PETRI is a digital Petri dish and visual proving ground inspired by John Calhoun's classic Universe 25 experiment. It provides a closed medium for observing how complex social behavior, hierarchies, and population cycles emerge from simple, explicit personality traits (aggression, intelligence, sociability, fear) under spatial and resource constraints.
Conceived as an in-silico Petri dish for artificial life and social physics, it builds on Calhoun's crowding research in a strictly controlled digital environment. Unlike the original mouse study with its biological assumptions, everything in this dish is completely transparent: every trait, metabolic rate, and social rule has explicit, observable, and tunable effects. Agents are currently driven by heuristic (rule-based) AI—a deliberate choice for the first phase to establish honest, reproducible, and efficient baseline dynamics before layering on cognitively rich LLM-agents.
A 2D bounded space (canvas) acting as a Petri dish with fixed surface area. Just like in Calhoun's closed mouse habitat, the dish dimensions do not expand with the population, creating inevitable crowding pressure, territorial competition, and density stress.
Each agent is a node defined by:
Goals are not set manually; they are derived from the agent's dominant trait plus random noise. This results in four emergent behavioral archetypes: Dominance (Aggression), Bonding (Sociability), Accumulation (Intelligence), and Survival (Fear). A goal sets priorities but doesn't block other actions—an aggressive agent can still pair up if the opportunity arises.
Directly modeling Calhoun's "behavioral sink": as colony density increases within local clusters, agents suffer from sensory overload and stress. This manifests as elevated energy drain, heightened aggression, withdrawal of social bonding, and reproductive decline. In Calhoun mode, the dish exhibits distinct phases and archetypes: dominant territory defenders, hyper-aggressive outcasts, maternal neglect, and asexual withdrawn entities ("the beautiful ones").
Heuristic simulation: fast (hundreds of ticks/sec in-browser), fully tunable, visually readable. Serves as the calibrated "physical substrate"—without it, any LLM layer on top would have no baseline for evaluating rational behavior.
Keeping survival mechanics heuristic (cheap/fast), LLMs are added where they provide quality, not just cost:
Targeted model calls for specific crossroads: birth-goals based on persona and family history; trade outcomes influenced by generated dialogue. Requires event-driven triggering (every N ticks or significant encounters) to maintain performance.
Standard agents remain heuristic (cheap, scalable to hundreds), while "heads of colonies" or leaders receive a full LLM-loop for strategic decisions: territorial expansion, war, alliances. The LLM manages a colony leader, not every single microscopic entity.
Headless version for batch runs with fixed seeds; systematic collection of metrics (Gini coefficient, strategy diversity over time, conflict frequency vs. density) for quantitative analysis. A/B testing of heuristic vs. LLM versions under identical conditions to verify if the LLM layer provides genuine emergent value.
Core Principle: If behavior is explicitly hardcoded into the prompt, it’s not a discovery, but a confirmation of design. Value emerges only when drives are minimal, but strategy and social dynamics are not.
PETRI is a community support token for PETRI — a way to back the project and help spread the word, not an investment.
A deflationary mechanic burns 1% of a pre-locked token reserve for every $20,000 in trading volume, permanently reducing total supply as activity grows. PETRI carries the risks typical of any speculative crypto asset.
ca: bpu74XUapF6hn1UYYKHY4TMv9RcDb87H7AZ9889pump