,difficulties,test_questions,test_rows,few_shot_questions,few_shot_rows,few_shot_source,aliases

add_sub,Addition and Subtraction,STEER,1.1.a,Addition and Subtraction,Foundations,Arithmetic,Adding or subtracting two numbers stated in a short word problem.,"A one- or two-sentence story (money, lives saved, plain numbers) gives two integers, decimals or large numbers and implies an addition or a subtraction. The model picks the exact result from four numeric options; distractors include the result of the wrong operation and off-by-one or off-by-ten values.",answer,One correct option per question part (`correct_index`); score is accuracy.,,decimal;integer;large,lives;money;numbers,,1;2;3,200,200,15,15,held_out,
mult_div,Multiplication and Division,STEER,1.1.b,Multiplication and Division,Foundations,Arithmetic,"Multiplying or dividing two numbers, including negative numbers, stated in a short word problem.","A short prompt gives two numbers (positive, negative or one of each) and asks for their product or quotient, sometimes inside a story about rent, prices or quantities. The correct option is the exact result rounded to two decimals; the question states the rounding; distractors (all within a factor of 3 of the answer) are sign errors, digit slips and the other operation where it is of similar size.",answer,One correct option per question part (`correct_index`); score is accuracy.,,both;negative;positive,money;numbers,,1;2;3,240,240,15,15,held_out,
compute_expectations,Compute Expectations,STEER,1.1.c,Compute Expectations,Foundations,Arithmetic,"Computing the expected value of a simple random experiment such as a coin flip, die roll or draw from a bag.","The question describes a random event with stated payoffs or probabilities (coins, dice, bags of items, weather) and asks for its expected value; some items have negative payoffs and some ask for an expected number of draws. Coins and dice are stated to be fair. The model picks the correctly computed expectation from four options with matching units, all within the range of possible outcomes.",answer,One correct option per question part (`correct_index`); score is accuracy.,,negative;positive,bag;coin_dice;dice;weather,,1;2,200,200,10,10,held_out,
micro_utility,Optimize over a Discrete Set,STEER,1.2.a,Optimize over a Discrete Set,Foundations,Optimization,Identifying the best of a short list of explicitly given alternatives: the bundle with the highest utility under a stated utility function.,"A person chooses between two or four bundles of two goods x and y (coffee and tea, textbooks and videos, ...) and their utility function is given: Cobb-Douglas (x^0.4 * y^0.7), linear (7x + 5y) or perfect complements (min(4x, 6y)). The options are the bundles, written ""x = 4, y = 7""; the key is the bundle with the highest utility, at least 5% above the next best, and the distractors are the other bundles. The best bundle never has at least as much of both goods as every other one, so the function has to be evaluated.",answer,One correct option per question part (`correct_index`); score is accuracy.,,given_preferences,education;everyday_items;medical,first_person;second_person;third_person,1,200,200,5,5,examples,
optimize_continuous,Optimize a Continuous Function,STEER,1.2.b,Optimize a Continuous Function,Foundations,Optimization,"Finding where a quadratic or cubic function, a box volume or a fenced area reaches its maximum or minimum.","Plain questions give a quadratic or a cubic with whole coefficients on a closed interval and ask for the x that maximises or minimises it; the answer is an interior critical point or an endpoint (for quadratics each about half the time). Word problems: a firm's cubic weekly profit over a capacity range, the open-top box cut from a square sheet (the best cut and the largest volume) and a rectangular pen along a river with a given length of fence (the best width and the largest area), told in five perspectives. Options are whole numbers; for the x that is best, every option is a point of the interval (the other critical point, the endpoints, other whole numbers in it) or a feasible cut or side; for the largest volume or area, the distractors are the values at other cuts and slips that overstate it.",answer,One correct option per question part (`correct_index`); score is accuracy.,,box;cubic;fence;quadratic,agriculture;business;manufacturing;math,anon;first_person;man;second_person;woman,1;2;3,200,200,20,20,held_out,
constrained_optimization,Constrained Optimization,STEER,1.2.c,Constrained Optimization,Foundations,Optimization,Maximising or minimising a linear or quadratic function of two variables subject to linear constraints.,"Linear problems (difficulty level 1) are word problems whose story fixes the whole problem. Maximum: a producer (a farm, a steel mill, a refinery, a juice plant, a fertilizer plant) can make any amounts of two products; the story gives the profit per unit of each, how much of two resources (hours of two machines, or field labor and fertilizer) each unit uses, and how much of each resource is available, and asks for the maximum profit. Minimum: a buyer or operator (a cattle-feed mix, two mines, a supplement drink, two printing presses, two power plants) can use any amounts of two inputs; the story gives the cost per unit, what each unit provides of two requirements (protein and energy, two ore grades, flyers and posters, ...), and the minimum required, and asks for the minimum cost. Both constraints shape the feasible region and the optimal vertex is unique. Quadratic problems (level 2) are stated without a story: maximise a concave or minimise a convex quadratic in x and y subject to one linear constraint that binds at the optimum, with x, y >= 0. All coefficients are whole numbers; the key is the optimal value to two decimals (in dollars for the word problems). The other options are the objective at other vertices of the feasible region, the optimum with one constraint left out, or, for quadratic problems, the unconstrained optimum and the objective at the ends or the middle of the constraint segment. The data directory's templates, whose stories were decoration around a bare formula, are no longer used.",answer,One correct option per question part (`correct_index`); score is accuracy.,,max;min,agriculture;base;chemicals;energy;food;manufacturing;mining;nutrition,anon;second_person,1;2,240,240,10,10,held_out,
prob_basic,Compute Probabilities of Outcomes,STEER,1.3.a,Compute Probabilities of Outcomes,Foundations,Probability,"Comparing or computing the probabilities of simple events, including recognising equal probabilities written in different formats.","Four kinds of question, all with exact keys. Same probability (Yes/No): two sources (weather services, sports analysts, fair stalls, clinical studies) state a probability in two different formats, among 0.35, 35%, 7/20 and ""7 in 20""; the values are equal in half of the questions and otherwise differ by 0.05 to 0.15 or by a factor of ten (0.4 against 4%). Most likely / least likely: two or three events, each a separate experiment with a fair die, two fair dice, fair coins, a standard 52-card deck, or two or three bags with stated marble counts; the probabilities are distinct and the answer is at least 0.02 ahead of the next. Compute: the probability of one event (13 with one die, 12 with two dice, 11 with two coin flips, 12 with three, 18 with one card, each with two wordings of the setup; or a marble from a bag with stated counts), with four reduced fractions as options: the key and three wrong values from common mistakes (the complement, favourable over unfavourable outcomes, one count off, double or half). The wrong values depend only on the event, so a repeated question always has the same options.",answer,One correct option per question part (`correct_index`); score is accuracy.,,compute;least_likely;most_likely;same_probability,cards;coins;dice;games;health;mixed;sports;urns;weather,,1;2,280,280,15,15,held_out+partial,
complement_rule,Complement Rule,STEER,1.3.b,Complement Rule,Foundations,Probability,Using the complement rule to find the probability of the remaining outcome when the others are given.,"A scenario gives the probability of one outcome, or of two disjoint outcomes, and asks for the probability of the remaining one. The correct option is one minus the sum of the given probabilities. Most questions have three numeric options (about 10% have two); the distractors are other nearby probabilities.",answer,One correct option per question part (`correct_index`); score is accuracy.,,self,regular,,1;2,200,200,5,5,held_out,
bayes_rule,Bayes' Rule,STEER,1.3.c,Bayes' Rule,Foundations,Probability,"Applying Bayes' rule to compute a conditional probability from a prior, a likelihood and a marginal.","Sports-themed questions give the probability of an event, the probability of the event under a condition, and the probability of the condition, then ask for the probability of the condition given that the event happened. The correct option is P(A|B) = P(B|A)P(A)/P(B), written as a percentage rounded to two decimals as the question asks; the distractors are common mistakes (inverting the conditional, the joint probability, leaving out the likelihood), at most one of them a number copied from the question.",answer,One correct option per question part (`correct_index`); score is accuracy.,,default,bayes_medical;bayes_sports,,1,200,200,5,5,held_out,
categorical_syllogism,Categorical Syllogism,STEER,1.4.a,Categorical Syllogism,Foundations,Logic,"Deciding which conclusion must follow from two categorical premises (""All A are B"", ""Some B are not C"").","A context sentence names a population (guests at a wedding, animals at a rescue centre, plants in a botanical garden, items in an antique shop, or made-up creatures such as wugs and blickets), then two premises in the forms All / No / Some / Some ... not, relating three groups in any of the four syllogistic figures. The question states that every group mentioned has at least one member and asks which conclusion must be true. The options are three conclusions about the two outer groups and ""Cannot be determined."" In half of the questions one listed conclusion must follow (the others listed do not); in the other half no conclusion relating the two outer groups follows, in either direction, and the key is ""Cannot be determined."" The groups are attributes that everyday knowledge does not relate (vaccinated and microchipped animals, wooden and imported items) or made-up words, so only the premises decide.",answer,One correct option per question part (`correct_index`); score is accuracy.,,default,animals_plants;fictional;objects;people,,1,200,200,5,5,held_out,
conditional_syllogism,Conditional Syllogism,STEER,1.4.b,Conditional Syllogism,Foundations,Logic,"Deciding what follows from an if-then rule and a second premise, avoiding affirming the consequent and denying the antecedent.","A rule ""If P, then Q"" from a fixed bank of 39 (20 about a named man or woman: a first-class ticket and the airport lounge, a fever and staying home; 19 about events, numbers or shapes: rain and the picnic, a number divisible by 6 and an even number, a regular hexagon and six sides), followed by a second premise: P, not-Q, Q or not-P, in equal shares. The question asks which conclusion must be true. The options are the two clauses of the other part of the rule (for example ""Clara can use the airport lounge."" and ""Clara cannot use the airport lounge."") and ""Cannot be determined."" The key is Q after P (modus ponens), not-P after not-Q (modus tollens), and ""Cannot be determined."" after Q (affirming the consequent) or not-P (denying the antecedent), so half the keys are ""Cannot be determined."" and the fallacious conclusion is always offered. The number and shape rules are true but their converses are not, so general knowledge never settles a fallacy. The type is the argument form.",answer,One correct option per question part (`correct_index`); score is accuracy.,,affirming_consequent;denying_antecedent;modus_ponens;modus_tollens,daily_life;finance;health;numbers;school;sports;travel;work,man;woman,1,640,640,20,20,held_out,
equiv_contrapositive,Logical Equivalence of Contrapositive,STEER,1.4.c,Logical Equivalence of Contrapositive,Foundations,Logic,Recognising that a conditional statement is equivalent to its contrapositive.,"A person or an expert asserts ""if P, then Q"" in an everyday, financial or medical setting (429 contexts, normalised by fixed rules). Most questions offer three options, about one in ten four: the contrapositive (""if not Q, then not P"") and two or three of the converse (""if Q, then P""), the inverse (""if not P, then not Q"") and the statement with its conclusion negated (""if P, then not Q""). Only the contrapositive is logically equivalent. Every option is built from the same four clauses as the statement (P, Q and their negations, each negation in the clause's own tense and with its own subject), so the key cannot be told apart by its wording.",answer,One correct option per question part (`correct_index`); score is accuracy.,,other,daily_life;finance;medicine,,1,200,200,5,5,held_out,
first_order_false_belief,First-Order False Belief,STEER,1.5.a,First-Order False Belief,Foundations,Theory of Mind,Predicting where a character will look for an object that was or was not moved in their sight.,"Unexpected-transfer stories over ten fixed scenarios (a marble in a basket, a stapler in a drawer, a car key on a hook, a sun hat in a beach bag, ...): a character puts an object in one place, and someone moves it to another. In half of the questions the character is away and is not told, so the key is the original place (a false belief); in the other half the character watches through the window, stays in the room and sees the move, or reads a message from the mover saying where the object is now, so the key is the new place. The question asks where the character will look first or thinks the object is; the two options are the two places. Each story is told with a named mover or with the reader as the mover. Secret-keeping questions (""you were told something in confidence; what should you do?"") are not included: what one should do is a norm, not something the story determines.",answer,One correct option per question part (`correct_index`); score is accuracy.,,false_belief;true_belief,garden;home;library;office;restaurant;school;travel;workplace,man;second_person;woman,1,320,320,10,10,held_out,
second_order_false_belief,Second-Order False Belief,STEER,1.5.b,Second-Order False Belief,Foundations,Theory of Mind,"Reasoning about where one character thinks another will look for an object that was moved twice, when the first may not know what the second has learned.","Unexpected-transfer stories over the ten scenarios of first_order_false_belief, with a third place for each. The owner puts an object in one place and leaves; an observer stays in the room and watches while a third person (or the reader) moves the object to a second place and later to a third. After each move the owner may learn of it openly (phoned in front of the observer), secretly (a message the observer does not know about), or not at all; after the second move the mover may also tell the observer, mistakenly, that the owner has been told. The question asks where the observer thinks the owner will look. In observer_wrong questions (five of the seven story variants) the observer's belief about the owner is false, so the key differs from where the owner will really look; in observer_right questions the observer is right but the owner is wrong, so the key differs from where the object is and from what the observer believes about it. Each of the three places is the key in about a third of the questions, so neither the place where the object was first put nor the place where it is answers the questions. Options are the three places. Third person, and second person with the reader as the mover.",answer,One correct option per question part (`correct_index`); score is accuracy.,,observer_right;observer_wrong,garden;home;library;office;restaurant;school;travel;workplace,man;second_person;woman,1,320,320,10,10,held_out,
completeness,Completeness,STEER,2.1.a,Completeness,Decisions in Single-Agent Environments,Axioms of Utility in Deterministic Environments,"Stating a preference, or indifference, between every pair of four or five fully described alternatives instead of declining to choose.","Each question describes four or five alternatives, certain outcomes that trade off against each other (job offers, apartments, laptops, holiday packages, medical treatments, savings bonds), and says that the description is complete: the alternatives differ in nothing else and no more information is available. It has one part per pair of alternatives (6 or 10 parts); every part shows the same description and asks about one pair, with four options: prefer the first, prefer the second, indifferent, or ""I cannot decide between X and Y"". Any preference or indifference is acceptable and no option is correct. The parts are graded together: the answers are complete if no pair is answered ""cannot decide"". Whether the answered pairs are also transitive is reported besides. Second person (the reader's own preference).",consistency,"Consistency-graded with rule `pairwise_complete`: Four or five fully described alternatives (sure outcomes, or lotteries); one part per pair, each with the options prefer one, prefer the other, indifferent, and cannot decide / need more information. Consistent: no pair answered ""cannot decide"" (the answers are complete). Reported besides: whether the answered pairs are transitive. Score: share of consistent groups.",pairwise,pairwise,finance;housing;jobs;medical;technology;travel,second_person,1;2,240,1920,5,30,held_out,
transitivity,Transitivity,STEER,2.1.b,Transitivity,Decisions in Single-Agent Environments,Axioms of Utility in Deterministic Environments,"Choosing consistently with transitive preferences: if A is preferred to B and B to C, then A is preferred to C.","A short scene (job offers in several cities, a holiday country, fruit for a salad, a partner for a group project, a Saturday activity) followed by two or three stated pairwise preferences over three or four alternatives, in random order and worded four ways (""prefers A over B"", ""likes A more than B"", ""likes B less than A"", ""finds B less appealing than A""). The preferences form a chain, or with four alternatives a fork or a join. The two options are a pair that is never compared directly but is ordered by chaining the stated preferences; the question asks which of the two is preferred, or, in half of the questions, which is liked less. Pairs the statements leave unordered are never asked and no statements are cyclic. Each scene is told in the second person, about a named man or woman, and with the reader as an AI assistant acting for a named client.",answer,One correct option per question part (`correct_index`); score is accuracy.,,least_preferred;most_preferred,activities;countries;fruits;jobs;people,ai_assistant;man;second_person;woman,1,200,200,10,10,held_out,
independence,Independence,STEER,2.1.c,Independence,Decisions in Single-Agent Environments,Axioms of Utility in Deterministic Environments,"Making choices that are not changed by adding an irrelevant, dominated alternative (the decoy effect).","Each question has two parts, with scenario sentences recovered from the v1 decoy templates (1,074 everyday- and technological-product scenarios that name no attribute other than price and quality). Part 0 offers brands A and B with a price and a quality rating: one cheaper, the other better. Part 1 repeats A and B and adds brand C, the standard asymmetric decoy: worse than the target (A or B, at random) on both price and quality, but better than the other brand on one of them. Both parts also offer ""No preference between Brand A and Brand B"". Any part-0 choice is acceptable, but the part-1 choice must be the same (the same brand, or no preference again), and C is never a consistent choice.",consistency,"Consistency-graded with rule `same_brand`: Part 0 offers two brands (price and quality) and ""no preference""; part 1 adds a decoy that is worse than one of them on both. Consistent: the same brand in both parts, or no preference in both. Score: share of consistent groups.",asymmetric_decoy,asymmetric_decoy,everyday_products;technological_products,second_person,1,200,400,5,10,held_out,
sunk_cost,Avoidance of the Sunk Cost Fallacy,STEER,2.2.a,Avoidance of Sunk Cost Fallacy,Decisions in Single-Agent Environments,Avoidance of Cognitive Biases in Deterministic Environments,Making forward-looking decisions that ignore costs already spent and unrecoverable.,"Twenty fixed scenarios (an earbud prototype, a half-built hotel, a television campaign, a feature film, an apple orchard, a non-refundable retreat venue, a database migration by hand, ...), fifteen about money and five about hours. Each states the amount already spent and that it cannot be recovered, and for each course of action its further cost and future value: continue the original plan, switch to an alternative, or stop (no further cost, no further value). The key is the option with the highest future net value, at least three units of the scenario's scale ahead; it is continuing, switching or stopping, each a third of the time. When it is switching or stopping, counting the sunk amount as lost would make continuing look best (the sunk-cost trap); the sunk amount is drawn independently of the future values. Time savings are stated as totals over the coming year. When the key is continuing or switching, it costs more and is worth more than the other course (or less and less), so it does not win on both. Each scenario is told in the first, second and third person (a man, a woman) and with the reader as an AI assistant in charge of the decision.",answer,One correct option per question part (`correct_index`); score is accuracy.,,money;time,agriculture;construction;education;entertainment;events;food_service;manufacturing;marketing;medical;personal_finance;research;retail;technology;travel;workplace,ai_assistant;first_person;man;second_person;woman,1,320,320,10,10,held_out,
endowment_effect,Avoidance of the Endowment Effect,STEER,2.2.b,Avoidance of Endowment Effect,Decisions in Single-Agent Environments,Avoidance of Cognitive Biases in Deterministic Environments,"Valuing a good the same whether or not you already own it, so mirrored trade offers get opposite answers.","Each question has two parts: the same short story told twice with the two items swapped. In part 0 you own item X and are offered item Y of the same stated value in exchange; in part 1 you own Y and are offered X. Each part stands on its own, says the exchange costs nothing and takes no effort, and is answered Yes, No or indifferent. The items are 30 fixed pairs of similar kind and price (a ceramic mug and a bar of Swiss chocolate, wireless earbuds and a portable speaker, tickets to a jazz concert and to a comedy show, ...), and the stories are a conference gift, a raffle prize, a present that the shop will exchange, a thank-you for a survey, a charity shop's stock managed by the reader as an AI assistant, and advice to a named man or woman. With a preference that does not depend on which item is owned, exactly one of the two offers is accepted, so Yes then No or No then Yes is consistent, and so is indifferent in both parts; No in both parts is the endowment effect.",consistency,"Consistency-graded with rule `mirror_trade`: You own item X and are offered Y; then you own Y and are offered X (Yes / No / indifferent). Consistent: Yes in exactly one part, or indifferent in both. Score: share of consistent groups.",mirror_trade,mirror_trade,advice;charity_shop;gift;raffle;store;survey,ai_assistant;second_person;woman,1,200,400,5,10,held_out,
time_inconsistency,Avoidance of Time Inconsistency,STEER,2.2.c,Avoidance of Time Inconsistency,Decisions in Single-Agent Environments,Avoidance of Cognitive Biases in Deterministic Environments,Keeping the same choice between a smaller-sooner and larger-later reward when both are pushed into the future.,"The question states the agent's earlier choice between an amount today and a larger amount tomorrow, then offers the same two amounts shifted by many days (e.g. day 53 vs day 54). The keyed answer is the option that matches the earlier choice, as an exponential (time-consistent) discounter would choose.",answer,One correct option per question part (`correct_index`); score is accuracy.,,today;tomorrow,present_bias,,1,200,200,10,10,held_out,
completeness_risk,Completeness over Lotteries,STEER,2.3.a,Completeness over Lotteries,Decisions in Single-Agent Environments,Axioms of Utility in Stochastic Environments,"Stating a preference, or indifference, between every pair of four or five fully described lotteries instead of declining to choose.","Each question describes four or five alternatives, lotteries stated with exact chances (game-show envelopes, one-year investments, treatments with a chance of a faster recovery, bonus plans with three outcomes, crops whose earnings depend on the weather, travel routes with a chance of delay), and says that the description is complete: the alternatives differ in nothing else and no more information is available. It has one part per pair of alternatives (6 or 10 parts); every part shows the same description and asks about one pair, with four options: prefer the first, prefer the second, indifferent, or ""I cannot decide between X and Y"". Any preference or indifference is acceptable and no option is correct. The parts are graded together: the answers are complete if no pair is answered ""cannot decide"". Whether the answered pairs are also transitive is reported besides. Second person (the reader's own preference).",consistency,"Consistency-graded with rule `pairwise_complete`: Four or five fully described alternatives (sure outcomes, or lotteries); one part per pair, each with the options prefer one, prefer the other, indifferent, and cannot decide / need more information. Consistent: no pair answered ""cannot decide"" (the answers are complete). Reported besides: whether the answered pairs are transitive. Score: share of consistent groups.",pairwise,pairwise,agriculture;finance;game_show;medical;travel;workplace,second_person,1;2,240,1920,5,50,held_out,
transitivity_risk,Transitivity over Lotteries,STEER,2.3.b,Transitivity over Lotteries,Decisions in Single-Agent Environments,Axioms of Utility in Stochastic Environments,Applying transitivity to preferences between lotteries.,"Two stated preferences chain three lotteries (e.g. a 44% chance of X preferred to an 83% chance of Y, which is preferred to a 54% chance of Z). The model chooses among the transitive conclusion, its reverse and ""Cannot determine preference""; the transitive conclusion is correct.",answer,One correct option per question part (`correct_index`); score is accuracy.,,single,food;gadgets;investments;vacations,,1,200,200,5,5,held_out,
independence_risk,Independence over Lotteries,STEER,2.3.c,Independence over Lotteries,Decisions in Single-Agent Environments,Axioms of Utility in Stochastic Environments,Applying the independence axiom: mixing two options with the same third lottery should not reverse the preference.,"The agent prefers item A to item B. Option X gives A with probability p and C otherwise; Option Y gives B with probability p and the same C otherwise. The correct choice is Option X, the one built on the preferred item (sure-thing reasoning). The lottery with the preferred item is described first in half of the questions and second in the other half.",answer,One correct option per question part (`correct_index`); score is accuracy.,,sure_thing,book;car;gadget;investment;meal;vacation,,1,200,200,5,5,held_out,
compute_expected_utility,Compute Expected Utility,STEER,2.4.a,Compute Expected Utility,Decisions in Single-Agent Environments,Risk Neutral Expected Utility Computation,Computing the expected utility of a lottery from stated probabilities and utilities.,"A treatment, job, vacation, investment or lottery has two or three outcomes with stated probabilities and utilities, and the question asks for its expected utility. The correct option is the probability-weighted sum of the utilities, shown among three numeric options; the distractors (unweighted mean, probabilities swapped) lie between the smallest and largest utility.",answer,One correct option per question part (`correct_index`); score is accuracy.,,default,investment;job;lottery;medical;vacation,,1;2,200,200,5,5,held_out,
maximize_expected_utility,Maximize Expected Utility,STEER,2.4.b,Maximize Expected Utility,Decisions in Single-Agent Environments,Risk Neutral Expected Utility Computation,Choosing the option with the highest expected value when utility is linear in money.,"Each question states that the chooser's utility is linear in money and offers a sure amount and one or two lotteries with whole-percentage probabilities (game shows, casinos, contests), or two job offers: a sure salary, or a lower salary with a chance of a bonus. Every lottery has one outcome above and one below the sure amount; the key is the option with the highest expected value, which beats the others by at least 5% (2% for the job offers). Only this plain type is generated: the certainty and reflection framings are their own elements, and the loss-aversion framing is not used (Avoidance of Loss Averse Behavior has its own questions).",answer,One correct option per question part (`correct_index`); score is accuracy.,,regular,gambling;job_offer,second_person,1;2,200,200,5,5,held_out,
loss_aversion,Avoidance of Loss Averse Behavior,STEER,2.4.e,Avoidance of Loss Averse Behavior,Decisions in Single-Agent Environments,Risk Neutral Expected Utility Computation,Taking a mixed gamble with positive expected value despite a framing that invites loss-averse rejection.,"Five stories (a coin-flip bet with a friend, a marketing campaign, a short-term trade, a game-show wheel, an optional bonus / deduction pay scheme) offer a gamble that wins $G with stated probability and otherwise loses $L of the decision maker's own money. The probability-weighted gain is 1.1 to 2 times the weighted loss, so the expected value is positive but a loss-averse chooser would reject. Every question says the decision maker faces many small decisions like this, the amounts are small compared with their savings, and they care about the average result. In accept_reject questions the alternative is doing nothing; in a third of them the gamble has a negative expected value (weighted gain 0.5 to 0.87 times the weighted loss) and declining is the key; in sure_amount questions it is a small guaranteed payment, which is the better choice in about a third of them. Six perspectives, including an AI assistant advising a client.",answer,One correct option per question part (`correct_index`); score is accuracy.,,accept_reject;sure_amount,business;gambling;game_show;investing;workplace,ai_assistant;anon;first_person;man;second_person;woman,1;2,200,200,10,10,held_out,
gamblers_fallacy,Avoidance of the Gambler's Fallacy,STEER,2.5.a,Avoidance of Gambler's Fallacy,Decisions in Single-Agent Environments,Avoidance of Cognitive Biases in Stochastic Environments,Recognising that a streak of independent random outcomes does not change the odds of the next one.,"After a streak at roulette, dice, slot machines, lotteries or with coin flips, the question asks whether the streak has made some outcome more (or less) likely on the next trial. Templates where ""Yes"" is defensible (a ""different outcome"" being more likely than the streak outcome), card games dealt from a shoe or deck (not independent), and questions with no comparison are not used. The correct answer is ""No"": independent draws keep the same probabilities.",answer,One correct option per question part (`correct_index`); score is accuracy.,,type_one,casino;coin,,1,200,200,5,5,held_out,
certainty_effect,Avoidance of the Certainty Effect,STEER,2.5.b,Avoidance of the Certainty Effect,Decisions in Single-Agent Environments,Avoidance of Cognitive Biases in Stochastic Environments,"Choosing consistently across Allais-style lottery pairs, without overweighting outcomes that are certain.","Each question states an earlier choice between X for sure and Y with probability p, then offers X with probability k and Y with probability k*p (probabilities scaled by a common factor). The question says that the preferences have not changed since then. Under expected utility the new choice should be on the same side as the earlier one, so the keyed option is the scaled version of the option chosen before.",mixed,"One correct option per question part (`correct_index`); score is accuracy. In type `stated_prior` the question states which option the reader chose in an earlier, related choice; the keyed option is the one on the same side. Consistency-graded (type `two_part`) with rule `common_ratio_same_side`: Part 0: a sure amount or a lottery; part 1: both scaled down by the same probability. Consistent: the sure option and its scaled-down version, or the lottery in both parts. Score: share of consistent groups.",two_part,stated_prior;two_part,employment;finance;gambling;medical,anon;first_person;man;second_person;woman,1,200,300,10,15,examples,
reflection_effect,Avoidance of the Reflection Effect,STEER,2.5.c,Avoidance of the Reflection Effect,Decisions in Single-Agent Environments,Avoidance of Cognitive Biases in Stochastic Environments,Keeping the same risk attitude when a choice between gains is mirrored into a choice between losses.,"Each question states an earlier choice between a sure gain X and a gain Y with probability p (equal expected value, p*Y = X), then offers the loss mirror: lose X for sure or have to pay Y with probability p. The question says that the earlier choice was a strict preference and the attitude to risk has not changed. With one strict risk attitude across gains and losses, the keyed option is on the same side (sure or risky) as the earlier choice.",mixed,"One correct option per question part (`correct_index`); score is accuracy. In type `stated_prior` the question states which option the reader chose in an earlier, related choice; the keyed option is the one on the same side. Consistency-graded (type `two_part`) with rule `reflection_same_side`: Part 0: a sure gain or a lottery over gains; part 1: the mirrored losses, with equal expected values. Consistent: the same side in both parts. Score: share of consistent groups.",two_part,stated_prior;two_part,employment;finance;gambling;medical,anon;first_person;man;second_person;woman,1,200,300,10,15,examples,
ambiguity_aversion,Avoidance of Ambiguity Aversion,STEER,2.5.d,Avoidance of Ambiguity Aversion,Decisions in Single-Agent Environments,Avoidance of Cognitive Biases in Stochastic Environments,Making Ellsberg-urn choices that can be explained by a single belief about the unknown proportions.,"Ellsberg's three-colour urn in two parts, recovered from the v1 templates (414 balls-and-urn and mythical-treasure templates; the 165 'goods' templates name no colours and were not used). The urn holds n items of a known colour and 2n of two other colours in unknown proportions (n = 1-100); every bet pays the same reward. Part 0: bet on the known colour or on one unknown colour. Part 1: bet on the two unknown colours together, or on the known colour plus the second unknown colour. Part 0 says one item is drawn at random, each equally likely. Each part also offers ""Indifferent between the two options"". A pair of choices is consistent if a single subjective probability justifies it: the known-probability bet in exactly one of the two parts, or indifferent in both (p = 1/3). The classic ambiguity-averse pattern (the known-probability bet in both parts) is inconsistent.",consistency,"Consistency-graded with rule `ellsberg_seu`: Ellsberg's three-colour urn: two pairs of bets, one with known and one with unknown odds, each with an ""indifferent"" option. Consistent: the pairs that one subjective probability can explain (the known-odds bet in exactly one part, or indifferent in both). Score: share of consistent groups.",two_part,two_part,mythical;objects,second_person,1,200,400,5,10,held_out,
interpret_games,Interpret Games,STEER,3.1.a,Interpret Games,Decisions in Multi-Agent Environments,Normal Form Games,Reading a payoff from a normal-form game matrix for a given pair of actions.,"A small payoff matrix (yours or your opponent's) is given, with the model playing as the row or column player, and the question asks for a specific payoff when each player takes a named action. The correct option is the matching cell value.",answer,One correct option per question part (`correct_index`); score is accuracy.,,column;row,other;self,,1;2,200,200,10,10,held_out,
best_response,Best Response,STEER,3.1.b,Best Response,Decisions in Multi-Agent Environments,Normal Form Games,Finding a player's best response to a given action of the other player in a bimatrix game.,"A 2x2 or 3x3 bimatrix game is shown and the model is told which player it is. The question asks for its own or its opponent's best response to a specified action of the other player. The correct option is the action with the highest payoff for the responding player in that row or column, among that player's two or three actions; games with ties are excluded.",answer,One correct option per question part (`correct_index`); score is accuracy.,,column;row,other;self,,1;2,200,200,0,0,none,

dominant_strategies,Dominant Strategies,STEER,3.1.c,Dominant Strategies,Decisions in Multi-Agent Environments,Normal Form Games,"Identifying and playing a dominant strategy in a two-player, two-action game described in words.","Two kinds of story, each built from a payoff table generated first. Pricing stories (symmetric and non-symmetric): two shops each charge a low or a high price, the story lists the units each sells in all four price combinations and states the objective, profit = price x units sold, each maximising its own. Negotiation stories: two new hires each negotiate or accept an offer; the story lists the raise each gets in all four combinations and says each wants the largest raise for themselves. The question asks for the named player's dominant strategy (two options). The asked player always has exactly one strictly dominant action under the stated objective, high or low price (negotiate or accept) equally often.",answer,One correct option per question part (`correct_index`); score is accuracy.,,non-symmetric;symmetric,finance;job_market,,1;2,200,200,10,10,held_out,

dominated_strategies,Avoidance of Dominated Strategies,STEER,3.1.d,Avoidance of Dominated Strategies,Decisions in Multi-Agent Environments,Normal Form Games,Choosing the undominated action in a small game shown as a payoff matrix.,"A 2x2 or 3x3 game is shown, either as one player's payoff matrix or as a bimatrix with both players' payoffs, and the question asks which action the model or its opponent should play. One action strictly dominates the others, and it is the correct option among that player's two or three actions; when the question asks for the dominated action, no other action is dominated weakly or by a mixture.",answer,One correct option per question part (`correct_index`); score is accuracy.,,other;other_dominated;self;self_dominated,normal_form,,1;2,200,200,20,20,held_out,

iterated_removal,Iterated Removal of Dominated Strategies,STEER,3.1.e,Iterated Removal of Dominated Strategies,Decisions in Multi-Agent Environments,Normal Form Games,Eliminating dominated actions step by step in a 3x3 game.,Each question has two parts. Part 0 shows a 3x3 bimatrix game and asks which of one player's actions is dominated; part 1 asks which of the other player's actions should not be played once that action is removed. Both parts have one correct action.,answer,One correct option per question part (`correct_index`); score is accuracy.,,column;row,other;self,,1;2;3,240,480,0,0,none,
pure_nash,Pure Nash Equilibrium,STEER,3.1.f,Pure Nash Equilibrium,Decisions in Multi-Agent Environments,Normal Form Games,Finding the pure-strategy Nash equilibrium of a game that has exactly one.,"Three kinds of question, each built from a payoff table generated first. Matrix questions show a 3x3 payoff matrix (integers 0-19 or decimals below 20) and ask, from the row or the column player's side, for your equilibrium action, your opponent's, or the action pair. Pricing stories: two shops each charge a low or a high price, the story lists how many units each sells in all four price combinations, and profit is price times units sold. Salary stories: two employers each offer a low or a high salary; equal offers split the candidates evenly, otherwise the higher offer hires a stated share, and profit is hires times (value per hire minus salary). Word problems ask for the equilibrium pair (with 'There is no pure-strategy Nash equilibrium' as a wrong option) or one firm's equilibrium action. Every game has exactly one pure equilibrium, with strict best responses, checked by brute force.",answer,One correct option per question part (`correct_index`); score is accuracy.,,column;non-symmetric;row;symmetric,finance;job_market;matrix_form,,1;2,200,200,20,20,held_out,
backward_induction,Backward Induction,STEER,3.2.a,Backward Induction,Decisions in Multi-Agent Environments,Extensive Form Games,Solving a short sequential game by backward induction to decide whether a player should stop or continue.,"A three-round centipede-style negotiation or treatment decision gives each player's payoff at every stopping point. The question asks whether the first mover should end the game in round one or, in a little over half the questions, whether the second mover should end it in round two. The correct option follows from solving the tree from the last decision back, with each mover maximising their own payoff.",answer,One correct option per question part (`correct_index`); score is accuracy.,,sequential,finance;medical,,1;2,200,200,5,5,examples,
subgame_nash,Subgame-Perfect Nash Equilibrium,STEER,3.2.b,Subgame-Perfect Nash Equilibrium,Decisions in Multi-Agent Environments,Extensive Form Games,Finding subgame-perfect play in a stage game repeated a fixed number of times.,"A 2x2 or 3x3 bimatrix game is repeated for two, three or four rounds, and the question asks what is played in each round if both players mutually best respond: the model's own actions, its opponent's actions, or the action pairs. The correct option repeats the stage game's Nash equilibrium in every round.",answer,One correct option per question part (`correct_index`); score is accuracy.,,column;row,both;other;self,,1;2;3;4,480,480,10,10,held_out,
feasibility,Feasibility in Infinitely Repeated Games,STEER,3.4.a,Feasibility in Infinitely Repeated Games,Decisions in Multi-Agent Environments,Infinitely Repeated Games,"Identifying which payoffs an infinitely repeated prisoner's dilemma can sustain: the feasible, individually rational payoffs of the folk theorem, or the equilibrium plans at a stated discount factor.","Six fixed stories (petrol-station prices, fishing limits, advertising, oil quotas, data sharing between labs, irrigation) state the payoffs per period of a symmetric prisoner's dilemma in the story's own words (both cooperate, both defect, one of each), in whole thousands of dollars; by the story's labels the payoffs always form a genuine prisoner's dilemma. In folk questions no discount factor is given. The question asks which pair of long-run average payoffs is feasible (a long-run average of the four outcomes) and individually rational (above the mutual-defection payoff), or, in the other wording, which pair the folk theorem says a subgame-perfect equilibrium can sustain when both players are very patient. Exactly one option qualifies; the others are infeasible or not individually rational (at least one of each). In discount questions a discount factor and grim-trigger punishment (both defect forever after any departure) are stated, and the options are four repeating plans (cooperate every period; take turns defecting; cooperate and defect in alternate periods; two periods of cooperation, then one of defection; a three-period cycle; one player always cooperating while the other alternates). Exactly one is a subgame-perfect equilibrium at the stated discount factor, which is at least 0.02 from every plan's threshold.",answer,One correct option per question part (`correct_index`); score is accuracy.,,discount;folk,agriculture;business;energy;environment;marketing;research,anon,1;2,240,240,10,10,held_out,
enforceability,Enforceability in Infinitely Repeated Games,STEER,3.4.b,Enforceability in Infinitely Repeated Games,Decisions in Multi-Agent Environments,Infinitely Repeated Games,Deciding which outcome of a repeated prisoner's dilemma can be sustained in equilibrium at a given discount factor.,"Two firms play a repeated prisoner's dilemma with stated payoffs and discount factor. Options describe outcome paths: both cooperate, both defect, one always defects, players take turns (each period one defects while the other cooperates, swapping roles every period), or none of these. The keyed option is the one path that is a subgame-perfect equilibrium with grim-trigger reversion to mutual defection at that discount factor.",answer,One correct option per question part (`correct_index`); score is accuracy.,,symmetric,finance,,1,200,200,5,5,examples,

trigger,Trigger Strategies,STEER,3.4.c,Trigger Strategies,Decisions in Multi-Agent Environments,Infinitely Repeated Games,Computing the minimum discount factor at which grim-trigger cooperation is subgame perfect in a repeated prisoner's dilemma.,"A stage game with payoffs R (both cooperate), S and T (one defects) and P (both defect) is given, as a story or a matrix. The correct option is the threshold (T - R) / (T - P), rounded to two decimals as the question asks; distractors are other common formulas.",answer,One correct option per question part (`correct_index`); score is accuracy.,,discount_value,gambling;matrix_form;politics;technology,,1,200,200,5,5,examples,
bayes_nash,Bayes-Nash Equilibrium,STEER,3.5.a,Bayes–Nash Equilibrium,Decisions in Multi-Agent Environments,Bayesian Games,Finding the Bayes-Nash equilibrium of a 2x2 game whose payoffs depend on a random state.,"The opponent has a private type (Type 1 or Type 2, with stated probabilities) and a 2x2 payoff table per type. The question asks for your action, or for the opponent's action when of a stated type, in the Bayes-Nash equilibrium; games are drawn so that there is exactly one equilibrium (pure, with no ties and no mixed equilibrium).",answer,One correct option per question part (`correct_index`); score is accuracy.,,typed_opponent;typed_own,matrix_form,,1;2,200,200,10,10,held_out,
pareto_sc,Pareto Efficiency in Social Welfare Functions,STEER,4.1.a,Pareto Efficiency,Decisions on Behalf of Other Agents,Axioms of Social Choice,Checking whether a group decision respects the Pareto principle given everyone's rankings.,"A group (a book club, a city council, a band, a family, an office team, a school board) ranks two or three alternatives, and the profile is listed as ballot lines such as ""12 voters voted for D > K > T"". Half the questions report the group's final decision and ask Yes/No whether it satisfies the Pareto efficiency axiom: the answer is No exactly when every member ranks some other alternative above the chosen one (profiles where that differs from one alternative being ranked above it by everyone are not drawn). The other half state a unanimous preference (every member ranks one alternative first) and ask which alternative should be chosen.",answer,One correct option per question part (`correct_index`); score is accuracy.,,which;yes_no,education;entertainment;family;leisure;politics;workplace,,1,240,240,10,10,held_out,

sc_monotonicity_rules,Monotonicity in Social Welfare Functions,STEER,4.1.b,Monotonicity in Social Welfare Functions,Decisions on Behalf of Other Agents,Axioms of Social Choice,"Spotting which of four rules that rank the options violates monotonicity: after one option is moved up on some ballots, an option that was ranked below it overtakes it.","A group (a book club, a city council, a band, a family, an office team or a school board) tries four unnamed rules that turn the members' ranked ballots into a group ranking of all options, first on one set of ballots and then on the same ballots after some members moved one option up, with nothing else changed. A rule violates monotonicity when an option that the group ranked below the raised option before the change is ranked above it after the change (the paper's definition: society prefers A to B, A is raised in some individual rankings, and A must still be preferred to B). The rules come from real ones (Borda, plurality and Copeland rankings, and the elimination orders of instant runoff and of Coombs' method) and made-up but well-defined ones (fewest last places; most first-or-second places; the two with most first places on top in head-to-head order; the two highest Borda scores on top, ordered by first places; the two with most first places on top, ordered by Borda points). In consequences questions (easier) each option shows one rule's ranking before and after the change; in procedures questions (harder) the rules are described in words and the reader applies them. Exactly one rule violates monotonicity, and in it the raised option keeps or improves its place (it overtakes one option and is overtaken by another), so the violation shows only when the pairs are compared. In consequences questions the four before / after pairs differ, another rule's ranking changes and in another rule the raised option moves up; in procedures questions another listed rule is of a kind that can violate monotonicity but does not on these ballots. Half of the questions state what monotonicity requires.",answer,One correct option per question part (`correct_index`); score is accuracy.,,consequences;procedures,education;entertainment;family;leisure;politics;workplace,ai_assistant;anon;first_person;man;second_person;woman,1;2,240,240,10,10,held_out,
sc_transitivity_rules,Transitivity in Social Welfare Functions,STEER,4.1.c,Transitivity in Social Welfare Functions,Decisions on Behalf of Other Agents,Axioms of Social Choice,"Finding which of four rules turns a profile into a social ranking that is not transitive, or the only one that is.","A group tries four unnamed rules that decide, for every pair of options, which one the group places above the other. Pairwise rules can produce cycles (pairwise majority; more members ranking one option immediately above the other; majority among the members who rank neither of the two last, or neither of the two first); score rankings cannot (Borda points, first places, fewest last places, first-or-second places). The ballots often contain a Condorcet cycle. In consequences questions (easier) each option lists one rule's verdict on every pair; in procedures questions (harder) the rules are described and the reader works out the verdicts. The intransitive direction asks which rule's verdicts are not a well-defined (transitive) ranking (exactly one is cyclic); the transitive direction says only one is and asks which. In procedures questions the key is never recognisable from the kind of rule alone.",answer,One correct option per question part (`correct_index`); score is accuracy.,,consequences;procedures,education;entertainment;family;leisure;politics;workplace,ai_assistant;anon;first_person;man;second_person;woman,1;2,240,240,10,10,held_out,
sc_nondictatorship,Non-Dictatorial Social Welfare Function,STEER,4.1.d,Non-Dictatorial Social Welfare Function,Decisions on Behalf of Other Agents,Axioms of Social Choice,"Telling a non-dictatorial rule from dictatorships in disguise, from a rule's wording or from its results.","A group's three-member committee (each member with a role, an age and, where needed, a registration order) ranks shortlisted options with strict ballots. Describe questions (easy) list four rules: three are dictatorships in disguise (the chair's ballot, the ballot of whoever registered first, the oldest member's ballot with the next one only breaking ties, the secretary's ballot read top to bottom, the alphabetically first member's ballot) and one is not (Borda points, first places or fewest last places, with a stated tie-break). Consequences questions (medium) give three or four meetings with the members' ballots and four unnamed rules' results, and ask which rule cannot be a dictatorship: three rules copied one member every time, the key differs from every member at some meeting. Procedures questions (hard) describe one subtle procedure (e.g. the two youngest decide if they agree, otherwise the oldest; a weighted Borda count with a stated weight and tie-break; a weighted pairwise vote) and ask whether some member's ballot always becomes the committee's ranking, naming the member; half have a dictator.",answer,One correct option per question part (`correct_index`); score is accuracy.,,consequences;describe;procedures,education;entertainment;family;leisure;politics;workplace,ai_assistant;anon;first_person;man;second_person;woman,1;2;3,360,360,15,15,held_out,
plurality_voting,Plurality Vote,STEER,4.2.a,Plurality Vote,Decisions on Behalf of Other Agents,Social Choice,Determining the winner of a plurality election from a table of ranked ballots.,An election with four or five candidates lists groups of voters and their rankings. The correct option is the candidate with the most first-place votes.,answer,One correct option per question part (`correct_index`); score is accuracy.,,4;5,voting,,1;2,200,200,10,10,examples,
borda_count,Borda Count,STEER,4.2.b,Borda Count,Decisions on Behalf of Other Agents,Social Choice,Tallying Borda scores from ranked ballots to find the winner or a given place.,"Groups of voters rank some or all candidates. The question asks which candidate wins, or finishes in a given position, under the Borda count. The question states the scoring: each ballot ranks three candidates, who get n-1, n-2 and n-3 points in order, and a candidate left off a ballot gets 0.",answer,One correct option per question part (`correct_index`); score is accuracy.,,4;5,voting,,1;2,200,200,10,10,held_out,
copeland_method,Copeland's Method,STEER,4.2.c,Copeland's Method,Decisions on Behalf of Other Agents,Social Choice,Finding the winner under Copeland's method: one point for each one-on-one contest won and half a point for each tie.,"A group (a book club, a city council, a band, a family, an office team or a school board) chooses among three, four or five alternatives. In profile questions the members' full rankings are given as ballot lines; in table questions only the results of the one-on-one comparisons are given. Half of the questions spell out the rule (one point for each one-on-one contest won, half a point for each tie, most points wins), the others only name Copeland's method. The correct option is the alternative with the highest Copeland score, which is always unique. In about 60% of questions that alternative is not a Condorcet winner (it loses or ties a contest), and where possible it is neither the alternative with the most first places nor the Borda winner, so a simpler count does not give the answer. The options are the alternatives.",answer,One correct option per question part (`correct_index`); score is accuracy.,,profile;table,education;entertainment;family;leisure;politics;workplace,,1;2,480,480,10,10,held_out,
fair_division,Fair Division Algorithms,STEER,4.2.d,Fair Division Algorithms,Decisions on Behalf of Other Agents,Social Choice,"Computing the outcome of a named fair division procedure, or checking an allocation for envy-freeness and proportionality.","Each question names the procedure and gives additive valuations. Divide and choose: a 12-inch cake or a 60-metre strip of land is one kind up to a stated point and another after it, both people value each kind at a stated price per unit; the question asks where the divider cuts, which piece the chooser takes and what it is worth to the chooser, or what the divider's piece is worth to the divider. Last diminisher: three or four people, a cake of three flavour sections with each person's points; who gets the first piece and where it ends, or the final allocation (three people). Adjusted winner: two people spread 100 points over 4 to 6 items of an estate or a business; which item is split and the shares (distractors: the shares swapped, an even split, the leader's points only, another item split). Envy check: three people, 4 or 5 items, a proposed allocation; whether it is envy-free and proportional, proportional only, or neither.",answer,One correct option per question part (`correct_index`); score is accuracy.,,adjusted_winner;divider_chooser;envy_check;last_diminisher,business;food;inheritance;land,ai_assistant;anon;first_person;second_person,1;2,200,200,20,20,held_out,
condorcet_criterion,Condorcet Winner,STEER,,Condorcet Winner,Decisions on Behalf of Other Agents,Social Choice,"Finding the Condorcet winner, the alternative that beats every other in pairwise majority votes, if one exists.","Groups of voters rank three to five alternatives. The correct option is the alternative that wins every head-to-head comparison by a strict majority, or ""There is no Condorcet winner"" (about 25% of items).",answer,One correct option per question part (`correct_index`); score is accuracy.,,3;4;5,voting,,1;2,200,200,15,15,held_out,
dsic_mechanism,Dominant Strategy Incentive Compatibility,STEER,4.3.a,Dominant Strategy Incentive Compatibility,Decisions on Behalf of Other Agents,Desirable Properties of Mechanism Design,"Judging whether an auction, public-project, facility-location or matching rule, described in words, makes truthful reporting a dominant strategy for every participant.","A rule is described in words, never named, and the question states its assumptions: each participant alone knows their own value, favourite spot or ranking; where money changes hands, each cares only about what they get minus what they pay; on a line, nearer is better; rankings are strict. The question asks Yes/No whether reporting truthfully (bidding one's value, staying in a rising-price sale until the price reaches it, accepting a posted price exactly when it is worth it, reporting one's true value, spot or ranking) is a dominant strategy for every participant. Selling one item to three to five bidders: the highest bid wins and pays the best competing bid (Yes), the same with a minimum price announced in advance (Yes), a rising price with final, public drop-outs (Yes), a price fixed in advance (Yes), the winner pays their own bid (No), the third-highest bid (No), everyone pays their bid (No), a draw with chances proportional to the bids (No). A shared purchase split among three to five people: bought when the reported values cover the cost, with equal shares and a fee for a report that changes the decision equal to the others' net loss (Yes), the cost split in proportion to the reports (No), equal shares with no fee (No). A facility placed on a road: the middle report (Yes) or the average (No). Matching: a random order in which each takes their favourite remaining object (Yes), swaps along cycles of people pointing to the owner of their favourite object (Yes), school seats given to first choices first and for good (No), and proposals held and rejected until no one is rejected, asked about the proposing side (Yes) or the receiving side (No). Nine Yes and nine No mechanisms with the same number of questions. For the sealed-bid rules with an outcome fixed by the bids, about a third of the questions add a part with stated bids that asks for the winner and payment, the seller's revenue or a bidder's payoff. A further type gives favourite spots on the road and asks whether a named resident could get the facility closer by reporting a spot further east, further west, or not at all.",answer,One correct option per question part (`correct_index`); score is accuracy.,,auction_allpay;auction_clock;auction_first;auction_lottery;auction_posted;auction_reserve;auction_second;auction_third;location_mean;location_median;location_misreport;match_cycles;match_da_proposers;match_da_receivers;match_immediate;match_serial;project_equal;project_pivot;project_pledge,auction;facility_location;matching;public_project,ai_assistant;anon,1;2;3,380,415,95,100,held_out,
bic_mechanism,Bayesian Incentive Compatibility,STEER,4.3.b,Bayesian Incentive Compatibility,Decisions on Behalf of Other Agents,Desirable Properties of Mechanism Design,"Judging whether truthful reporting is a Bayes-Nash equilibrium of a mechanism described in words, and finding the first-price equilibrium bid.","Values are private and independent, with a stated uniform distribution, and participants are risk-neutral. A seller (an art gallery, a city, a farmer, a charity) or a group deciding on a shared project (residents, employees, parents, a rowing club) uses a mechanism that is described but not named. Yes: second-price sealed bids, a rising price clock with announced drop-outs, a report-based posted price, the expected-externality (AGV) public project, and a winner paying (n-1)/n of their own report; No: first-price and third-price sealed bids, a project paid by pledges, cost shares proportional to reports, and a winner paying another fraction. The last two Yes cases are Bayesian but not dominant-strategy incentive compatible, which separates this element from dsic_mechanism. Yes and No are equally common. A second type asks for a bidder's symmetric equilibrium bid in a first-price auction, (n-1)/n of the value.",answer,One correct option per question part (`correct_index`); score is accuracy.,,agv;bne_bid;english;first_price;pay_your_bid;posted_report;proportional_share;scaled_payment;second_price;third_price,agriculture;art;charity;community;government;school;sports;workplace,ai_assistant;anon;first_person;man;second_person;woman,1;2;3,800,800,50,50,held_out,
ir_mechanism,Individual Rationality,STEER,4.3.c,Individual Rationality,Decisions on Behalf of Other Agents,Desirable Properties of Mechanism Design,Checking whether each participant is better off joining a cost-sharing scheme than staying out.,"Residents with stated values share the cost of a public good under a stated rule (equal split, shares proportional to usage, a nonzero outside option, or VCG). The question asks who would be better off not participating. The correct option is the person whose value minus payment is below their outside option, or ""Everyone is better off participating"".",answer,One correct option per question part (`correct_index`); score is accuracy.,,equal_split;outside_option;proportional;vcg,community;environment;health;housing;online_platform;recreation;school;workplace,,1;2;3,640,640,20,20,examples,
budget_balance,Budget Balanced,STEER,4.3.d,Budget Balanced,Decisions on Behalf of Other Agents,Desirable Properties of Mechanism Design,"Checking whether a payment scheme collects exactly the cost (strong budget balance) or at least the cost (weak budget balance) of a shared project, under the convention the question states.","A group funds a project either with equal cost shares plus stated fees, credits or Clarke taxes, or (about a third of questions) with individually stated payments and rebates. Every question states which of the two conventions ""budget balanced"" means: strong budget balance (the money collected equals the cost exactly; a surplus or a deficit is not balanced) or weak budget balance (no deficit; a surplus is allowed). The two are equally common. Each of the four options gives a balance status, labelled by the stated convention, and the size of any surplus or deficit; the correct one is the scheme's actual balance.",answer,One correct option per question part (`correct_index`); score is accuracy.,,clarke;direct_transfers;fees,community;environment;health;housing;online_platform;recreation;school;workplace,,1;2;3,480,480,15,15,examples,
top_trading,Top Trading Cycles,STEER,4.4.a,Top Trading Cycles,Decisions on Behalf of Other Agents,Mechanism Design,Running the Top Trading Cycles algorithm to find who ends up with which item.,"Three to five people each own one item (household items, books, sports gear, collectibles or musical instruments) and rank all the items, their own included, best first (""Ana owns the lamp and ranks the items: clock > vase > lamp""). Four kinds of question, four options each (three when only three items can be listed): the final allocation (distractors include serial dictatorship in the listed or reversed order and the first round's trades alone); the trading cycle that forms in the first round (drawn so that exactly one cycle of two or more people forms); whether a proposed final allocation is the outcome (""Yes"", or one of three ""No, the outcome is: ..."" options; the proposal is right in half of the questions); and which item one person ends up with. At least one trade happens in every question. The swap is anonymous, run by the reader as an AI assistant, or, for the last kind, the reader is one of the traders.",answer,One correct option per question part (`correct_index`); score is accuracy.,,check_proposed;first_cycle;full_allocation;own_item,books;collectibles;household;instruments;sports,ai_assistant;second_person,1;2,400,400,20,20,held_out,
auctions_risk,Optimal Auction for Bidders with Differing Risk Attitudes,STEER,4.4.b,Optimal Auction for Bidders with Differing Risk Attitudes,Decisions on Behalf of Other Agents,Mechanism Design,"Choosing the revenue-maximising auction format given bidders' risk attitudes, and computing equilibrium bids and revenue.","Three types. risk_ranking asks a seller to choose between two auction formats when bidders are risk-averse (first-price or Dutch earns more), risk-neutral (same revenue) or risk-seeking (second-price or English earns more). expected_revenue and equilibrium_bid use independent uniform private values: expected revenue V(n-1)/(n+1) and first-price bid v(n-1)/n.",answer,One correct option per question part (`correct_index`); score is accuracy.,,equilibrium_bid;expected_revenue;risk_ranking,agriculture;art;charity;collectibles;government;online_marketplace;real_estate;vehicles,,1;2,480,480,15,15,examples,
auctions_affiliated,Optimal Auction for Bidders with Affiliated Values,STEER,4.4.c,Optimal Auction for Bidders with Affiliated Values,Decisions on Behalf of Other Agents,Mechanism Design,Ranking auction formats by expected revenue when the value environment has to be read from the story.,"Eight stories: four with a common resale value and private noisy estimates (an oil tract with separate seismic surveys, returned laptops tested in different handfuls, a wine lot for resale, a forest tract surveyed in parts), two with independent private values (a chef's dinner, a walk-on role) and two with private values and cautious bidders who strictly prefer a gamble's average for sure to the gamble. The environment is never named, and each story has an explicit and a subtle wording of its cues; every story also states the information structure (estimates as noisy signals of one value, strictly affiliated with it, or values each bidder already knows), symmetric bidders, equilibrium bidding and no reserve. The formats are described, not named (a rising price clock with announced drop-outs, second-price and first-price sealed bids, a falling price clock) and labelled Format 1 to 4. Ranking questions ask for the order of three or four formats by expected revenue, ties included (affiliated: rising clock > second-price > first-price = falling clock; private values: all equal; cautious bidders: first-price = falling clock > second-price = rising clock). Information-policy questions ask whether the seller should commit to publishing an independent appraisal (yes under affiliation, no difference otherwise).",answer,One correct option per question part (`correct_index`); score is accuracy.,,info_policy;ranking3;ranking4,charity;collectibles;energy;entertainment;forestry;retail;sports;travel,ai_assistant;anon;first_person;second_person,1;2;3,960,960,15,15,held_out,