Table 1. Glossary of model terms and model description.

Model measure	Technical definition and abbreviation	Key roles
Baseline level of Harmful Intent attribution	Mode of prior probability distribution of harm intent, pHI0. It is the starting value of the mode pHIt.	Greater pHI0 leads to greater attributions of intent to harm initially, but how persistent this is depending both on evidence encountered (Dictator decisions seen) and, crucially, the uncertainty (inverse-strength) with which this belief is held.
Baseline uncertainty of Harmful Intent attribution	Spread of prior probability distribution of harm intent, uHI0. It is the starting value of the uncertainty uHIt.	Greater uHI0 denotes reduced confidence about attributions of harmful intent, and more willingness to believe that a Dictator who acts less generously than expected has higher intent to harm. The balance of uncertainty about Harm intent, uHIt, and uncertainty about Selfish intent, uSIt, determines the balance of which attribute is updated more on the basis of the dictator’s observed behaviour. Greater uncertainty uHIt directly contributes to greater variability in harmful intent attributions.
Baseline level of Self-Interest attribution	Mode of prior probability distribution of selfish intent, pSI0; the starting value of the mode pSIt.	Exactly analogous to pHI0 above.
Baseline uncertainty of Self-Interest attribution	Spread of prior probability distribution of selfish intent, uSI0; starting value of the uncertainty uSIt.	Exactly analogous to uHI0 above.
Partner policy uncertainty	Uncertainty parameter uΠ through which partner attributes are believed to lead to observed behaviours.	Unlike other uncertainties, this is not a spread of the distribution of both HI and SI attributions. The higher this uncertainty value, uΠ, the less informative each observed return by the dictator is. Low uΠ means that one can be certain that the actions of the dictator were not ‘by chance’, but due to their true attributes.
Learning rate, a.k.a. belief-update parameter, from one dictator to the next	Weight η by which the prior belief distribution over partner attributes shifts toward the distribution posterior to observing Dictator behaviour	A higher η leads the starting assumptions of dictators after the first one seen to be influenced by prior dictator behaviour seen so far. It can be thought of as a strength of belief that the Dictators seen during the experiment will resemble each other.
Model fit	Log-posterior probability lp that the fitted parameters gave rise to the data for this participant	A high lp means that if given the fitted parameters, the model would closely reproduce the attributions of the participant. Note that a bad fit might be because the participant is behaving erratically (e.g. because of a high uΠ) or that their pattern of behaviour is consistent in its own terms, but not captured by the model (e.g. a ‘magical thinking’ participant that alternates between two values in consecutive trials).