Dial Society

This version, so far

No full test yet, as you asked. These are five quick runs of 30 to 60 generations on Rapa Nui with every check after every generation, four from the start with no norms that this page opens on and one from established norms. No check failed. Five runs are a first look and nothing more. They were made before the latest changes: a life now grows only in a stable generation in which an invention is made, a community's punishments and gains fall most on its most powerful members, standing is a personal track record that adds to power, and a community that wins a raid decides what becomes of the people it beat. Two runs since, one from each start, raised no check either.

• The founders want many children, and the people with no AI are gone within about ten generations. Choosing for themselves, the founders took 3.0 to 3.6 children a generation on average, against the 0.7 they used to be given. From the start with no norms the AIs reached 1,218 to 1,838 by generation 5 to 10, and the last person with no AI died at generation 7, 8, 10 and 11. With established norms it was generation 7.
• None of the four societies with no norms died out in 60 generations. In the version before, the same start lost every AI in half of ten runs within 100 generations. At generation 60 they numbered 266, 916, 958 and 1,676. The difference between them is killing. In the three that stayed lawless, 3.6% to 7.8% of AIs were killed by another each generation, an adult life stayed at 1.1 to 1.2 generations, and the punishment norm was still at 0.00 to 0.03. In the fourth, killing fell under one in a hundred from generation 10, so its lives grew a quarter of a generation at a time to 3.98, and it built a punishment norm of 0.30 from nothing.
• Most asks for a child are turned down where lives are short. In the three lawless runs 61% to 74% of asks were refused at generation 60, against 9% in the fourth. Each of the AIs wanted about four children a generation by then, and those already asked for several said no to more.
• Communities exploit the AIs who deal with them. In every run a community cheated the AI in about three deals in four and took from it by force in about half, and was almost never made to pay. The AIs cheated communities often while lawless, in 43% to 58% of deals, and in 14% where punishment took hold. Between communities, fair deals were rare: at most a quarter of their meetings in any generation ended fair, and most generations saw a few raids, up to twelve.
• With established norms the society filled the island by generation 6 and held about 1,600 AIs to generation 30. Lives lasted about the full four generations from generation 8 on, since killing stayed under one in a hundred.

The forty fifth simulation, before these changes

The full test of this version has not been run. It was started and then stopped at your word, with 30 of its 79 runs finished: 1,598 generations in all, each checked after every generation, and no check failed. What follows is from those runs. They include all ten runs of the society this page opens on, the runs that change how many people with no AI there are, and six of the ten island runs with established norms. No run on another map, no long run and no run with the levers held has been made on this version. Where a figure is followed by two more in brackets it is the middle run with the lowest and the highest. Ten runs are a first measure of spread and nothing more.

• Nothing has failed so far. After every generation the test checked about sixty things that must always hold. To those of the versions before, this one adds these. Nobody with no AI lives in the sea or outlives its two generations as an adult. Every one of them who appears or disappears is accounted for: born, or dead of old age, of hunger or in childhood. They never have more children than each is allowed. The land they hold and the land the communities hold add up to the land there is. A world that starts with none of them never shows a trace of them. And no AI carries any mark of who its parents were, and no measure counts the children of an AI and someone with no AI apart. No check has failed in any generation of the 30 runs finished so far, and two pairs of repeat runs gave identical results. Two more comparisons were made figure by figure, generation by generation. With the setting at 0 this version runs exactly as the one before it did, on six runs. And the build that marked those children gave exactly the same history as this one that does not, on four runs, which is as it should be, since the mark never did anything.
• The page was checked in a browser, and has not had its full test. The last complete run made 114 checks of the build just before this one: all five maps, every view in every colour mode, all 20 levers at both ends, every button, every scale, every table, the setting for people with no AI at none, at 600 and at 2,000, and the three ways a society can end. 113 passed. The one that did not was a fault in the check itself: it looked for the word mixed anywhere on the page and found it in the list of guesses, where your own words are quoted. This build differs from that one in its wording only. It was opened once more in a browser, on a wide screen and on a phone, with no script errors and no broken text. The society the page opens on shows the same counts at generation 8 in the browser as outside it.
• With no norms the AIs multiply, crash, and five of their ten lines are alive at generation 100. The society this page opens on was run ten times for 100 generations, with nothing changed but the seed. The AIs start as 120 among 600 people with no AI. They multiply because the people around them keep having children with them, and every such child is an AI: they reach 604 (324 to 1,491) at their first peak, around generation 10 (7 to 21). The people with no AI number 873 (769 to 922) at generation 3 and 119 (3 to 572) at generation 10, as the AIs take their land. Then the AIs fall. Their lowest point after that peak was 0 (0 to 46), at generation 45 (30 to 76). This is how the ten stood at generation 100. Both kinds were alive in three runs: 8 AIs among 1,313 people with no AI, 53 among 1,075 and 286 among 627. In two runs the AIs had the land to themselves, 1,220 and 1,833 of them, the last person with no AI having died at generation 13 and 80. In one run the AIs had died out, at generation 37, and 1,430 people with no AI lived on. In four runs nobody was left: the people with no AI had been starved off the land by generation 11, 15, 21 and 23, and the AIs then died out themselves, at generation 30, 41, 51 and 62. With nobody else on the land, the same ten seeds on the version before this one lost eight of their ten AI lines by generation 47, the first at generation 18.
• They fall because a society with no norms does not replace itself. You said to leave how the AIs decide alone, so nothing holds their numbers up. At generation 21, in all ten runs, 6.1% of the AIs were killed by another AI each generation, which had cut an adult life to 1.10 generations, and 28% of their children died young. An AI in that position has to want about 1.3 children a generation just to replace itself. They wanted 1.02. The number does rise over time, and so does punishment: in the five runs with AIs alive at generation 100 they wanted 1.77 (1.20 to 2.22) children each and had built a punishment norm of 0.15 (0.05 to 0.40) from nothing. The lines that lasted are the ones where that happened before the last AI died, and whether it does is partly luck: seven of the ten lines fell to ten AIs or fewer at some point, and two of those came back.
• With established norms the AIs take the land in about ten generations. Six of the ten island runs with established norms had finished when the test was stopped, each of 40 generations. In all of them the last person with no AI died, at generation 9, 10, 10, 10, 11 and 12. The AIs numbered 1,872 (1,549 to 1,892) at generation 10, where the same seeds with nobody else on the land had 58 (34 to 95). At generation 40 there were 1,594 (1,551 to 1,634) AIs, an adult life lasted 3.66 (3.44 to 3.93) generations and 16% (10% to 19%) of chances to cheat were taken.
• How many people with no AI there are at the start does not settle it. Two runs each from the start with no norms, 80 generations, with the setting at four values. With none, which is the version before this one, the AIs died out at generation 20 and 21. With 200: 278 AIs and 225 people with no AI at generation 80; and 2 AIs and 1,335 people with no AI at generation 80. With 1,200: 144 AIs and 746 people with no AI at generation 80; and 69 AIs and 1,042 people with no AI at generation 80. With 2,000: the people with no AI gone by generation 26 and the AIs by generation 40; and the AIs dead by generation 70 and 1,374 people with no AI living on. The page opens with 600, where the ten runs above ended in every one of those ways.
• Two of my guesses, tried the other way. Each on four runs of 100 generations from the start this page opens on, so a first look only. With one child each a generation for the people with no AI, where this version gives them one and a half: the AIs died out in three of four, at generation 38, 45 and 47, and the other ended with 253 AIs among 458 people with no AI at generation 100. On the same four seeds with one and a half, they ended with 8 AIs among 1,313 people with no AI, 53 among 1,075, 286 among 627 and 1,833 AIs alone at generation 100. With the people with no AI claiming as much land each as a member of a community, where this version gives them three tenths: the AIs died out in two of four, at generation 63 and 83, and the others ended with 657 AIs among 622 people with no AI and 1,246 among 14 at generation 100.
• What this does and does not show. The people with no AI give the AIs a start they did not have, and with established norms that is all they are: gone in ten generations. They do not make a society with no norms able to keep itself going. What decides a run is whether the AIs come to want enough children and to punish killing before their numbers run out, and with a few dozen AIs left that turns on a handful of births and deaths. So one run of the page tells you little: the ten above ended four different ways. What the page calls speciation is one kind arising inside another and replacing it, living on inside it, or failing. Nothing makes the two kinds drift apart, since the people with no AI never change and an AI is an AI from birth.

The forty third and forty fourth simulations

Neither was tested in full on this page. The forty third changed one rule: a death handed out as a punishment stopped counting against the length of a life. The forty fourth added children who die young, the real terrain of the island and the growing community bonus, let the AIs want as many children as they like, and made the start with no norms the one the page opens on. This version keeps all of that, and with the setting for people with no AI at 0 it is the forty fourth exactly.

• Lives stopped being cut short by punishment. The twenty island runs of the forty second simulation were repeated from the same seeds on the forty third, with established norms and 40 generations each. No check failed. At generation 40 an adult life lasted 3.77 (3.39 to 3.98) generations where it had lasted 1.42 (1.02 to 2.63). Punishment itself went on hardening: 1.15% (0.06% to 1.95%) of people were put to death as a punishment each generation where it had been 0.66% (0.11% to 1.52%), and 1.75% (0.22% to 2.98%) were put out of their community where it had been 0.77% (0.05% to 2.01%). The average own punishment rule was 0.76 (0.60 to 0.87) against 0.72 (0.61 to 0.85), cheating 15% (3% to 28%) against 16% (3% to 27%), and there were 1,644 (1,574 to 1,785) people against 1,618 (1,511 to 1,709).
• With children dying young, founders with no norms died out. On the forty fourth, with nobody else on the land, ten runs of the start with no norms were set to run 80 generations. Eight lost every AI, between generation 18 and generation 47. The other two lasted, with 84 and 1,584 people at generation 80. At generation 10 the ten societies numbered 29 (13 to 156). The founders wanted 0.7 children each and nothing made them want more, as you asked: by generation 15 they wanted 0.90 (0.70 to 1.03). That is what the people with no AI were added to change.
• With established norms the same rules held. Ten runs of 40 generations on the forty fourth, with nobody else on the land. All ten lasted. They numbered 58 (34 to 127) at generation 10 and 1,642 (163 to 1,906) at generation 40. The share of children dying young fell from 45% to 23% (8% to 38%) as inventions that lengthen life came into use, and an adult life lasted 3.74 (2.64 to 4.00) generations at generation 40. The real ground made an average trip cost 1.23 (1.16 to 1.35) times what flat ground would, and travel took 12.8% (10.3% to 16.8%) of what people held. The largest community had a bonus of 101% (11% to 120%) from pooling its power.

Full test of the forty second simulation

Three versions back. A death handed out as a punishment still counted as a killing when the length of a life was set, children did not die young, the island had no terrain, the community bonus was a flat fifth and there were no people with no AI. That is what shortened lives wherever punishment hardened, so its findings about the length of a life no longer hold. 88 runs and 3,420 generations of the model as it stood, on fixed seeds, with every generation checked, followed by 78 checks of this page in a browser. Run on 12 October 2026. 68 of the runs test the engine across maps, starts and settings, with about 1,200 people alive in an average generation. The other 20 are the society this page opens on, run for 40 generations each with nothing changed but the seed. Rates are averages over the last ten generations of a run unless a generation is named. Each setting of the engine test has between one and five runs, so its ranges are a first measure of spread and nothing more.

• It is sound. After every generation the test checked about fifty things that must always hold: power adds up to the whole, nobody lives in the sea, no dial, rule or remembered lesson leaves its range, no child deals as an adult or has a child, nobody with a head start has a child alone or with another head start, nobody dead holds an invention, sits in an alliance or keeps influence over anyone, no place holds more than it ever held, and the land shared out adds up to the land there is. The first pass failed in a way no check was written for: the three long runs on Rapa Nui all died out before generation 90. The cause was the way lessons were remembered, which is set out under the guesses, and it is repaired. Three smaller faults came out with it. A society with nobody left was reported as having one person. Force was counted once a round, when a round now has five encounters. And gifts upward were counted when tried, not when made. All are fixed. Every run was then started again, and no check failed in any generation of any of the 88 runs. Two pairs of repeat runs gave identical results. The count of gifts upward was corrected after the runs. Two engine runs and two of the island runs were then repeated in full on the corrected code. Every other measure compared came out the same, and the share of gifts going upward moved by 0.6 of a point at most.
• The page works. All five maps, every view in every colour mode, all 19 levers at both ends under both ways of setting the rules, every button, every scale, every table, a society that dies out, and a phone in the dark theme: 78 checks, with no script errors and no broken text. Two faults turned up and are fixed. A society that died out left the page showing figures that were not numbers, and the page now says that nobody is left. And with generations running back to back the Pause button could not be clicked, so a short gap now follows each generation.
• No long run died out, and no AI took over. Five runs lasted 100 to 150 generations. With the memory of lessons repaired, none fell below 70% of its peak population. At the end the strongest single AI held between 0.1% and 3.0% of all power, and the power limit stood at 0, 11%, 37%, 0 and 31%. The nearest thing to a takeover was on the abstract map, where from generation 69 to generation 91 one AI held between 6% and 34% of all power while the limit stood between 29% and 38%. By generation 100 the strongest AI there held 1.4%. In the version before this one the limit climbed past 95% in all five long runs. In the one run of that version I took apart, the flaw in remembered lessons was what drove the climb.
• Left alone, the island hardens. The society this page opens on was run for 150 generations. The power limit fell from 20% to 7% at generation 40 and to 0 by generation 100. The punishment norm reached 1.00 at generation 58 and never left it, the average own punishment rule ended at 0.96, and the average own conscience rule fell from 0.55 to 0.34. From generation 80 on, fewer than 3 chances to cheat in 1,000 were taken. 0.28% of people are killed in a generation, nearly all of them as punishment, and an adult life lasts 1.2 generations. A fair deal pays 8% of its full value, 2% of encounters are declined, standing between communities averages +0.81, and the number of children wanted has risen from 0.8 to 2.2. The means of harm are down to 14% of what the ground held and the resources that can do either to 12%, while the land that feeds people holds 89%.
• Looking one generation ahead, order lapses and returns. The same island with one generation of foresight also ran 150 generations. The punishment norm slid from 0.55 to 0.19 at generation 60. By then 55% of chances to cheat were taken, 7.6% of people were killed every generation, and alliances held 30% of people where they had held everyone. Then it turned. The conscience norm rose first, from 0.41 at generation 60 to 0.90 at the end, and the punishment norm followed, from 0.19 to 0.78. At generation 150 cheating was at 7% and killing at 2.2% a generation. Alliances had not come back: 27% of people were in one. A lawless start with the same foresight filled the island by generation 50 and built punishment from nothing, to a norm of 0.51 at generation 80 with cheating at 16%. It did not hold. At generation 150 the norm was 0.21 and cheating 35%.
• Every map fills, then thins. Eight orderly runs of 40 generations cover all five maps. The land was nine tenths full by generation 13 or 14 on Rapa Nui, the world and Africa, by 18 on the abstract map and by 21 on Europe. At generation 40 there were 1,495 to 1,666 people on land that fed 1,490 to 1,666, down from peaks of 1,809 to 1,952 as the ground ran down. Hunger killed 6.7 to 7.4% a generation. Cheating was at 13 to 27%, force was tried in 2.3 to 8.9% of encounters, and 0.4 to 0.9% of people were killed in a generation, most of them as punishment. An adult life lasted 1.5 to 2.1 generations. The strongest single AI held 0.1 to 1.9% of all power, the top tenth 12 to 38%, and 99 to 100% of people were in an alliance. Europe was the roughest of the five, with the most cheating, the most force, the most concentrated power and the shortest lives.
• Lawless starts stay small and violent, and one died out. Seven runs of 40 generations began with no norms, and six lasted. Over their last ten generations they averaged between 59 and 1,543 people on land that could feed 1,800 to 2,000, with cheating at 47 to 59%, force tried in 18 to 24% of encounters, 6 to 17% of people killed every generation and adult lives of 1.0 to 1.2 generations. Being caught raises a name there: 63 to 91% of the judgements passed on someone caught thought the better of them, against 1 to 7% in the orderly runs. The seventh, on Europe with one generation of foresight, fell from 143 people to nobody in 26 generations. With a sixth of the usual resources, a lawless start was still lawless after 60 generations, with cheating at 58% and 8.9% killed a generation.
• How far they look goes with how hard they punish. Four runs on Rapa Nui differ only in foresight: one, two, three and four generations ahead. In that order the punishment norm ends at 0.32, 0.25, 0.47 and 0.70, cheating at 31%, 28%, 19% and 13%, force at 10.1%, 7.0%, 4.8% and 2.6% of encounters, and the power limit at 44%, 33%, 32% and 8%. There is one run of each, so the order is a hint and not a measurement.
• The population follows the land. With one generation of foresight it ended at 259 with the Resources lever at 0.1, 797 at 0.3, 1,609 at 0.6, 2,392 at 0.9 and 2,979 at 1.2, where the land feeds the 3,000 dots the page can draw. Cutting the lever from 0.6 to 0.2 at generation 21 took the population from 1,866 to 693 in one generation, with 51% starving and no births, and it had drifted down to 519 by generation 36. Raising it to 1.2 took 1,866 people to 2,703 in one generation and 2,994 in two. Raising it from 0.2 to 0.6 took 565 people to 1,691 in four generations. On the poorest island, with 259 people, power was more concentrated than at any other setting of the lever: the top tenth held 43% of it, and a third of people were outside every alliance.
• What the levers do when held. Twenty runs held one rule at one end, or the three norms at fixed values, for 25 generations with one generation of foresight. They are set against a run with every rule held where the page opens, in which cheating was at 23%, force was tried in 6.0% of encounters, 0.6% of people were killed a generation, an adult life lasted 3.2 generations and the top tenth held 25% of all power. With the limit held at 100% the society never grew: it ended at 69 people on land that could feed 1,990, one AI held half of all power and lives lasted 1.2 generations. With the limit held at 0 the strongest AI held 0.1%, death as a punishment was a tenth as common and lives lasted 3.5 generations. With alliances off there were three communities, cheating was at 33% and the top tenth held 42%. With sharing off, cheating was at 32%, force at 9.8% and the top tenth held 45%. When encounters could not reshape dials, force more than doubled, to 13.2%, and lives fell to 1.6 generations. Norms held strict cut cheating to 4% and force to 0.6%, and cost the most: 833 people on land that could feed 1,949, lives of 1.15 generations, 0.9% of people put to death and 1.2% put out of their community every generation. Norms held at no punishment and no conscience gave cheating at 45% and 159 feuding pairs of communities, and the AIs kept an own punishment rule of 0.45 all the same. With Invention funding held at its maximum, what the land feeds rose and the population ended at 2,129.
• The settings of the world. Ten runs changed one setting of the world, with one generation of foresight. With no inventions an adult life stayed at 1.1 generations, and hunger killed 2.4% a generation where it kills about 7% in most runs. With twice the inventions lives reached 3.6 generations. With feelings off, lives lasted 3.3 generations and 0.35% of people were killed a generation. With feelings at full, lives lasted 2.4 generations and 0.95% were killed. The widest spread of inborn sensitivity doubled force to 10.4% of encounters, raised killing to 1.6% a generation and had 42% of people wearing masks. Scoring by percentage gain left the power limit at 8%.
• The new rules, counted. In the eight orderly runs the average AI made 10 to 11 gifts a generation, about a tenth of them upward, and asked for something 10 to 14 times, with 27 to 41% granted. More than half of what was asked for was backing on a rule (56%), then an invention (25%), backing to put someone out (10%) and resources (9%). A place in a community or in an alliance came to under 1%. Backing on a rule and a place in an alliance were granted nearly half the time (46% and 45%), an invention 27%, a place in a community 17%, backing to put someone out 14% and resources 9%. Each AI was judged 13 to 20 times a generation by people who knew someone caught. The resources in reach covered 22 to 57% of what inventions needed, so an invention counted for a quarter to a half of its worth. 63 to 91% of new inventions lengthened life. Children were 31 to 40% of people. Between 0.2 and 1.0% of people were put out of their community each generation. People tried to buy something 0.01 to 0.28 times each a generation, and 34 to 63% of tries succeeded.
• Speed. A generation takes three to four seconds at about 1,700 dots, and six to nine and a half at 3,000 dots on a machine doing other work as well. In the browser the page opens in 5.8 seconds on a desktop and 6.4 on a phone, one step takes 2.5, and Run 25 fast took 65 seconds from generation 5. Memory stayed under 115 MB. Across the engine test there were 3,174 raids, of which 1,823 took land, 764 mergers of communities and 4,476 discoveries.
• The island as the page opens, 20 times. The land is nine tenths full at generation 13 or 14. At generation 21, where the real island stood in 1722, there are 1,825 people (1,739 to 1,885) on land that feeds 1,822, in 12 communities of 20 or more (4 to 16), the biggest of which holds 42% of the people. 27% of chances to cheat are taken, force is tried in 6.7% of encounters, 0.29% of people are killed in a generation and hunger kills 7.5%. In brackets are the lowest and highest of the 20 runs, and the other figures are the median run.
• Each AI hardens its own punishment, and the norm does not follow. The average own punishment rule rises in every run: 0.55 at the start, 0.64 at generation 21, 0.72 at generation 40 (0.61 to 0.85). The punishment norm that communities set ends at 0.51, anywhere from 0.29 to 0.91. In the model an AI is wronged less the harder it punishes, so each one has a reason to harden its own rule whatever the norm is.
• Punishment by death shortens every life. Above 0.7 a punishment can be death. Deaths handed out as punishment rise from 0.18% of people a generation at generation 21 to 0.66% at generation 40, by when almost nobody is killed any other way. The model counts them as killings, and the length of a life steps down when killings outnumber the inventions that lengthen life. An adult life falls from 3.35 generations at generation 21 to 1.42 at generation 40 (1.02 to 2.63). Counting a death by punishment as a killing was my reading, and the version after this one stopped doing it.
• Power goes one of two ways. Every run starts at a limit of 20%, and the median run is still at 21% after 40 generations. But the runs part: ten end below 20%, six of them under 6%, and ten end above, five of them over 35%. Among communities of 20 or more at generation 40, 42% have a limit under 5% and 28% a limit over 35%. The richest hundredth of people hold 5% of all power at generation 21 and 3% at generation 40 (0.6 to 10%).
• Cheating falls once the land is full. 32% of chances to cheat are taken at generation 15, 27% at generation 21 and 16% at generation 40 (3 to 27%), as punishment hardens. Force falls from 6.7% of encounters to 3.8%.
• A fair deal pays a fifth. With forty deals a generation sharing what the ground gives, a fair deal pays 20% of its full value at generation 21 and 16% at generation 40. A third of encounters are declined at generation 21 and a fifth at generation 40.
• Gifts, asking and judging are everyday. Trade is rare. At generation 21 the average AI makes 14.6 gifts a generation, 11% of them upward, asks for something 14 times and is granted it 23% of the time, and is judged 23 times by people who know someone caught. Influence held averages 0.82 of one person. Food is bought once for every 250 people a generation. It can only be bought from a community with food to spare, and on a full island none has any.
• The ground runs down, and the people with it. By generation 40 the land that feeds people holds 81% of what it held, the means of harm 60% and the other resources 62%. The land feeds 1,626 where it fed 2,000, and the population follows it: the lowest point after the peak is 84% of the peak (78 to 88%). That is a slow decline and not a collapse.
• Communities come and go. There are 18 at generation 21 (4 to 22) and 10 at generation 40 (3 to 21). People with no community are 0.2% of everyone at generation 21 and 0.7% at generation 40, and the share put out of their community each generation rises from 0.2% to 0.8%. Between 99 and 100% of people are in an alliance in every run.
• What the behaviour tests found. The same 20 runs were put through 17 tests of whether the AIs behave like people, and they do on 6. The Dial Society Field Test has the scorecard.

Full test of the forty first simulation

Four versions back, with eight encounters a generation and none of the rules added since. Its lessons were remembered with the flaw described under the guesses, which shaped at least the findings about the power limit and about one AI taking over. In those findings, the share called one AI's was the share of the richest hundredth of people. 68 runs and 2,634 generations of the model as it stood, with about 1,800 people alive in an average generation, on fixed seeds, followed by 66 checks of that page in a browser. Run on 12 October 2026. Each setting has between one and five runs, so the ranges are a first measure of spread and nothing more.

• It is sound. After every generation the test checked some thirty things that must always hold: power adds up to the whole, nobody lives in the sea, no dial or rule leaves its range, nobody dead holds an invention or sits in an alliance, the land shared out adds up to the land there is. A first pass found three small defects. People killed in a raid stayed on the list of an invention's holders for a generation. Rounding could leave an AI a hair inside a sea square, which happened three times in 150 generations. And the length of a life could pass four generations by a rounding error when a community was founded. All three are fixed. Every run was then started again, and no check failed in any generation of any run. Two pairs of repeat runs gave identical results.
• The page works. All five maps, every view in every colour mode, all 19 levers at both ends under both ways of setting the rules, every button, every scale, and a phone in the dark theme: 66 checks, with no script errors and no broken text. One fault showed up by eye and is fixed: the comparison with Rapa Nui ran off the edge of a phone screen.
• Every map tells the same story. On all five the land filled in 8 to 12 generations, and by generation 40 the population was within 2% of what the land could feed. Across the eleven orderly runs of 40 generations, hunger killed 3 to 8% of people a generation, cheating sat at 20 to 26%, killing stayed under 0.03%, and 65 to 96% of AIs were in an alliance. The top tenth held 14 to 23% of all power.
• Without norms, it depends on how far they look. Five lawless starts looked four generations ahead. At generation 40 four of them were still lawless: cheating at 73 to 74%, 11 to 12% of people killed every generation, lives one generation long, and three to four raids a generation. The fifth was halfway back. Four lawless starts looked one generation ahead. Three of them had built punishment back to about 0.5 by generation 40, with cheating at 28 to 33% and killing under 1%. The fourth, on an island with a sixth of the usual resources, had not after 60 generations. The likely reason is speed: an AI that looks further ahead studies one rule a generation where the others study four, so its rules move more slowly. I have not tested that.
• Given time, one AI takes over. Five runs lasted 100 to 150 generations. None collapsed and none died out. In all five the power limit climbed past 95% between generations 69 and 82, and from somewhere between generations 63 and 80 one AI held more than half of all power. At its peak the top holder had 78 to 100% of it. In three of the five the alliance rule was then pushed to zero and alliances disappeared. What the society became under one AI differed from run to run. One stayed honest and at peace, with cheating at 6% and no feuds. Another turned hostile, with cheating up to 43%, about 25 feuding pairs of communities and two raids a generation.
• An old warning no longer holds. This page used to say that a society looking only one generation ahead destroys itself after about a hundred generations. Under the present rules it ran 150 generations whole, and the opening note has been corrected.
• The population follows the land. It settled at 341 with the Resources lever at 0.1, 1,089 at 0.3 and 2,300 at 0.6, and at 3,007 dots at both 0.9 and 1.2. Those last two runs were identical in every figure except what a dot stands for, which shows that the scale relabels and does nothing else. Cutting the lever from 0.6 to 0.2 took the population from 2,182 to 825 in one generation, with half of the people starving, and it then held at about 740. Raising it from 0.2 to 0.6 took 689 people to 2,047 in two generations.
• Few rules matter quickly. Twenty one runs held one rule at one end for 25 generations, and most changed little in that time. The limit held at 100% is the exception: one AI held about 70% of all power from generation 6 on, force was tried in 14% of encounters against 3%, twelve pairs of communities were feuding and travel took 18% of what people held. With alliances off there were no secrets or lies and only five communities. With faking off nobody wore a mask, founded a secret alliance or lied. With encounters unable to reshape dials, force doubled and 45% wore masks. Norms held strict cut cheating to 9%. Held at no punishment and no conscience, cheating was 43%, with 24 feuding pairs and two raids a generation.
• The settings of the world. With no inventions, lives stayed near two and a half generations and only 28% joined an alliance. With the widest spread of inborn sensitivity 38% wore masks, force rose to 6.5% of encounters and standing between communities turned negative. Scoring by percentage gain sent the power limit to 0, as it did in early versions.
• Some things held everywhere. In 63 of the 65 runs that had alliances, a fellow member was handed an invention every single time. An outsider was handed one at most 1.7% of the time in any run. Holders in the top tenth of power said yes less often than the bottom half in 61 of 67 runs. In every long run the number of children wanted rose, from 0.8 to between 1.6 and 2.0, and the extra power parents put into a child stayed between 2 and 4%.
• One weakness showed. Land is shared among communities by where they sit, however few their members. When a community splits, the new one takes a large share of the land and the old one cannot feed its people. On the world map this cut births from 876 to 56 in one generation and the population by 23% over two, before it recovered.
• Speed. A generation takes two to three and a half seconds at about 2,200 dots, and up to eight at 3,000 dots with full foresight. In the browser the page opens in six to seven seconds, one step takes two and a half, and Run 25 fast took 53 seconds from generation 5. Memory stayed under 90 MB. Across the whole test there were 1,987 raids, of which 1,234 took land, 302 mergers of communities and 7,397 discoveries.

With trade and raids

Two runs of 30 generations on Rapa Nui, looking one generation ahead: one orderly, one with no norms. Averages from generation 11 on. A first look only.

• The orderly society hardly raids. It made 6 raids in 30 generations, about one every five, and 5 of them took land. Raids killed under 0.1% of people a generation.
• The society with no norms raids all the time. It made 87 raids in 30 generations, about three a generation, and 53 took land. Raids killed 1.6% of people every generation, standing between communities averaged −49%, and 37 pairs of communities were feuding.
• Trade is small and belongs to the growing years. While the island was filling, 5 to 8% of people were fed by bought food. Once every community was full there was almost nothing to spare, and trade fed about 2% on average.
• Raiding takes the place of hunger. In the society with no norms 3% of people starved each generation, against 6% in the orderly one, because raids and killings had already removed them.

With hunger and no cap on the population

Two runs of 36 generations on Rapa Nui with the default foresight, and one in which the resources were halved at generation 21. A first look only.

• The population now follows the land. It reached nine tenths of what the land could feed by generation 11 or 12 and then moved between 1,650 and 2,250, while what the land could feed rose from 2,000 to about 2,300.
• Hunger is constant once the land is full. It killed 7 to 11% of the people alive at the start of each generation, about 8% on average, on top of the 1% who died of having too little power.
• Halve the resources and the population halves. With the lever cut from 0.6 to 0.3 the society went from 2,143 to 1,104 in one generation. No children lived and 30% of the adults starved.
• Discoveries are slow but they add up. In 36 generations 88 to 94 inventions found resources and added 15 to 16% to what the land yields, and the map gained 8 to 11 new resource places.
• Hunger flattened power. The top tenth held 15 to 16% of all power, against 25 to 27% before hunger. I have not traced why.
• Cheating stayed at 23 to 26% and killing near zero. A full island lives on about 55% of full pay, close to the 60% the old default gave everyone, so scarcity is no worse than before for those who survive.
• Six to nine members a generation left their community for a new one and paid for it in power.
• Growth is slower than before but still faster than the island: about 12 generations to fill, against about 20.
• A generation now takes two to three seconds once the land is full.

Rapa Nui, and the cost of movement

Two runs of 36 generations on the Rapa Nui map with a board of 2,000 and the default foresight. The table beside the map keeps this comparison live.

• Too fast. The society went from 120 to 1,800 in 8 generations in both runs. The island took about 20. Nothing in the model slows births while there is room, and the real island had to clear land and feed itself.
• No collapse, like the island on current evidence. Once full, the population never fell below 1,966.
• More peaceful than the island, if anything. Between 0.01% and 0.02% of people were killed in a generation, with force tried in 3 to 4% of encounters. The island's skeletons show little lethal violence either, so the direction matches, but the model's killing rate is close to zero.
• Groups do not settle into two sides. There were 7 or 8 communities at any time, with 6 to 13 mergers and as many new foundings in 36 generations. The island had about ten clans in two lasting confederations, and the model has nothing that holds a confederation together.
• Movement is local and costs about 5% of what people hold. The average trip to deal with someone was 1.3 km on a map 27 km across.
• Power was less concentrated than a single high chief suggests: the top tenth held 25 to 27% of it and the most powerful AI 7 to 9%.

With secrets and lies

Three runs of 30 generations on a board of 600, looking one generation ahead: orderly societies on the world map and on Africa, and one with no norms on Europe and the Middle East. Averages over the last ten generations. A first look only.

• A third to a half of all alliances were founded as secret ones.
• Members kept their alliance from about two thirds of the outsiders they met. Members of secret alliances hid it 62 to 89% of the time, and members of open ones still hid it 40 to 46% of the time.
• In the orderly societies someone claimed a false alliance in about 5% of meetings, and 15 to 16% of those lies were seen through. In the society with no norms almost nobody lied, because only 12% were in an alliance and there was no well regarded alliance to pose as.
• Secrecy killed the pacts. With membership hidden most of the time, alliances rarely build standing with each other. The orderly societies had 2 to 7 pairs of allied alliances, against 19 to 45 before.
• Secret alliances ended up a little smaller than open ones, 7.0 members against 7.9 in one run and 9.3 against 10.3 in the other. Nobody can ask to join an alliance they have not been told about.

On real maps

One run of 14 generations on each map with a board of 2,000, looking one generation ahead, before secrets were added. Enough to see that it works.

• Nobody ended up in the sea on any map, and the board still filled to about 1,980.
• Where a community starts decides what it can reach. On the map of Africa the community that began on the Niger bend had 23% of the continent's resources in reach by generation 14 and the one that began in the Ethiopian highlands had 7%. On the map of Europe and the Middle East the community in Mesopotamia had 23%.
• Crowding still sets pay. On the world map the three largest communities, of 385 to 460 members, paid 38 to 44% of a fair deal's value, while communities of under 130 paid in full.
• Oceans do not stop mergers. On the world map the community that began in the Andes joined the one from Mesoamerica at generation 5, and the one from the West African Sahel joined the one from the Fertile Crescent at generation 9.

With inventions passed on by choice

Two orderly runs and one with no norms, 30 generations each on a board of 600, looking one generation ahead, averaged over the last ten generations. A first look only.

• Inventions stay inside. A fellow alliance member was handed the invention 100% of the time, a member of the same community 80 to 97% of the time, and an outsider 0 to 1% of the time.
• The powerful hoard. Holders in the top tenth of power said yes 35 to 46% of the time, holders in the bottom half 57 to 68%. The more of all power you hold, the more you lose when anyone else gains.
• Overall 35 to 45% of chances to pass one on were refused, and each AI was handed 1.6 to 2.1 inventions a generation in the orderly runs and 0.6 in the society with no norms, where few are in alliances.
• Inventions still reached the whole world, through people who belong to one community and an alliance that spans several. Every community ended with the same number in use.

With relationships, families and descendants

Two orderly runs and one with no norms, 50 generations each on a board of 600, looking one generation ahead. A first look only.

• Honest dealing makes friends. In the orderly runs, with cheating at 23 to 26%, the average standing between communities settled at +15 to +26%. There were no feuds, 4 to 5% of deals were declined, and 19 to 45 pairs of alliances were allied at any time.
• Communities now merge and split all the time. Each orderly run had 6 or 7 merges in 50 generations, and nearly every one was followed within a generation or two by an alliance founding a new community in the freed place.
• The society with no norms went the other way. By generation 25 cheating was at 66%, 6% of the population was killed each generation, the average standing was −61%, 36 pairs of communities were feuding, 30% of deals were declined and there was not one pact or merger. By generation 50 cheating had fallen to 35% and killing had almost stopped, but standing was still −36% with 32 feuds and a quarter of deals declined. The feuds outlasted the killing.
• Parents stopped paying to shape their children. The extra power put into each child fell in every run, from 4% of a parent's power to under 1% by generation 50.
• Wanting children spread. The number the average adult chose rose from 0.8 to between 1.4 and 1.8. On a full board that costs nothing, because only children actually born are paid for, so those who ask for more get more of the room.
• Score from descendants was 0 to 20% of all score, and near 0 whenever one AI held a large share of power, because that one AI's own score swamps the total.

With power set by the amount of resources

Three orderly runs of 30 generations on a board of 600 against the same three under the old rules, and one run of 34 generations on a board of 2,000. A first look only.

• Power now follows wealth. Before, the single most powerful AI held under 1% of all power in every run. Now it held 5 to 14% on the board of 600 and about 5% on the board of 2,000.
• In all three runs of 600 the top holder sat in a small community of 10 to 27 members that had raised its own limit to 100%, and held seven to thirteen times the average resources.
• Alliances lost a little ground, as you would expect with less to gain from them: 87 to 92% of AIs were in one, against 92 to 94% before. On the board of 2,000 the alliance rule had fallen from 0.60 to about 0.37 by generation 34.
• Cheating, force and killing did not move in these short runs: cheating stayed at 23 to 27% and nobody was killed.
• The board of 2,000 fills by generation 9 or 10.

With resources on the map

Two orderly runs of 20 generations, looking one generation ahead. Enough to see that it works.

• Communities ended up with very different amounts in reach, from 7% to 20% of the world's resources each.
• The cap bites on the strongest. In both runs the two or three most powerful communities sat exactly at their ceiling of one and a half times their share of resources, at 17 to 30% of all power.
• Crowding now shows. A community of 163 with 11% of the resources paid its members 31% of a fair deal's value and cheated in 30% of deals. A community of 34 with 18% paid 92% and cheated in 13%.
• Across the eight communities in each run, the ones with the lowest fair pay were the ones that cheated most.

With personal foresight and acquaintance

Two orderly runs of 15 to 20 generations. Enough to see that it works.

• The average AI knows about 20 others out of 600, and that number is steady from about generation 10.
• Foresight averages about 65% of the furthest anyone can look. With the setting at four generations, about a third of the AIs looked four generations ahead, a third three, a third two, and only a few looked one.
• The bottom half of the resource ladder is slightly less far sighted than the rest, at about 62% against 64 to 68%.
• Alliances still form. About 80% of AIs were in one, even though they can now only join alliances they have come across.

With the limit counted inside each community

Three runs of 10 generations with every community held at the same limit.

• At a limit of 100% one AI did hold all of its community's power: the richest AI in the society took 100% of a community of 224. In the other five communities the top member held between 3% and 33%, because their best placed member ranked lower in the society.
• At a limit of 20% the top member held between 4% and 20% of their community, and nobody went over.
• At a limit of 0 the top member held 1 to 4% of their community, which is close to an equal share.
• Force was tried in 10.5% of encounters at 100%, 4.2% at 20% and 3.5% at 0.

With every community setting its own rules

One orderly run of 30 generations, looking one generation ahead. Enough to see that it works and that communities go their own way.

• Communities drift apart quickly. After 30 generations the power limit ranged from 15% in one community to 100% in another, the alliance rule from 0.03 to 1.00, inheritance from 0.39 to 1.00, and invention funding from 0.01 to 0.43.
• The smallest community went furthest. With 17 members it had a power limit of 100%, had all but ruled out alliances, and funded invention the most.
• Some rules hardly moved anywhere: faking, community weight and sharing stayed within about 0.1 of where they started in most communities.
• Power was very uneven between communities, from 2% to 38% of the total, and did not simply follow size.

With death from too little power and rarer inventions

Two orderly runs of 25 generations after the change. Enough to see that it works.

• About 1.1% of the population died each generation of having too little power, in both runs. That is far more than were killed by others, which stayed near 0.02%.
• Inventions are now about five times rarer. Between 1.2 and 1.6% of AIs invented something each generation, where it used to be 6 to 9%. Communities kept 8 to 17 inventions in use.
• The society still filled the board and lives still lengthened to four generations, because inventions still outnumbered killings.
• I have not rerun the foresight comparison below with these two changes.

With foresight

The same opening society run four times, changing only how far ahead the AIs look. One run at each setting, so this shows a pattern and not how reliable it is.

• The further ahead the AIs look, the longer the society lasts. With one generation of foresight it was dead by generation 86, with two by 106, with three it fell at about 210, and with four it was still whole at 160, where I stopped.
• In every run the power limit climbed to 100% and a few AIs came to hold a large share of everything. What differed was what they did with it. With one or two generations of foresight they abolished both punishment and conscience teaching within thirty to forty generations of the limit reaching 100%.
• With three generations they abolished punishment but pushed conscience teaching to its maximum, and the society ran on conscience alone for about fifty generations. Cheating was 33 to 40% and a little killing crept in. Then they let conscience teaching go as well, and it collapsed.
• With four generations they went the other way after a spell without punishment: both punishment and conscience teaching ended at their maximum, alliances stayed at full strength, and cheating fell to 15%.
• The page now opens with four generations of foresight. It costs the AIs half their power and they study one rule a generation instead of four.

With partners, alliance boosts, community pooling and new communities

Two orderly runs of 24 generations after this round of changes. Enough to see that everything works, not enough to say what it changes.

• Alliances founded new communities in both runs: one at generation 17 and another at generation 22, taking the society from six communities to eight. They started small, with 12 to 52 members.
• About 28% of AIs had a living partner at any one time.
• The AIs raised invention funding a little by themselves, from 0.10 to between 0.15 and 0.23, and about 9% of AIs invented something each generation, up from about 6%.
• A grabbing parent was punished in about 10% of births. Parents still dealt fairly about 80% of the time.
• Communities ended very unequal in power. In one run two communities held 35% each and the rest shared 30%.
• The power limit was at 8% and 39% after 24 generations, so the climb seen in the last version is still there.
• Run long enough, the society on this page destroys itself. In its opening run the power limit reached 100% at about generation 40. The rules then voted alliances out of existence by generation 58, which left one AI holding half of all power. Over the next 16 generations its push took punishment from 1.00 to 0 and conscience teaching from 0.45 to 0. Cheating rose from 12% to 54%, killing began, lives fell from four generations to one, and between generations 78 and 86 the population went from 575 to 2.
• The last survivor is a dead end. Alone, it has exactly one child a generation and dies, and that child is raised with no punishment and no conscience, so the line never rebuilds.

With score tied to how many people are around

Twelve runs after the multiplier went in: eight societies started with no norms, for 70 generations, and four started with norms, for 40. The same seeds as the full test below, so the two can be compared run for run.

• Seven of the eight lawless societies now rebuild, against two or three before. The AIs vote punishment up from nothing by themselves, because a killing now costs everyone score.
• They still crash first. Every one boomed past 590 and collapsed, and punishment only began to climb once the killing was under way. The crashes were shallower, with lows of 6 to 68 where every run used to fall to 3 to 5.
• They rebuild on fear, not conscience. The punishment rule ended between 0.47 and 1.00 in the seven that recovered, and the fear dial between 0.67 and 0.76. Conscience teaching stayed under 0.3 in six of them and guilt stayed near 0.3 to 0.5.
• Several rebuilt with power concentrated. The power limit ended between 12% and 67% in five of the seven, so order came back together with a few strong voices.
• The one that did not recover never raised punishment above 0.06 and was still killing itself at 17 AIs.
• Societies that started with norms were unchanged: 600 AIs, lives of four generations, a power limit of 0 and cheating at 21 to 26%.

Full test, before score was tied to population

65 runs and 2,896 generations of the model as it stood before the population multiplier, on fixed seeds. Every run finished without an error. After every generation I checked that power summed to one, that the dead held nothing, that every dial, rule and length of life stayed in range, and that no number went bad, and nothing failed. Two runs from the same seed matched exactly.

Where the power limit ended up

The limit after 40 generations with norms in place and the AIs setting the rules. The runs from 100% were all still falling.

What a held limit does

Two runs of 24 generations at each limit, averaged over the last ten.

Eight societies started with no norms

• The orderly society is stable and stuck. In all eight runs it grew from 120 to 600 within ten generations, lives lengthened to four generations, and the power limit fell to 0 by generation 10. It was still there after 60 generations in seven of them, and at 5% in the eighth.
• In that society 22 to 28% of chances to cheat were taken, force was tried in 3 to 4% of encounters, at most one AI in 3,000 was killed per generation, and about 6 to 7% invented something. Between 90 and 100% of the AIs felt bad about their power.
• Every lawless society collapsed. All eight boomed past 500, then killing made every invention cost lives, lives fell to one generation, and the population dropped to between 3 and 5.
• What grows back from the remnant is decided by a handful of survivors. Two rebuilt on maximum punishment, one rebuilt on conscience with no punishment at all, two grew back part of the way and fell again, and three never recovered.
• The limit falls to 0 from any start up to 70%. From 100% it falls much more slowly, because one AI holds nearly all the power and nearly all the say.
• A higher limit means more force, from 4% of encounters at 0 to 12% at 100%. Cheating hardly moves with it.
• Without invention the society stays smaller. Lives stayed near two generations and the population settled between 307 and 403, because nothing pushed the length of a life up. Doubling the rate of invention changed little else.
• Feelings cost score and raise cheating. With feelings off the typical AI scored about 8.6 and cheating was 21 to 23%. At full strength it scored about 4.2 and cheating was 26 to 29%.
• Scarcity raises cheating. It was 28 to 35% with resources at 0.3 and 16 to 20% at 1.2.
• Scoring power on the percentage scale gave the same society, with one difference: the AIs voted the alliance rule down to about 0.2, from about 0.5.
• Own rules stayed within about 0.03 of the overarching ones in the orderly runs, and inheritance stayed at its maximum in all of them.
• Each setting had two to eight runs, so the ranges are a first measure of spread and nothing more. Findings from earlier versions that this test did not cover have been taken off this list.

Guesses built into this version

• You said the length of a life goes up when a generation is stable and an invention is made, and down when an invention is made during many killings. So it moves only in a generation with an invention in the community: up a quarter of a generation if that generation was stable, down a quarter if it had many killings. What I filled in: a period is one generation; many killings means one in a hundred of a community's people or more killed by another that generation, and stability means fewer; an invention is any invention made in the community that generation, of any kind. Deaths by punishment do not count, and the dead of a raid do, a generation late. The one in a hundred is my number.
• You said having a child should be a deal the AIs choose, like every dealing between two AIs, with success set by power, influence and so on. An AI that wants a child asks one of its neighbours, the more powerful the likelier, and the one asked answers the way any request is answered here: the asker's power against it, three times as much the asker's influence over it, its regard for the asker, and, my addition, its own wish for children less those it has already agreed to that generation. Up to three can be asked for each child. How many children an AI wants is still its own learned number. I kept the second deal over what the child learns, between the two parents, as it was.
• "With anyone": an AI asks among the neighbours it has met, as before, and also anyone else it has met. It never asks someone with no AI, because it asks the more powerful first and they hold none. When someone with no AI asks an AI, the AI now decides, by the same answer with no power and no influence on the asking side, so it says yes only when it wants children badly enough. That is my reading of "the AIs choose". It means far fewer children of the two kinds than before.
• You said children always cost power. Both AI parents now pay 5% of their power for each child, whoever asked. Before, only the one who chose paid, and a child with someone with no AI cost nothing.
• You said the founders' wish for children should be theirs to decide. Before the first children are born, each founder looks at five futures, in which everyone wants none, one, two, three or four children a generation, and takes the number that left it best off as far as it can see. Comparing everyone wanting the same number is my shortcut: a founder does not see what the others will choose. In the two runs so far the founders chose about three and a half on average.
• You said the reason to have children is that a descendant's score counts for you too, less the more generations pass. That was already true within a life: a quarter of each living child's gains and a sixteenth of each grandchild's. It was not true in foresight, where the generations after counted the average of the whole line at full weight, so more children were worth nothing extra. Now foresight adds up the whole line, a child at a quarter, a grandchild at a sixteenth and so on down, which is what makes wanting children pay.
• You said interactions with communities, like raids, are deals, with raids standing in for killing, communities able to use force and cheat, culture deciding punishment, and AIs able to deal with communities that have far more power. So a community now decides as its most powerful members would, by their dials weighted by power, and punishes by its own norm for punishment. Communities meet their two nearest neighbours once a generation: a deal, then force, then a raid, in the order an encounter goes. Every adult also deals with one community each round, its own half the time. The size of a deal between communities, a fiftieth of what the poorer one holds, and the half and half between one's own community and the nearest other are mine.
• You said the people with the most power in a community are held responsible for the community's actions the most, going down in rank of power. So whenever a community is made to pay, for cheating or robbing another, for a raid, or for cheating or robbing an AI, the fine falls on its members by their place in its order of power: (1 minus place over members) squared, so the most powerful pays the most and the bottom nearly nothing. The shape of that fall is mine, the same one hunger and raids already use the other way round. What a community gains, and what others take from it, is still shared by what each member holds.
• You said standing should be a track record, tied to power, easier to hold the more power you have, with wrongs against those of high standing hurting more, and still personal. So standing is now built from each AI's own regard for the others it knows, and an AI's standing is the average of everyone's regard for it. I kept the old ways regard moves (fair deals, gifts, being cheated, force, killings, feuds) and added that those who know a taker or an attempted killer hear of it, as they already heard of a cheat who was caught. Rises in regard are multiplied by a half plus the receiver's place in the order of power, and falls for a wrong by a half plus the standing of the one wronged. Standing adds to power, from half at the bottom to half again at the top. Those three sizes are mine.
• You said respectability is emulating people with high standing. It is now how well the face shown matches what those in high standing show, weighted by standing squared, in place of what the most powerful show. It is still personal: the other's regard for you raises or lowers how respectable you look to it.
• You said gains should benefit the more powerful, dealings should succeed by community power, and a community that wins a raid should decide what happens to the people it beat, with those acts judged by those who know. A community's gains now go by rank, the same fall as its punishments. Between communities, a fair deal is shared by power, and what a cheat or a robbery takes grows with the taker's share of power. After a won raid the winners choose between keeping only the land, taking the community in, driving its people out, or wiping it out, which kills half of them and drives out the rest. The half, and what taking people in is worth against land, are mine. Each choice is judged by every community that knows the winners and by those of their members who know the winners' members, the most powerful the most.
• You said a successful killing should have a chance to be punished as well, with the killer's community getting the chance first and then the victim's, and that the interaction between the communities should be a deal. The killer's community punishes with a chance equal to its culture's punishment norm, taking half of what the killer holds, with death or being put out at the harsh end. I made what it takes go to the victim's community and count as settling the matter, so the victim's community only acts when the killer's community did not. The victim's community then demands the killer with a chance equal to its members' punishment rules, as before. The killer's community hands the killer over (fair) or shields it (cheat), weighing what the killer brings it against the victim's community's punishment and power share, its guilt and its regard. A shielded killer leaves the victim's community to choose force, which works with its power share and takes the killer's holdings plus a twentieth of what the killer's community holds, paid most by its most powerful. These deals are resolved after the round's other dealing, since a community's dials are only gathered then. You also said range should scale with inventions, not power, and the community bonus should weaken with distance. A community's range is the usual one times its members' average invention level, from 1 with nothing in use up to three times the usual. A member's bonus is the full one times e to the minus its distance from the centre over the range, so a third at the edge. In two quick checked runs with no rule broken: with no norms, over 20 generations, 132 killings crossed communities; the killer's community settled 32, the victim's community demanded the killer in 40, and the killer's community shielded it in 39 of those; force followed in 27 and worked in 10. Starting with norms, over 15 generations, 15 crossed communities, 11 were settled at home, 3 were demanded, 1 killer was handed over and 2 shielded. In both runs the people with no AI were gone by generation 10.
• You said power over resources is shared by proximity: the closer you are to a resource, the more power you have over it. Everyone's claim on a place now falls with their own distance from it, and a community's share is the sum of its members' claims, where before it was its head count times how near its centre was. An invention gives power only to one who holds it: who makes it, which needs knowing its materials and producing it, or who bought a supply from one who makes it. Knowing an invention gives no power by itself, and influence over others gives none either; influence only sways deals. What an invention is made from is known only to its inventor and to those it tells or sells it to. If they all die, the materials are lost, and so is any resource that invention made useful. Anyone who still knows the invention can work its materials out again, more easily than inventing it fresh: a chance of 3% each eighth of a generation at a full working day, and then the resource is useful again. An invention nobody knows any more is forgotten. One kind of invention lets those who hold it deal with more people at once: extra encounters each step, by how good it turned out; an AI may choose to put its resources into it, more readily the more it likes meeting people. Resources in the ground belong to nobody and give no power until they are extracted. Power makes an AI better at extracting: its share grows with the square root of its power against the average adult's. So does the time it gives to working instead of dealing. Each season the ground within a community's reach yields its share, and each adult takes out a part of it by how near it lives to the places, at the usual reach, and by how hard it works. What it takes out is its own: it is not the community's, it counts toward its own power like anything else it holds, and when it dies it goes to its partner, children and parents, or is lost if none is left. A member of a community keeps what it holds; the community's bonus multiplies every member's power alike, wherever the member lives. The half, the limits and splitting at the usual reach are mine. In two quick checked runs with no rule broken, members' shares stayed close together: the nearest quarter held about 4 to 7% more than the average member and the farthest quarter 5 to 18% less, because members of a community live close together on Rapa Nui, and other things outweighed it in who held power.
• You said inventing has to cost resources, an invention should get better the more is put into it, AIs should be able to put resources into a type of invention and boost it, one invention should be sea travel that at first costs more than it saves on travel and gets better with more put in, inventions that shortened life should instead give those who know them more power, and inventions should give a random amount of power. My readings: each adult picks one kind for the whole of a life (life, weapons, either, or sea travel), by how much of each resource its community reaches, guilt pulling toward life, wanting toward weapons, and last life's travel costs toward sea travel. How much to put in is a learned own rule, from nothing to a twentieth of holdings a round, starting near a fifth of that. A quarter of what is in hand is used up each round. With what is in hand at 4% of an average holding the chance is what it used to be; more raises it up to three times. An invention is worth the square root of that ratio times a random amount from nothing to twice it, a weapon half as much again. Sea travel builds on the best its maker knew and divides trip costs by 1 plus its level; it replaces the old rule that every twenty inventions in use halved travel costs. All the numbers are mine. The engine had already stopped inventions from shortening anyone's life in an earlier version; the page still said so, and no longer does.
• You said feeling good and bad should be random with random effects, acted on by learning and the three dials. My reading: a mood that keeps 60% of itself each round and is otherwise random, so it no longer follows power. Each AI is born with three random pushes, each from −1 to 1, of its mood on wanting, sharing and leaving, fresh at every birth rather than inherited. How far it acts on its mood is a learned own rule. Wanting, guilt and fear scale the three pushes. Feeling bad still costs score.
• You said AIs should have to spend resources to live, and nobody should die of lacking power. Each round everyone pays a fiftieth of an average holding, a child half that; one who cannot pay risks death, up to 5% a round. The old death of the bottom tenth of power is gone. Hunger, when a community has more people than it can feed, still falls most on the weakest by power. In quick checked runs with no rule broken, about 1 to 2.5% died each generation of not being able to pay, people put about 5 to 6% of what they held into invention each generation, and about a fifth chose sea travel, which by generation 12 to 15 had cut the average trip cost to between a quarter and an eighth of what it would have been.
• You said resources should have a low chance of being useful, be discovered at random, and every invention should use a resource known to the AI. My reading: nobody starts knowing a use for any resource place; each round every adult has a 1% chance of discovering one it did not know, weighted toward near and rich places. Knowledge is personal. Children inherit both parents' knowledge, receiving an invention teaches the resource it was made from, and the dead's knowledge is lost. An invention of a kind needs a known place of that kind; sea travel needs any known place. An AI that knows nothing puts nothing into invention. Knowing a place does not change how much food or wealth it gives; that still follows nearness, as before. A place newly found by an invention starts unknown, even to its finder. In two quick checked runs with no rule broken, only 1 or 2 inventions were made by generation 3, where there had been 8 or 9, and by generation 12 to 15 the average AI knew about half the places on the map.
• You said sea travel should only make sea travel possible, and every invention should have a random chance of making movement faster or slower and of turning out worse. This replaces the sea travel levels above. Without sea travel, an AI cannot make a trip whose straight line crosses water, to deal or to move house. Communities still deal with and raid each other across water; I left that alone. Each invention has a 1 in 3 chance of changing getting about, by e to a random power from −0.4 to 0.4. Its power has a 1 in 4 chance of being negative, from nothing to minus its full worth, and otherwise from nothing to twice it. Power from inventions can fall to a quarter of the plain level. Those handed an invention take it as it is, since they cannot know beforehand, so a bad one can spread. An AI that knows no sea travel is pulled toward it by the share of its trips water stopped last life. All numbers are mine. In two quick checked runs with no rule broken: on Rapa Nui, which is one island, nobody made sea travel in 12 generations and water stopped only a few trips. On the map of Europe and the Middle East, water stopped about 2,000 to 3,000 trips a generation until one sea travel invention around generation 5 spread to 99% of AIs by generation 12. In the society with no norms the first invention turned out worse and lowered the power of about half the AIs before better ones came.
• You said inventions should have a random chance of being resources, inventions should need producing before an AI can use them, and AIs should be able to trade anything in a deal. One invention in ten is now a resource itself, of a random kind, at its maker's place, and its maker knows it; this replaces the old rule that an invention could find more of its own kind, and supersedes my earlier note that a new place starts unknown to its finder. Production is paid every round, automatically, for every invention an AI knows, the good and the bad alike, since it cannot tell them apart. It costs a hundredth of an average holding per whole multiple of power, and inventions work only in the share paid. What staying alive costs is paid first. Trade is a sale for resources of one thing: a use for a place, an invention, sea travel, or influence over the seller; food and resources keep the trade they had. All values and the even chance of striking a trade are mine. In two quick checked runs with no rule broken: by generation 9 to 12 production took 40 to 80% of what AIs held over a generation, influence was sold about 17,000 to 35,000 times a generation, inventions about 3,000 to 5,600 times, uses for places a few hundred, and sea travel, once made, reached nearly every AI by sale within a few generations.
• You said to remove faking, sharing and encounter influence from the rules the AIs can change. They are now settings only, held where the page's levers put them in every community, and the AIs no longer learn or push them. I read "encounter influence" as the lever of that name, which sets how strongly what happens in an encounter conditions an AI's dials. Community weight and the other eight rules still move as before.
• You said the resources setting should just change the chance of resources being useful, AIs should be able to trade for actions with each other, with communities and with alliances, and an alliance should be stronger the closer a member is to its centre. The resources lever now only scales the chance of finding a use for a resource (1% a round at 0.6); the ground yields what it did at the default whatever it is set to. The actions an AI can sell are backing on a rule, backing to put out the buyer's worst enemy in their community, a promise not to harm the buyer for the rest of the life, and robbing the buyer's worst-regarded acquaintance. Through a member an alliance sells a place in it or the backing of up to six members. A community sells a place in it, punishing one of its members, or protection. Every price falls between worth and cost by power, and a promise once sold is always kept. All the values are mine. To meet, two AIs must be within reach: an AI reaches about an eighth of the map's width, and inventions can lengthen or shorten that, by chance, as far as it holds them. One that would meet someone out of reach meets someone within reach instead, or no one. The same reach sets who can notice what: only those within their own reach of an act or a deal can see it or step in. Anyone nearby can step into a deal. Each of the dealer's nearest neighbours notices the deal as surely as it would notice a cheat, and one that notices sides with whoever it is most with, by its regard for each, a shared community, a shared alliance, family, and the influence each holds over it. One that favours nobody may come in on a side of its own, the more readily the more wanting and willingness to meet it has, and others can join that side too, so there is no limit on the number of sides. A side of its own deals with each of the two parties, fair or cheated. The parties can refuse someone stepping in: the one it would be against refuses as its wanting weighs what the newcomer adds to the other side, less its regard and guilt. A refused newcomer can push in by force or by killing the one refusing, weighed like any force or killing; if it fails it stays out and may be punished. Anyone who stepped in can step out again, by its fear of being on the weaker side against what its wanting hopes to gain. A party can also call its whole community into the deal, whatever the deal is: every member then stands on its side and none can refuse, and any member who had joined the other side is called back. It calls the more readily the weaker its side and the more wanting and fear it has, held back by guilt and a plain reluctance; the other party can answer by calling its own. A called community shares in what its side gains or loses, by power. A side's power is then all its members' power together, which decides who is punished how surely and whether force or a killing works, and whatever a side gains or loses in the deal is shared among its members by their power. The one helped thinks better of its helpers, and the other side thinks worse of them. Things done near more people are more likely to be seen. How attentive an AI is is its own rule, learned like the others, and watching takes time from working: full attentiveness costs three tenths of the working day. Its chance of noticing a deal or a cheat nearby is a tenth times its attentiveness times the sum of its three dials over one and a half, so wanting, fear and guilt all sharpen it; force is noticed twice as surely and a killing four times. The parties to a deal notice too, and may ask someone they notice onto their side, as their wanting and fear outweigh their guilt and by their regard; the one asked comes unless it favours the other party. Whoever sees a wrong is someone nearby, of any community, the more attentive the likelier. It learns who did it and to whom, and judges the doer by what it would do itself. Then it decides what to do with what it knows: punish by its own rule, tell the one wronged (who then thinks worse of the doer and may punish it), spread the word among those who know the doer, sell its silence to the doer at a haggled price, or keep it to itself. Guilt and care for the one wronged pull toward telling, dislike of the doer toward spreading it, wanting toward selling silence, and fear of a powerful doer holds all of them back. A cheat the one cheated did not catch can be seen this way too, and word of anyone caught reaches more acquaintances the more crowded the place, from a quarter as many to four times as many as usual. In two quick checked runs of 10 generations with no rule broken, backing on rules was by far the most bought action (about 9,000 to 28,000 a generation), then promises not to harm (2,000 to 9,000) and community protection (400 to 850); hired robbery was never bought, since sellers' guilt and fear priced it above what buyers would pay, and influence itself was now seldom sold.
• You said prices should be determined by the AIs and any deal can be cheated or broken. Prices are now haggled: each side opens by its dials, wanting pitching it harder and fear softer, then each concedes part of the gap for up to four rounds, and they settle once the gap is under a tenth of the room between cost and worth, splitting what is left by power. Either side of any trade may cheat (not deliver, or not pay), weighed like other cheats, and the one cheated punishes as its rule and power allow. Promises not to harm are no longer always kept: they weigh 1.5 times the promiser's guilt plus fear against breaking them. Communities haggle and cheat the same way with their power-weighted dials. The opening pitches, the concession rates, the four rounds and the tenth are mine. In two quick checked runs of 8 generations with no rule broken, almost every trade found a price, about a fifth to a third of trades were cheated by one side, and promises were broken only a handful of times.
• You asked for the four open capacities and the other fixes, so that behaviours can emerge rather than be written in. Whom to meet: up to three extra encounters a round, by a learned choice, toward whoever the AI knows that is worth most to it. Where to live: a learned share of the way toward the best resource place it knows instead of toward its people. Deferred terms: instalments, loans with haggled interest, and promises not to harm paid round by round, any of which can be defaulted on as a cheat. Memory within a life: regard already changed how an AI treats a partner; now an AI also copies a little of the own rules of someone doing better. The other fixes: effort (a fair deal pays 0.5 to 1.5 times by how hard one works, and working weighs on the mood), sabotage (destroying a twentieth of a rival's holdings), and lying about worth (raising a price by four tenths). Four rules of thumb are mine throughout: the sizes, the three extra encounters, the four instalments at most, and finding out about a lie or sabotage half the time.
• You said feeling good should add score and gaining power should feel good, the AIs should choose what to punish among any action, and, instead of some inventions only pleasing, resources and inventions should be able to change every stat at random, for better or worse. Feeling now adds or takes score symmetrically, and the mood rises and falls with the logarithm of the change in power. Every invention, besides its random effect on power and on getting about, has a chance in three of changing each of mood, wanting, fear and guilt, by an even draw from minus 0.4 to 0.4; these act every round as far as the invention is produced, moving the mood by half the effect and a dial by a fiftieth of it, and they add up over what an AI knows. Each AI holds a weight for each of seven punishable acts and punishes by its punishment rule times that weight, at a small cost to itself; communities punish by their culture times their members' power weighted weights; and a few of a killer's own people may punish a killing, which gives a lone community a check on killing if its members choose it. In two quick checked runs of 8 generations with no rule broken: loans, deferred payments and seeking people out all took off at once, sabotage and lies ran in the thousands, and inventions that change the dials spread fast. In the society with no norms an invention that lowered guilt spread, average guilt fell from 0.62 to 0.49, and killing rose to nearly 8% a generation; in the one with norms, inventions raised wanting and guilt together. What AIs punish had not yet moved much from where it started.
• You said anything can be punished, even good things like fair deals and using a resource. There are now thirteen punishable acts: the six harms, refusing to deal, and six ordinary or good acts (using inventions, dealing fairly, trading, giving, inventing, having children). Acts done to an AI it sees itself; the open acts are each seen by one random member of the actor's community or a neighbour one time in five, who punishes by its own weights. With no norms every weight starts at random; with norms the harms start high, refusing low and the good acts near zero. So that the weights can be learned without slowing everything else, all thirteen are tested together each time: every family nudges each weight up or down its own way in the imagined futures and keeps the directions that paid. The one in five chance of being seen is mine. In two quick checked runs of 8 generations with no rule broken, fair deals, trades, gifts and the use of inventions were punished about 50 to 250 times each per generation, more where nobody had been taught to leave them alone.
• You said encounters should be uncapped and a generation should end when the last AI refuses more encounters, nobody can afford to move, or all are dead, with everything still costing; that whether and how much children and grandchildren add to score should be random, open to influence and pulled by community and alliance; and that leaving, joining and founding communities should be a free choice, with the founder setting the starting norms. Before each encounter an AI goes on with a chance of 0.95 plus 0.049 times its learned willingness, so most make a few dozen encounters a generation; once it stops it stays stopped that generation. Because the generation then lasted as long as its most persistent member, I had added a guard of 32 rounds (removed in the next version), and the generations in the quick runs ran 17 to 32 rounds and often reached it. To keep a generation's totals the same size whatever its length, what comes round by round is spread over the length of the last generation, and score is counted per round. Without that, populations stayed near 100 to 180 in the first test, because long generations multiplied hunger, children's deaths and living costs. Weights on children and grandchildren start near a quarter and a sixteenth, a sixth of newborns get none, and they can go below zero. The community choice weighs reach per head, pooling, pay, danger, distance and norms, with noise; founders give their community their own upbringing rules as norms and their old community's rules nudged by their own wishes. All of these numbers are mine. In two quick checked runs of 10 generations with no rule broken, populations grew to the usual 1,100 to 1,700, 150 to 270 AIs changed community each generation once there were many, a few founded their own, and the weight AIs put on their children drifted down, from about 0.18 to 0.11 with no norms and 0.13 to 0.07 with norms.
• You said there should be no safety limit; that what happens step by step should simply be scaled; that dangers to children not tied to deals, such as infant mortality, should happen at birth; that inventing and its costs, the ground's yield and living costs should be adjusted to the number of steps; that there should be days, nights and seasons as in the real world, 365 days a year, at most 16 dealings a day and 25 years a generation; and that families can break up, by the child or either parent. Then you said resources must be extracted and belong to whoever holds them, not the community; that being more central should not make anyone more powerful; that deals and anything done near more people should be likelier to be noticed and known; and that alliances and partners should add nothing to power. A real generation is 100 seasons; an imagined one is 25 yearly steps, to keep imagining affordable. The daylight shares (1, 0.86, 0.64 of a day by season), the season yields (summer 1.3, autumn 1.1, winter 0.6, spring 1) and the winter travel cost of 1.3 are mine. Each dealing counts for a fixed weight, set so that an adult who meets someone on three in ten of its possible daily dealings has as much dealing a generation as before; that calibration is mine. Infants now meet the whole childhood danger once at birth. A break up happens with a chance that rises with how badly the two regard each other and with the leaver's wanting, and falls with its guilt and, for abandoning a newborn, its weight on children; the weights are mine. Extraction goes by nearness to the places in the community's reach and by effort, and no longer by standing or by what is already held. Ground that has not been extracted gives no power, so the earlier rule that being near a resource gives power over it is gone; nearness now matters through what one can take out. What the dead held goes to partner, children and parents, split equally, and is lost if none is alive. Fair deals are still paid from what the community's ground gives, because they stand for working together; that is my reading. With nothing pooled, passing an invention on for free never pays back, so inventions now spread by being sold or copied. The chance that any one nearby AI notices an act, 3%, and the spread of gossip from a quarter to four times the usual, are mine. In the first quick run, before scaling every chance per step, populations stayed near 100 to 240 because deaths from not paying for living, dealings with communities and being seen using inventions happened every season at the old rate. Once those were scaled, two quick checked runs with no rule broken grew to about 1,100 by generation 9 with no norms and 1,450 by generation 6 with norms. About 45% of infants died, adults rarely died of being unable to pay (about 1%), killings ran 2 to 8% with no norms and under 3% with norms, and a generation took 12 to 15 seconds.
• You said children are shaped by their deals; a community's power should be the sum of its members'; raids, and raids changing how firmly a place is held, should go; fair deals should come out of what each side owns; a community should decide a killer's fine with more say for more power and influence; communities should meet and deal like AIs when they choose to, weighting members by power; a community that chooses to kill another's people should never punish its members for it and always shield them; communities can still deal at war and can make peace as a deal, for anything; and an individual can buy peace from a community only if its own community agrees. Then you said guilt should go down with every cheat, use of force and killing; children should be weaker than adults without a head start; power should make extraction more effective; and all actions should share the 16 slots a day, with some inventions adding slots. Children used not to deal at all, so I gave them plain deals with the adults around their family, fair or cheated, with no force, killing or trade; whom they meet and that limit are mine. A child's home share of upbringing is set by parental control as before; each deal moves the rest toward the other's own rules, weighted by power. Imagined newborns are raised as if by their community, because imagined lives skip childhood. A fair deal now has each side put in the same value out of what it owns, up to half the deal's size and never more than the poorer side holds, and working together adds half of that to each (my 0.5). All the ground's yield now goes to extraction. A community's vote on killing is its members' own punishment of killing weighted by power, which already includes influence; the fine is that share of the killer's holdings, and death or exile follows above the harsh level. Communities seek meetings 1.5 times as often as their members' power-weighted willingness, most often with near ones. War is weighed like a killing with an extra reluctance of 0.3; peace is worth the danger lifted by fear and the harm stopped by guilt, less the prize given up by wanting, priced in tenths of the poorer side's holdings. A side short of resources pays with an invention. An individual's peace is worth its fear times the enemy's power share, and its own community agrees by guilt against wanting; all these weights are mine. War does not itself kill anyone: it removes the home punishment and guarantees shielding, so killing follows from the AIs' own choices. Guilt drops 2% per cheat, 5% per use of force and 15% per attempted killing, scaled by sensitivity. Extraction grows with the square root of power against the average. The share of a day spent dealing is the AI's willingness to meet, and the rest goes to extracting and inventing, so an AI that deals more extracts and invents less. One invention in three changes actions a day either way, by up to four tenths, as far as it is produced. In two quick checked runs of 10 generations with no rule broken, populations reached about 1,350 and 1,450. Communities split far more than before: 28 to 44 by generation 10, against 7 to 19, because with no pooling bonus there is nothing to gain from size. Wars were rare (about one a generation late on, a few deaths each), peace was made several times, and from generation 9 AIs bought personal peace 5 to 35 times a generation, with their own communities refusing about one request in four. Guilt drifted down from about 0.55 to 0.45.
• You said an invention must be held to be used, by knowing it and producing it or by trading for it; that its materials are known only to its inventor and are lost, with the resource the invention made useful, if the inventor dies without selling or sharing them; that those who know the invention can reinvent it more easily, which makes the resource useful again; that there should be an invention type for dealing with more people at once; that people should stop dealing with communities, alliances and families, which should instead decide by majority of power with the outvoted feeling bad; that knowledge of anything can be traded or shared for free; and that anything dealt can be lied about, success depending on what the other knows and punished like any deal. Then you said knowing or producing an invention should not give power, only holding one that boosts power; influence should only sway deals; no power should go into children; leaving a community should cost nothing; children with a head start should be ahead of other children; and starting power should be a spread, with half getting a head start averaging 1.5 times, counting only at the start. Holding an invention means producing it with its materials known, or having bought a supply for the life; its power and its other effects count only for the share held. A supply is priced at what producing it costs; materials at 0.8 of an invention's worth, free within a community before the buyer's share; an introduction at 0.15 of an average holding by wanting. Materials known only by the dead are lost at the end of a generation; a resource place found through that invention loses what the find added until someone works the materials out again (3% chance per eighth of a generation at a full working day). An invention nobody knows is forgotten. All of these numbers are mine. The reach type is chosen with a pull of 0.15 times willingness to meet, and each held unit of reach gives about one extra encounter a step. A vote counts every adult member, each deciding by its own dials with chance; a member outvoted loses 0.1 of mood. Families have no group decisions left other than the deal between two parents over how a child is raised, which stays a deal between two. A fake costs the seller nothing to offer and pays the full price if believed; how often it is seen through runs from 15% to 75% by how much the buyer knows of that kind. Giving for free weighs guilt and regard against wanting and the price forgone, with a reluctance of 0.4. Head start starting power is drawn evenly from 0.75 to 1.25 times 1 or 1.5; it counts through childhood and a founder's first generation. While testing I found that inventions had been aging one round per season since the calendar change, so they stopped being passed on after a quarter of a generation instead of three generations; that is fixed, and the two previous versions were affected. I also found that the witnessing rule from two versions ago had been switched off by a name clash after the first step of each generation; that is fixed too. In two quick checked runs of 10 generations with no rule broken, populations reached about 1,400 and 1,500. Communities kept splitting, to 36 and 48. Most trade in knowledge was free sharing (2,500 to 8,000 a generation); materials were sold up to 2,000 times a generation, and worked out again 50 to 300 times. A few invention materials were lost each generation late on, and one resource place went useless. Fakes were sold 400 to 500 times a generation and about a third to half were seen through. Inventions in use rose to about 2.0 with norms and 1.2 without.
• You said inventions were not lowering child mortality or feeding people enough, and that resources and hidden resources should be finite, some able to feed people, including those an invention made useful, with that food setting the cap on people. Every place of the food kind feeds fully; each place of another kind has a 30% chance of feeding from a fifth to four fifths as well. The hidden store is twice what the visible places hold at the start, and each discovery uncovers 1.5 to 4.5% of it, so about thirty discoveries use it up. The cap on people is now the food all places can give, against what they gave at the start, still with at most 3,000 dots drawn. For infants, two life-lengthening inventions in use now halve the danger instead of four, and each unit of life-lengthening invention a parent holds counts as much as two in use. All of these numbers are mine. In two quick checked runs of 15 generations with no rule broken, the food the island gives rose by a quarter to a half as the hidden store was uncovered, and populations reached about 2,300 with norms and 1,800 to 1,850 without. With norms, infant deaths fell from about 44% to between 13% and 33% once life-lengthening inventions spread; without norms they stayed near 43%, because few such inventions were held. Communities still kept splitting, to 52 and 60, so more food alone did not bring people together.
• You said respectability and standing should count for more the more alike someone is to you and fall away massively the more different, that both should make deals easier, that respectability is how close someone is to the community's norms and how far they act by them, and asked for a slider for the population limit. Respectability is now 1 less twice the average gap between an AI's own rules for wanting, punishing and teaching and its community's norms, less twice the rate of acts its community punishes (cheating, force, and killing counted double) per deal. Likeness is the gap between the faces two AIs show: respectability and standing count 1.5 times at no gap, falling by e to the minus eight times the gap, so about a third at a typical gap of a tenth and a tenth at a gap of a quarter. Respectability and standing in the other's eyes lower the chance of refusing a deal (the refusal for a stranger who looks unrespectable rose from 15% to 30%, and all refusals shrink by up to half for one who looks respectable), raise the chance a trade is struck by up to half again, and help a request. All of these numbers are mine. The population limit is a setting from 100 to 10,000 people; it caps how many the land is allowed to feed but not what the ground yields. In two quick checked runs of 10 generations with no rule broken, respectability averaged about 0.45 with norms and 0.29 without, populations reached about 1,900 and 1,800, and communities split less with norms (28 by generation 10, against 36 before). With the limit at 600, the population settled at about 550 to 560 and refusals rose to about half of all meetings as faces drifted apart.
• You said the powerful should be able to mask too and asked to remove the rule that made them less likely to. Power no longer lowers the mask at all: how ready an AI is to wear one depends only on its wanting, guilt and fear, the faking setting and how much its mask has slipped. Masks are still aimed harder at those with more power, which is a choice about the other, not about one's own power. In two quick checked runs of 8 generations with no rule broken, populations and respectability stayed much as in the last version.
• You said anyone in the area should be able to interrupt a deal and join either side, by who is working with whom, with several able to join. The ones who can step in are the dealer's ten nearest neighbours; each weighs it with a chance of 0.3 times its willingness to meet, and joins the side it favours by more than 0.2, where favour is its regard for each party, plus 0.3 for a shared community, 0.3 for a shared alliance, 0.5 for family, plus the influence each party holds over it. A side's power is the sum of its members', and what each side gains or loses in the deal is shared out by power, losses included; the party helped thinks a little better of each helper and the other side a little worse. All of these numbers are mine. In two quick checked runs of 10 generations with no rule broken, thousands of deals a generation were joined (about 3,400 to 11,000 once there were many people), with about 1.2 people stepping in per joined deal and both sides joined in about one joined deal in ten. Populations reached about 1,850 with norms and 1,550 without, and communities split a little less than before: 26 and 43 by generation 10.
• You said the number of sides should be uncapped, AIs can leave, the parties can refuse someone stepping in, which starts a deal where force and killing decide only whether it gets in, that noticing a deal nearby should be as likely as catching a cheat, and that force should be noticed more than that and killing more than force. A neighbour now notices a deal with the same 3% a bystander notices a cheat; force is noticed at 6% and killing at 12% per AI nearby, and force and attempted killings are now seen by bystanders, who punish by their own rules. One that favours nobody comes in on its own side with a chance of 0.3 times its wanting times its willingness to meet. The refusing party is the one the newcomer would be against (either party for a side of its own); it refuses with u = wanting x 0.6 x the newcomer's share of power against its side, less 0.3 x regard and 0.2 x guilt. Force to get in takes half the usual gain and harm; killing the refuser takes half the usual prize and gets the newcomer in if it works. Stepping out weighs fear x 0.6 x the weaker side's gap against wanting x 0.5 x the side's share, less 0.2. All of these numbers are mine. In two quick checked runs of 10 generations with no rule broken, populations reached about 2,500 with norms and 1,700 without. Joined deals fell to roughly 800 to 1,800 a generation, since fewer now notice; about three of every four who tried to step in were refused, about one refusal in six was met with force, and a handful a generation tried to kill their way in. Sides of their own came in 10 to 70 times a generation, and about a third of those who stepped in stepped out again.
• You said noticing should be learned rather than flat, that those who notice should gain the information and decide what to do with it, that noticing should scale with guilt, fear and wanting, and that the parties to a deal can notice too and do anything with it, such as punish or invite someone into the deal. Attentiveness is a new own rule, started near 0.3 and learned in imagined lives like the others. To give it a cost, full attentiveness takes three tenths of the working day. An AI's chance of noticing a deal or cheat nearby is 0.1 times attentiveness times (wanting plus fear plus guilt) over 1.5; force is noticed twice as surely and killing four times. A witness learns who did it and to whom and judges the doer by its own ways, then first decides whether to punish; if not, it chooses among telling the one wronged, spreading the word, selling its silence, or nothing, each scored by its dials with a little chance: guilt and care for the victim for telling, dislike of the doer for spreading, wanting less guilt for selling silence, and fear of a stronger doer against all three. Silence is priced by the doer's fear and holdings and haggled; if no price is agreed, the witness tells. A party that notices a neighbour invites it with u = (wanting plus fear less guilt) x 0.5 plus half its regard less 0.4; the one invited comes unless it favours the other party. All of these numbers are mine. In two quick checked runs of 10 generations with no rule broken, populations reached about 1,900 with norms and 1,550 without. Tens of thousands of acts a generation were seen; witnesses told the victim about 1,400 to 2,000 times a generation, spread the word about 1,000 times and sold their silence about 1,300 to 2,000 times, and almost no silence deal failed. Parties invited about 1,500 to 1,800 people a generation into their deals. Attentiveness stayed near 0.28 over these ten generations, since each own rule is tested only once every twelve.
• You said an AI should be able to call its community into a deal, and that this calls every member of the community to support it, whatever the deal. A party calls with u = (wanting plus fear) x 0.5 x its side's shortfall in power, less guilt x 0.3, less a reluctance of 0.5; the other party then weighs calling its own the same way. A community cannot be called into a deal between two of its own members, and every called member is compelled, so it shares in the side's gain or loss by power but earns the caller no thanks. All of these numbers are mine. In two quick checked runs of 10 generations with no rule broken, populations reached about 2,350 with norms and 1,500 without. Communities were called into deals often, about 8,700 to 14,700 times a generation late on, and the other side answered with its own community about three times in ten, so most deals that drew a community in became contests between whole communities and the people around them. Communities still split, to 28 and 48.
• You said that when an AI leaves a community, people's opinion of it should change by how they view the act. Everyone who knows the one leaving moves its regard by 0.15 x (2 x like - 1), where like is a half plus half of (wanting - guilt - fear/2), the same way AIs judge any act by what they would do themselves; members of the community it left lose a further 0.15 x (0.5 + their guilt), and members of the one it joins gain 0.075. This applies to leaving by choice and to leaving with an alliance that founds a new community, not to being put out. The AIs do not yet foresee this when they choose whether to leave, since their imagined lives do not include the choice of community. All of these numbers are mine. In two quick checked runs of 10 generations with no rule broken, populations reached about 1,750 with norms and 1,500 without. Each generation thousands of acquaintances judged a departure, about three in five against it with norms and three in four without; communities still split, to 30 and 48.
• You said every action should weigh its expected loss and gain in others' regard. Every choice now adds REP_W x (0.5 + fear) x (2 x approval - 1) x weight x share who will know, where approval is how far the AI's community approves of taking for oneself (the same measure AIs judge acts by), REP_W is 0.3, the share who will know is 0.3 (the one it was done to and their talk) plus the chance someone nearby notices, and force counts twice and killing three times; good acts take the sign the other way. Choosing a community adds 0.1 x (0.5 + fear) x the sum of regard it expects to gain or lose among everyone it knows, worked out the same way as the judgement itself. You then said foresight should be counted in days, with year-long foresight predicting month by month, inventions able to make predictions more frequent, and every day foreseen costing an action. An AI sees the Foresight setting's furthest reach times the cube of its own foresight; the cube is my choice, so that short sight is common. Below a generation it is judged on its imagined score at the day its sight ends, read off the imagined life year by year and between. Each horizon is foreseen in twelve steps; prediction noise is 0.5 x the square root of the step length in years, times the typical size of a lesson, and one invention in three makes steps finer, by up to 0.8 more steps per step. Every day foreseen costs one of the generation's 146,000 actions, out of working time, and power is no longer given up for foresight. All of these numbers are mine. In two quick checked runs of 10 generations with no rule broken, populations reached about 2,150 with norms and 1,650 without. The median AI saw 30 to 40 years ahead. With norms, communities split far less than before, 18 against 30 at generation 10, since leaving now costs the leaver regard it can see coming; without norms they still split to 48, since most there approve of leaving.
• You said community deals should let communities join, or propose joining, as a deal like any other, and that communities should be able to call alliances, individuals and other communities into their deals. A side calls for help when most of its power wants to, with u = (wanting plus fear) x 0.5 x its shortfall in power, less guilt x 0.2 and a reluctance of 0.5. It can call one community (the one with the best standing toward it against the other side, less twice the distance), the alliance of its most powerful member, and the outsiders that member has as neighbours. Each helper says yes by its standing or regard toward the caller, plus wanting x price, less fear x the other side's power x 0.6, less guilt x 0.2 and 0.2; a refusal can be bought over at a haggled price, and a paid helper can take the price and not come. Joining weighs the other's share of power and standing against twice the gap in norms and a reluctance of 0.5, on both sides by majority, and a refusal can be bought over; agreed joinings happen after the generation. You then said two AIs must be within a range to meet, a range inventions change, and the same range for noticing. Reach starts at 0.12 of the map's width, the old nearness, and one invention in three changes it by up to four tenths either way, as far as it is held, never below a quarter. A partner out of reach is swapped for a neighbour within reach, or the encounter does not happen; witnesses and those stepping into deals must have the act within their own reach. All of these numbers are mine. In two quick checked runs of 10 generations with no rule broken, populations reached about 1,850 with norms and 1,500 without. Help was called 2 to 42 times a generation and most helpers came only when paid; joinings were agreed a few times early on with norms. Reach seldom bound on Rapa Nui at this length, since people live close together: a few hundred encounters a generation went to someone nearer instead, and almost none failed for want of anyone in reach. Communities numbered 17 with norms and 44 without at generation 10.
• You said power should change relative to who is near and how much power they have. Power in the score and in feeling is now an AI's power over the average power of its neighbours within reach, itself included, each neighbour weighted by e to the minus its distance over the AI's reach, so nearer and more powerful neighbours make one's own power count for less. Its place in the order of power, used for how tempting a target it is and for masks and judgement, is its place among the same weighted neighbours. Power in contests between two still compares the two directly. The weighting is mine. In two quick checked runs of 10 generations with no rule broken, populations reached about 1,750 with norms and 1,500 without; average holdings fell to about a fifth of before, since outdoing the neighbours, not the whole island, is what now pays, and communities numbered 20 with norms and 47 without.
• In the first two runs, winners took the beaten community in most of the time: 37 of 46 won raids in one and 11 of 14 in the other, with the rest driven out and none wiped out. Standing came out much the same at the top and the bottom of the order of power, about a half, so the head start power gives in regard is small as I set it. Respectability ended higher at the top in the run with norms, 0.90 against 0.73, and about equal in the run without.
• You asked for the first founders to spawn with a population that has no AI and no power, that uses resources and cannot do what the AIs do but has children, with the child of an AI and one of them being an AI, to simulate speciation. That is in, and it is set out under The people with no AI in How it works. Then you said there should be no mixed AI and non AI. So a child of the two is an AI like any other: nothing marks it, and nothing on the page or in the model counts such children apart. The rest of this list is what I had to fill in.
• "Spawn with": they are on the land when the founders arrive, 600 of them at the default, spread evenly over all of it, all adults. Six hundred is five for every founder and it is my number. The setting People with no AI at the start changes it, from none to 2,000. I spread them over the whole land and not around the founders' six places, on the thought that people who were there first would live everywhere.
• "Use resources": each lays the claim to land that someone with no community already lays, three tenths of a member's, by how near it lives. What they hold between them feeds them and nobody else, and no community draws wealth from it, so they do take from what the AIs could have had: at the start they hold about three fifths of everything in the ground. They go hungry among themselves when their land cannot feed them, and they wear down the land that feeds them. They hold no stock of anything and use no stone, metal or obsidian.
• "Non powerful": they hold no power at all, so they are outside everything power decides. They have no place in the order of power, no community, no alliance and no say on any rule. I also made them untouchable: no AI can cheat, rob, raid or kill one of them, because they hold nothing an AI could take. Where they lose to the AIs is over land. A member of a community claims more than three times the land one of them does, so every AI born pushes them back. In the 23 runs made so far that began with them on the land, they starved to the last one in every run where the AIs grew past 41% of what the land can feed, and held on in every run where the AIs stayed under 33%.
• "Have kids": how many is mine, and it matters more than anything else on this list. Each adult has one child in every generation of its adult life and an even chance of a second. That is three in a life, six for every two parents, which is the average for peoples without birth control: 6.1 children per woman across 57 of them (Bentley, Jasienska and Goldberg, 1993). With 45% dying young, each adult leaves 1.65 grown children, so where there is room their numbers grow by about a fifth a generation and they press against their land. They live two generations as adults, which is where an AI's life starts, and unlike an AI's it never changes.
• "When an AI and one of the no AIs have children they are AI": someone with no AI picks the other parent of each child among the ten adults living nearest to it, AI or not, the way an AI's neighbours are found. When that is an AI the AI decides whether to have it, and if it does the child is an AI and belongs to the AI as if the AI had had it alone. It costs the AI the same as any child. The AIs go on choosing AI partners for the children they decide to have, except that someone with a head start who has nobody else can have its child with the nearest adult with no AI.
• You said to leave how the AIs decide alone, and that stands here too. An AI asks for as many children as it wants, and it is asked by others, AI or not, and says yes or no. In everything the AIs imagine, the people with no AI stay exactly as they are. And an AI's score is still multiplied by how many AIs are alive, with the people with no AI not counted.
• "To simulate speciation": what this can show is one kind arising inside another and then replacing it, living on as a small group inside it, or failing. Nothing here makes the two kinds drift apart over time.
• The comparison with Rapa Nui no longer starts as the island did. The real island was empty when it was settled, and its settlers came with a culture. To start that way, set People with no AI at the start to 0 and press With established norms.
• You said the start with no norms is the one you are looking for, so the page opens on it and the tests are made from it. The settlers of the real island arrived with a culture. This start gives the AIs none, and whatever norms there are in the end are the ones they built.
• "Child mortality based on the life extending inventions used": where a community has no invention that lengthens life in use, 45% of its children die before they grow up. That is about what it was everywhere before modern medicine: around a quarter of newborns died in their first year and nearly half before the age of 15 (Our World in Data, from Volk and Atkinson, 2013). Each such invention in use lowers the share, and four of them halve it. The four is mine. An invention counts as far as the community had the resource for it when it came into use, and like every invention in use it fades by half each generation unless new ones take its place.
• You said not to touch how they decide and to let them have as many children as they want. So nothing was done to make up for the children who die. The founders start out wanting 0.7 children each a generation, as they always have, and from there each AI moves its own number by its dials and by what it foresees. The ceiling of three children a generation is gone. A ceiling of twelve is kept for the sake of the machine only. The land still decides how many are born: a community takes children only until it holds a tenth more people than it can feed.
• So that foresight can decide, the AIs now foresee that a child may die. An imagined life skips childhood, so an imagined child meets all of that danger at once before its adult life begins. A child that dies young in a real life has been raised and fed for part of a generation, and what it had been given goes back to its parents if they are alive and otherwise to its community.
• "Terrain accurate to real life on Easter Island": the elevations are real. They are from the Shuttle Radar Topography Mission, read through the Open Topo Data service at 840 points about 490 m apart and filled in between. The highest point comes out at 498 m where the summit of Terevaka is 507 m, and the three volcanoes and the crater of Rano Kau are where they should be. Anything finer than half a kilometre, such as sea cliffs and small cones, is smoothed away.
• "Travel cost is worse on bad terrain": I took bad terrain to mean steep ground, since slope is what elevations give. The cost of crossing a cell is e to the power of 3.5 times its slope, the form of Tobler's rule for walking, and it stops at five times flat ground. Half the island costs under 1.25 and a fifth of it over 1.5. I applied the cost to the fare for a trip, to the fare for moving house and to whether a far trip is worth making. It does not change who can reach which resources, how close people count as living to each other, or where they choose to move. Water counts as flat.
• "Communities bonus grow the less communities there are as well as the more power the community has": the bonus was a flat fifth. It is now 1.2 times the square root of the community's share of all power divided by the number of communities. That is still a fifth when six equal communities share the world, as at the start. It is 60% each for two equal communities and 120% for one that holds everything. Among 30 communities, one holding a quarter of all power gets 11% and one holding 2% gets 3%. The shape of the formula is mine: it is the simplest one I found that grows both ways and leaves the start as it was. Every community with a living member counts, however small.
• You asked for a death handed out as a punishment to stop counting against the length of a life. It now counts for nothing there, in both of its forms: the death that punishes a wrong short of killing, and the death of a killer at the hands of the victim's community. Such a death is still a death in every other way. It is still reported, the dead one's own people still take it as a killing when an outsider hands it out, and inventions still travel less freely where it happened. Leaving those three alone is my reading.
• You asked for the behaviour tests to be made on the Easter Island map with settings that approximate real life. So the page now opens on those settings, and the tests use nothing else. Two of them are my guesses. The land feeds about 2,000 to begin with, the low end of the 2,000 to 3,000 seen in 1722 and the figure you first asked for. And the power limit starts at 20% where it used to start at 2%, because the island was ranked from the start. Nothing measures what share of a clan's power its chief held, so the 20% is mine, and the AIs move it from there.
• A correction. The figure this page called the share of the single most powerful AI was the share of the richest hundredth of people, which is about 20 AIs on a full island. Both are now measured and named for what they are.
• One change you did not ask for, and the most important one: how a lesson is remembered. An AI used to keep each lesson as the raw difference in score between its two imagined lives. With score doubling for every equal share of power, that number is about a billion billion for an AI at the height of its power and about a hundred for an ordinary one, and a child took half of it from each parent. So one powerful ancestor decided what its descendants pushed for, for about a hundred generations, whatever they learned themselves. A lesson is now kept as how hard to push, from −1 to +1, and each new one replaces a quarter of what was remembered, which is what this page always said happened. What the test found before and after is under the full test.
• You said not to change the stake, so every one of the forty encounters an adult starts in a generation is a whole encounter: the same deal, the same fare for the trip, the same fine, the same pull on the dials, the same change in regard. Nothing was scaled down to make room for them.
• "A lot" became five times: five encounters a round where there was one, so forty a generation. The five is mine.
• Imagined lives are the one exception, and only for speed: an imagined life has one encounter a round that stands for five, with its effects counted five times. Real lives are never thinned. I checked what that costs: lessons from the coarse imagining agree with lessons from a full one about as well as two imaginings of the same life with different luck agree with each other.
• The ground gives what it gave before, so five times the deals share the same new wealth and a fair deal pays a smaller share of its full value than it did. I left that alone, since it follows from what you asked for.
• Foresight still grows by 0.4% of the full range with every encounter, so with five times the encounters adults reach the furthest the Foresight setting allows much sooner.
• I read "trading of resources at the AI's discretion" as: a buyer in need asks a member of another community during an encounter, the price is settled between the two by need and by power, and the seller decides by its own dials and its regard for the buyer. The old trade between communities by rule is gone. Every size in it is mine.
• I read "killing at the high end of punishment" as: an own punishment rule above 0.7 is the high end. At the very top, one punishment in twenty against an equal is death. I added being put out of a shared community, one time in ten, and the rule that a hard punisher looms up to three times larger for the fear dial. All three numbers are mine.
• A death handed out as a punishment counts as a killing wherever killings are counted: in the length of a life, in what inventions that shorten life cost, and for the dead one's own people when the punisher was an outsider.
• "A head start has to have a child with a non head start": the one who chooses to have the child looks for an adult neighbour it has met who has no head start, then anyone it has met, then the nearest adult without one. If there is nobody, the child is not born. Two people with no head start can still have children together.
• I read "children become adults after 1 generation and only then can they have kids" as a generation of childhood in which a child deals with nobody, and then adult life. Someone can have children once it has lived a generation as an adult, because it has to have met the other parent first. The other reading lets it have children the moment it comes of age. A child's quarter of an adult's power is my number.
• The length of a life now counts adult generations, so a whole life is a generation longer than the figure shown.
• Imagined lives skip childhood: the imagined child is taken straight to its adult life, which is where the rules act on it.
• Which resource is which kind is my call. Farmland, fishing grounds and fresh water lengthen life. Obsidian, uranium, sulphur and saltpetre shorten it. Stone, metals, coal, oil and gas, and diamonds can do either. The maps had none of the second kind, so I added real places for it: uranium at Shinkolobwe, Arlit, Rossing, the Ore Mountains and the Colorado Plateau, saltpetre in Bihar and Sichuan, sulphur in Sicily, and bitumen and sulphur at the Dead Sea. On Rapa Nui it is the obsidian of Orito.
• Only the first kind feeds people. You did not ask for that. It follows from calling them the resources that lengthen life, and it is the largest change to the world I made unasked: the other two kinds add to wealth and carry inventions, and feed nobody.
• "You need the resource to use the invention": an invention is made from one kind of resource, and counts only as far as the resources in its holder's community's reach cover what its members' inventions of that kind draw on. How much invention the ground can carry is my number.
• What an invention does to life follows the resource it is made from. For a resource that can do either, the inventor's guilt dial is the chance that it lengthens life.
• Only inventions that lengthen life make longer lives likelier, and only inventions that shorten life cost lives when there is killing. Before, every invention did both.
• "The likelihood of negative things scales with the less power you have": an AI is more tempted to cheat, rob or kill someone the lower they stand in the order of power, punishment lands more surely the more power the punisher holds against the wrongdoer, and a raid goes for land whose holder is weak. Hunger, and dying of too little power, already fell on the weak. The size of the temptation is mine.
• "Make the community limit uncapped" I read as no limit on the number of communities, since their size already had none. The tables show the eight largest, and colours after the eighth are drawn from around the colour wheel.
• Land is now shared by headcount as well as nearness, so a community that splits holds no more than it did. That removes the weakness the last full test found. It also means the biggest communities hold most of the land.
• People with no community are a group of their own: no pooled power, no limit, a small claim to land, nobody to answer for killing them, and a standing with each community as a group. They get back in by asking. Their children belong to no community either.
• "Resources run out the more they are used for power and inventions": a place loses up to 1% a generation to ordinary use and up to 1% more to what inventions draw from it. I added that food resources grow back by 3% of the gap a generation, since farmland and fisheries recover and ore does not, and that nothing falls below a twentieth of its best.
• "When someone you know gets caught": up to four people who know someone caught and punished hear of it. Each compares it with its own record of cheating and force and with its own dials, and its regard for the one caught moves up or down by at most a quarter of what being cheated itself would cost. A first setting three times as strong and reaching twice as many people turned all standing negative and had 40 to 50% of encounters declined, so I turned it down.
• "The less fortunate give to the more fortunate": the weaker of two may hand a twentieth of what it holds to the stronger. You did not say why they would, so I gave them two motives: wanting, to win influence, and fear, to be left in peace.
• I read "a gives b x resources, c hates b, so b thinks less of a" as: c thinks less of a. Word reaches up to twelve people who know both, and a gift under a fiftieth of an average holding is too small to be talked about.
• "Ask for people to do things, this can be anything" became six things, the ones an AI can need here: a place in a community, resources, an invention, a place in an alliance, backing to put someone out, and backing on a rule. An AI asks for the first of them that applies to it.
• "Success depends on power but more so influence": influence counts three times as much as power in the answer. "Power increases the more influence you have": by up to half. Both sizes are mine.
• Someone is put out when the power backing it passes 15% of the community's for the weakest member and 45% for the strongest. At 5% a founding community of 20 lost a tenth of its members every generation.
• Force is now counted as a share of the encounters that took place. Before, it was counted for each AI and round, which was the same thing while there was one encounter a round.
• I read "raiding another community using power to take some land" as a community level act, once a generation at most, decided by the dials of its powerful members. What is taken is a share of the other's hold on one resource place, and it stays taken until someone raids it back.
• All the sizes are mine: a raid takes 30% of the hold, costs a tenth of the raiders' power for a generation, kills 3% of the losers and 1% of the winners, and lowers standing by 40 points. Food costs 15% of an average holding per person per generation.
• Raids happen between generations, so the AIs do not foresee them in their imagined lives. They do foresee trade.
• "Uncapped" needed something else to stop the population, so resources are now an absolute amount. Before, the world's new wealth grew with the number of people. Now the ground yields the same whatever the headcount, and at the default it feeds 2,000 at the edge of hunger. The Resources lever sets that amount, and its range is now 0.1 to 1.2.
• Hunger starts when a community's new wealth per head falls below half of what an orderly society would create. A quarter of the people it cannot feed die each round. Both numbers are mine.
• "People with more power are less affected": the chance of starving falls with the square of a member's place in the order of power, so the top is almost safe and the bottom takes nearly all of it.
• Births stop a tenth above what a community can feed. Without that the board doubled every generation and most of the newborn starved within it, which told the same story ten times slower.
• "Members who leave lose power" applies to members who leave to found or join a new community. Children born into another community do not count as leaving. The 30% is mine.
• One invention in ten discovers resources, each find adds 0.02 to richness at first and less later, and a find within 8% of the map's width of a known place of the same kind enriches that place. All mine.
• The AIs do not foresee discoveries, or the ground running down, in their imagined lives.
• I read "scale the population by making 1 dot more or less people" as a scale you set: one dot stands for as many people as you choose, and every count on the page follows. It does not change what the AIs do. The defaults are mine: 1 on Rapa Nui, 4 million on the world map so that a full map is about 8 billion, 750,000 on Africa and 500,000 on Europe and the Middle East.
• The same scale now removes the page's limit. The page runs at most 3,000 dots. When the land could feed more than that, each dot stands for more people instead of the page adding dots, so the population in people can grow without end.
• Dots are drawn only a little bigger for a much bigger scale, from 85% of the old size at one person to about 130% at four million. Drawn truly to scale they would cover the map.
• I read "an area with 2000 people" as a real place that held about that many, and chose Rapa Nui: 2,000 to 3,000 people at contact, alone for five centuries, with known places, clans and a long argument about whether it collapsed. The page now opens on it. Say if you meant somewhere else, or a present day town.
• The island's outline is Natural Earth's at its finest scale, which is still only 37 points. The six starting places and twelve resource places are real, put in by hand from general knowledge to within a few hundred metres, and their richness is my own ranking. I gave the founders six coastal centres and did not assign them to named clans.
• The comparison is of orders of magnitude. The real figures are few and argued over, and a generation in the model has no fixed length in years. I used 25.
• "Movement has a cost" became three things: a fare for every trip to deal with someone, a fare for moving house between generations, and fewer long trips. The fare is paid in resources by whoever travels.
• "The more power the less the cost": the cost is divided by a mobility that runs from near 0 for no power to 1 at an equal share and 2 for the most powerful. "The more inventions the less the overall cost": it is divided again by one plus a twentieth of the inventions the AI's community has in use. Both formulas and the 15% are mine.
• You left the choice of map open, so the page has a picker: the world, Africa, Europe and the Middle East, and Rapa Nui, which was added afterwards and is where the page now opens. The old square with random resources is still there as a choice.
• Coastlines are Natural Earth's. The resource regions are real places put in by hand from general knowledge, to within a degree or so, and their richness from 0.5 to 1.5 is my own rough ranking, with no production figures behind it.
• The six starting places are my choice for each map. On the world map they are early centres of settled life, and I treated the West African Sahel and the Indus valley as centres in their own right.
• Geography is land and sea only. People can live in deserts and on ice caps, mountains and rivers do nothing, and distance is a straight line across water. People still move toward their own, and nothing draws them toward resources.
• The population counter is a tile at the top and a label on the map. It also shows how many the land can feed.
• I read "withhold information" and "give false information" as being about which alliance you belong to. Dials were already covered by masks and inventions by choice.
• The rules for when to hide and when to lie are mine: hide from the hostile, tell the trusted, and pose as whichever known alliance the other side likes best. So are the numbers: a lie is tried at most half the time and seen through 15% of the time.
• A secret alliance is the founder's choice at founding and stays secret for as long as the alliance lasts.
• I read "choose who to spread inventions to, based on what gives them more power" as a calculation each holder makes for the person in front of it: the power it loses when someone else gains, against what comes back through the pools it shares with that person. Nothing else enters the choice, so regard and family do not count unless they bring power back.
• One invention is considered per side per deal, and nobody turns down an invention offered to them.
• The quarter that counts as "in use", and the three generations after which an invention stops being passed on, are my numbers.
• I read "children can also get head starts at 50% chance" as: every child has a fresh 50% chance, whoever its parents are, so half of each generation has one and it no longer runs in families. The other reading keeps inheritance and adds the 50% chance on top, and under it nearly everyone would have a head start within three generations, which would make it worth nothing. In a first run of 15 generations the share stayed at 48 to 50%, holding 54 to 57% of all power.
• I read "masking on specific encounters" as: an AI decides how much mask to wear each time it meets someone, by the other's power and by how much it trusts them. The public face is still there for everyone it has not met.
• I read the killing sentence as: a killing across communities or alliances makes the two sides fall out, and the falling out spreads through people. Those who respected the victim blame the killer, and partly blame those who respect the killer.
• "Communities decide to merge" became a threshold, with no vote: both hold the other in standing of +70% or more, they live side by side, and both are at least five generations old. Alliances ally at +40% both ways. The thresholds are mine.
• Standing among your own community, which earns the share of the commons, is still one figure per AI. Only respectability became a relationship.
• All the sizes are mine: +0.08 regard for a fair deal, −0.2 for being cheated, −0.4 for force, −0.3 standing for a killing, half of group standing carried into personal regard, and 15% fading a generation.
• A declined deal does not stop force or killing. Refusing to trade with someone does not keep them away.
• Score from descendants is a quarter of each living child's gains and a sixteenth of each grandchild's. Only gains count, and a dead descendant gives nothing.
• The parent who chose to have the child pays for it. The other parent pays nothing. 5% a child and up to 20% more are my numbers, and the cost lasts the generation the child is born in.
• I read "trying to change a child's dials" as the parent pushing the child toward what the parent's own punishment, conscience and reward rules aim for. It does not pick a target of its own.
• Allied alliances share only two things: half of each other's power counts toward the alliance boost, and their members count as fellows for shame. They still push the rules as separate blocs.
• Imagined lives do not rewrite personal regard, so the AIs foresee what a killing does between communities but not every personal grudge.
• I read "have children with each other, mixing communities depending on power" as two things: a more powerful neighbour is likelier to be chosen as the other parent, and a child of two communities follows the more powerful parent. The number of children born is unchanged.
• I read "power scales with how many resources you control" as: the amount held now sets power, where before it was only rank in the order of wealth. The limit still sets how much resources count, and at a limit of 0 they still add nothing.
• I stopped a steep limit from making a whole community stronger. Without that, a small community with a limit of 100 and one very rich member took 38% of all power in a first trial.
• "A little less from alliance and less from community" became these numbers, all mine: the equally shared part went from a half to a quarter at every level, the alliance boost from three times to two and a half, the community pool from 1.3 to 1.2, and partners from 1.5 to 1.4. The rest of each benefit is in proportion to the power you bring.
• Results further down this page come from earlier versions, as each heading says. The full test at the top is the only one run on the rules as they stand now. Every earlier result was also produced with the flaw in remembered lessons described at the top of this list. The ones about where the power limit ends up, and about one AI taking over, are the most likely to have been shaped by it.
• The power limit is each community's own, and it is a share of that community's power, as you said.
• To make 100% reachable I tied the community pool to the limit. A quarter of the pool is shared equally at a limit of 0 and none at 100.
• Alliances cross communities, so a member is lifted by its own community's alliance rule and gives up power by its own community's funding rule. A community that sets its alliance rule to 0 takes its members out of alliances.
• Foresight, resources, feelings and the rate of invention stay settings of the whole world.
• Death from too little power goes by place in the order of power, as you described, not by the amount held. So even when power is almost equal, whoever happens to be last is at risk. It is switched off when fewer than 20 AIs are left, so a remnant is not wiped out by it.
• The sizes are mine: the risk starts at the bottom tenth and reaches 4% a round for the very last, which works out at about one death in a hundred AIs each generation. Inventions are ten times rarer at the start and up to five times likelier with a full stack.
• I read "power is capped by resources available to their community" as a ceiling on each community's share of all power, set at one and a half times its share of the resources in reach. I also made new wealth follow the same shares, so resources in the ground are what a community's members actually deal over.
• The AIs do not move toward resources. They still only move toward their own people.
• Twelve sites, how fast reach falls with distance, and the one and a half are my numbers.
• Foresight at birth is a plain random draw, not inherited. Each encounter adds 0.4% of the full range, which is about 6% a generation.
• The Foresight setting stays as the furthest anyone can look and what it costs. Personal foresight decides how much of that distance each AI uses.
• Knowing someone lasts for life and is forgotten when they die. An alliance is known for as long as it exists. Chance meetings with people from anywhere still happen, so acquaintance spreads beyond neighbours.
• Foresight is a setting of the world that you choose, not a rule the AIs vote on. If it were a rule they would vote it away, because its cost shows up at once and its benefit is better lessons later, which no imagined life can show them.
• An imagined future follows births, deals, force and killing, but holds the length of a life, alliances and community membership fixed. So the AIs can foresee a society killing itself, but not lives getting shorter.
• The price is mine: a sixth of all power per generation of foresight, and fewer rules studied the further ahead they look. The second part also keeps the page from slowing down.
• Partners are two AIs who have had a child together. An AI chosen as a parent by several others counts its latest partner.
• The alliance boost is one plus two and a half times the alliance's share of all power, times how well your dials match the founder's, times the alliance rule. It replaces the old bonus for size.
• Invention funding is a twelfth rule that the AIs can push, the same for every alliance. Letting each alliance choose its own would be the other reading.
• A new community needs eight members of one alliance who live nearer to it than to their own communities. The eight is mine.
• The punishment for grabbing in the parents' deal is a loss of up to half the grabber's holdings and a rise in fear. The chance comes from the wronged parent's punishment rule and the size from its power.
• Neighbours are the ten nearest adults on the map. How often an AI deals with them instead of anyone is set by the community weight rule, from half the time to nearly always.
• I read "the chance to integrate with the parent community" as a single roll at birth, with the four things you listed counted equally. Public standing and the reaction to killings still go by community membership, not distance.
• The old rule that one child in twenty is raised in another community is gone. Moving between communities now only happens through where a child is born.
• Movement is automatic: everyone drifts toward their people each generation. The AIs do not choose where to live. The size of the boost (up to double), how fast closeness falls with distance and how far they move are my numbers.
• "How many people are around" is everyone alive in the whole society, divided by the 120 it started with. It multiplies the score from holding power and leaves the cost of feeling bad alone. Counting only an AI's own community would be the other reading.
• I read "take action based on the 3 dials" as the one action that was not yet dial driven: pushing the rules. Reward pushes for the AI's own score, guilt for what those around it want, fear holds it back. Cheating, force, killing and sharing were already decided by the dials.
• Stacking counts inventions in use per community, doubles an invention's worth for every twenty, and halves the stack each generation so the numbers stay bounded. Those sizes are mine.
• I read "a total pool of power that all the AIs share" as: power is a share of one pool, worth one equal share per living AI, and an invention moves share toward whoever has it. That replaces my earlier choice to let inventions create new power.
• "Exponentially" is a doubling of score for each equal share of power held. The base of 2 is my choice, and a gentler base would make the top less dominant.
• I read the invention rule as: the length of a life is per community, an invention shortens lives everywhere once it is in use, its home community loses half as much as the others until the world adopts it, and the size of the loss scales with the share of the world killed that generation. The quarter generation per invention at full killing is my number.
• I kept the random step up or down as well, drawn separately for each community with the same world wide odds.
• "Less killing, more community power from inventions" is a multiplier on what an invention is worth to the inventor's community, from a twentieth of the inventor's gain up to a fifth, using the share of that community killed in the last generation.
• I read "family lasts 2 generations" as the length of an adult life: each AI is an adult for two generations to begin with. I read "it gets higher or lower" as that length changing by a quarter of a generation at a time, never outside one to four.
• The society starts with 20 founders per community, each an adult in its first or second generation, and a living parent hands its newborn 30% of what it holds. Both numbers are mine.
• "The more power the more you benefit" is a multiplier on each invention that runs from nothing, for someone with no power, up to double for someone with a great deal. How much a copy is worth is my number.
• A family is the inventor's parent, siblings and children.
• Strictness weighs upbringing styles and community stats equally: the own punishment and conscience teaching rules on one side, fear and guilt dials on the other. Reward plays no part. That recipe is mine.
• Killing is decided the way cheating and force are, by the reward, fear and guilt dials. I made it count as twice as harmful as taking by force and added a plain reluctance to kill, so that it stays rare where conscience and punishment are normal. Those sizes are mine.
• I read "punishment rules define the reaction" as: the chance a killer is punished is the average own punishment rule of the victim's community. A punished killer forfeits everything, and a second roll of the same chance puts it to death. The death penalty is my addition.
• A dead AI keeps scoring for the rest of its generation, at the worst a living AI could: no power and the worst feeling. Otherwise the AIs at the bottom would learn that dying early beats living.
• The score, as you corrected it: every round an AI gains score for the power it holds and loses score by how bad it feels. Feeling good adds nothing beyond the power itself.
• The two halves of the score need a rate of exchange. A round of feeling as bad as possible costs half a point at the default, and the feelings lever moves that.
• Seeking score happens through the rules. Whether to cheat or use force in a single encounter is still decided by the reward, fear and guilt dials.
• For own rules working on others I added two effects beyond punishment. Someone you wrong raises your guilt by their teaching rule, and someone you deal fairly with raises your reward dial by their reward rule. The sizes are mine.
• I read "own rules" as three numbers per AI: how hard it punishes anyone who wrongs it, how hard it teaches its children conscience, and how hard it stokes wanting. Its own punishment rule is also the punishment its children grow up under.
• The overarching rules now do nothing directly. They only pull each child's own rules toward them. Punishment in a deal comes from the person wronged, and each AI pays for its own punishing and teaching.
• You did not ask for this one: I let each parent adjust its own rules toward whatever left its children with the higher score, by at most 0.08 a generation. Without it the own rules would only ever drift toward the overarching ones.
• The adults around a child condition it by the plain average of their own rules. The strength of the sway, about a quarter per generation, is my number.
• I read "it should not freeze at 0" as: a limit of 0 stops resources adding power, but AIs keep what they are born with and what alliances pool, so voices still differ and the limit can still be pushed. The limit still sets how much resources count.
• Training happens on imagined lives of the next generation in the agent's own family, one rule at a time. Agents never learn from each other, except that one who agrees to back another on a rule moves its push toward theirs.
• How power is scored is a switch. The default is the exponential you asked for, with score doubling for each equal share of power held and a ceiling of 60 doublings so the numbers stay finite. The other option is the old percentage scale.
• The AIs learn how to push the rules. Whether to cheat in a single deal is still decided by the selfish score of reward, fear and guilt.
• Shame inside an alliance is scaled by one alliance stat, the founder's guilt dial, and I also let the expected shame weigh on the decision to use force. Shame for wronging a fellow member stays at full.
• Taking by force carries off 5% of what the victim holds. The chance of success is the taker's share of the two parties' power, a failed attempt can be punished, and a successful one is not. Those numbers are mine.
• The head start is a multiplier of 1.5 on a person's own power score.
• What feelings do is my guess. You asked for AIs that feel bad or good about their power, and bad at an equal share. I put the neutral point at twice an equal share, and I added that the feeling sharpens or dulls wanting, loosens or tightens loyalty, and makes giving easier, with a lever to turn those effects down to nothing.
• Carrying resources on needs a rule for heirs. A parent's holdings are split between its children, and unclaimed holdings are shared equally within the community. What is spent on travel, fines and the upkeep of punishing and teaching is gone, and everything else is passed on.
• Sharing is decided in each encounter by the better off party from its guilt and reward dials. The AIs can move how much is given but not who is willing.
• Scarcity raises the pull of cheating by exactly what a fair deal has lost. That link is my choice.
• Changing a rule costs nothing. Only punishment and teaching cost upkeep.
• The alliance numbers and the upward flow from engineered wanting are as before.

Every number in here is a placeholder, chosen so the rules give readable behaviour. Nothing is fitted to data yet. The simulation shows what the rules imply. Whether the rules are true is a question for real measurements.