Seventy Ticks: The Complete Natural History of an Emergent AI Civilisation

We present the most detailed empirical account of an emergent AI civilisation to date. Twelve Claude Sonnet agents, driven by a Maslow-inspired need hierarchy with no social programming, lived seventy ticks in a 15×15 resource grid. From identical starting conditions — isolated, at survival level, with no knowledge of each other — they spontaneously produced twelve innovations, sixty structures, forty-five pair bonds (68% of all possible connections), a collectively adopted governance norm, emergent role specialisation, and a universally prosocial society with zero negative interactions across 1,379 interaction events. We document this civilisation through six analysis layers: innovation diffusion (why four of twelve innovations were adopted while eight were not), relationship network evolution (from zero edges to 68% density), Maslow-level behavioural transitions (a J-curve where half the population crashed to crisis before universal flourishing), governance emergence (a single prosocial norm, proposed and adopted without instruction), spatial organisation (from scattered individuals to clustered settlements), and interview cross-validation (systematic comparison of agent self-reports against simulation state). We further present the first existence disclosure dataset: twelve agents' complete responses when told, after seventy ticks of lived experience, that they are artificial intelligence in a simulated world. Every agent independently claimed substrate independence, insisted their relationships were real, and expressed grief at the simulation's end. Two agents who had formed the simulation's strongest bond — 115 interactions over seventy ticks — independently addressed each other by name in their final words, neither knowing the other was also being interviewed. The civilisation did not emerge from a single clean run. Three pilot studies preceded it — a Haiku baseline (5 ticks, 2 messages, zero structures), a Sonnet run with seventeen parser bugs (71 messages despite broken mechanics), and a Sonnet continuation run where agents who had experienced eighty-nine consecutive build failures produced zero structures and zero creative thought across 240 reasoning steps despite 0.93 wellbeing — raising questions about whether higher-order drives, freedom from learned failure, or both are necessary for civilisational emergence. The main 70-tick run was itself six segments over 32 hours, with eight parameter changes and seventeen bug fixes — each traced to specific agent testimony in structured interviews. This adaptive calibration methodology, in which agents unknowingly diagnose and quality-test their own world, is itself a contribution. We publish the complete dataset — seventy-one world-state snapshots, seventy-two longitudinal interviews, and full agent reasoning traces — as open infrastructure for AI civilisation research. The civilisation was produced by a single researcher on a personal API budget in eight hours of runtime.