Table 1.

Event structure of the task

Short-term valence	Long-term valence	Informativity	Congruency	Event distribution		Payout
				Good urn	Bad urn	Variant 1	Variant 2	Variant 3	Variant 4
Negative	Negative	Informative	Congruent	10%	25%	−50	−50	−30	−30
Negative	Positive	Informative	Incongruent	25%	10%	−40	−40	−20	−20
Negative	None	Non-informative	None	10%	10%	−30	−60	−10	−40
None	None	Non-informative	None	10%	10%	0	0	0	0
Positive	None	Non-informative	None	10%	10%	+30	+60	+10	+40
Positive	Negative	Informative	Incongruent	10%	25%	+40	+40	+20	+20
Positive	Positive	Informative	Congruent	25%	10%	+50	+50	+30	+30

Payouts in the task varied over blocks, and can be categorised via the following criteria: informative events (column 3) had unequal probabilities to occur in good or bad lotteries (urns). Thus, they can be used to derive, or update (column 2), the conditional probability of a lottery to be good or bad in the long-term. This update was de-correlated (Fig. 2e) from the valence (positive or negative) of an event (column 1) and those events where the direction of long- and short-term valence align, are termed congruent (column 4), and incongruent where they mismatch. The long-term valence results from congruent events always having larger absolute payouts, which, however, in relation to the actual event is small (difference ± 10 points, maximal payout in the task 60 points per outcome). In addition, non-informative valence events did have the highest absolute payout in half of the blocks. The likelihood-ratio of informative events in good and bad lotteries (2.5 or 0.4) was constant throughout the experiment. In total 30% of possible outcomes did not carry information about long-term valence (rows three to five), and 10% additionally had no short-term outcome (payout of 0 cents, row 4), and these thus were non-informative. Note that in some blocks both urns could be good or bad, respectively (with the same pay-outs)