Reciprocity needs you to meet the same person twice. Most of the people you deal with you will never see again, and they get helped anyway. The missing mechanism is that somebody is watching. I help you, it is noticed, and later somebody else helps me.
Everyone carries a score. Helping raises it, refusing lowers it. When you are asked for help you can see the other person's score, most of the time, and you decide on that. You will never meet them again.
A strategy here is one number: the score you insist on before you will help. Set it below the bottom and you help everybody. Set it above the top and you help nobody. In between you are choosing, and the choosing is what turns out to matter.
Alexander (1987) argued that this is where moral systems come from. Nowak and Sigmund (1998) made the claim precise enough to test, and found that it holds only above an exact line.
Alexander's claim was not modest. He argued that moral systems are indirect reciprocity: that what we call conscience is a mechanism for maintaining a reputation worth having.
Direct reciprocity, the kind Second Guess models, only works between people who meet repeatedly. That covers a village and not a city. Indirect reciprocity lifts the restriction: the person I help need never help me, provided somebody notices and the noticing spreads. It is the only well-understood mechanism that scales cooperation past the people you know.
Helping costs you something and is worth more to them. Nowak and Sigmund's result is that helping evolves only when the chance of your reputation being known exceeds that cost-to-benefit ratio. Below the line, building a reputation does not pay for itself and nothing gets started.
The switch above is sharp, and it sits a little higher than the algebra says: with a ratio of a quarter the population needs word to travel closer to two fifths of the time. That gap is not noise. Refusing to help someone with a bad score lowers your own score too, so a discriminator pays for discriminating, and the clean condition does not account for it. That is the same moral hazard discussed below, showing up as a number.
What survives is not generosity. It is generosity that checks first.
Watch which strategies carry the cooperation and they are the discriminators, not the unconditional helpers. That is worth sitting with if the model is being offered as an account of morality: the thing that evolves is conditional, calculating, and indifferent to whether the person in front of it deserves anything in any sense other than the score.
There is a second problem the model makes visible only if you look for it. Refusing to help someone with a bad score lowers your own, so a discriminator is punished for discriminating. Brandt and Sigmund called this the moral hazard in image scoring, and it is why later work replaced bare scores with standing rules that distinguish justified refusals from unjustified ones. Those rules need considerably more cognitive machinery, which weakens the claim that this is where morality came from.
Scores here are accurate, public, and free. Nobody lies about anyone, nobody has to spend anything to find out a score, and there is no gossip in the ordinary sense — no rumour, no motive, no reason to shade the truth about a rival. Every one of those is where the interesting difficulty in real reputation lives.