BDPD⁰ — A Computational Laboratory for Generative Agents, Commons Dilemmas and More
Abstract
Common-pool resource dilemmas generate predictable collective failures — overextraction, wealth concentration, and irreversible collapse — whose integrated dynamics under agent heterogeneity and information asymmetry remain poorly understood. We present BDPD (Be Different Play Differential), an open-source, dual-mode computational laboratory comprising a configurable multi-agent simulation platform and a card game, The Forest of Humbaba, designed as an experimental and didactic testbed for the integrated study of these phenomena. Through systematic parameter sweeps (5–200 runs per cell depending on outcome variance), we demonstrate four principal findings.
Discontinuous fragility threshold. In the slow-regeneration regime (logistic rate \(r \le 0.32\), representative of slow-growth common-pool resources such as commercial fisheries and forests (Li et al. 2026)), the collapse threshold exhibits a discontinuous step: a single aggressive agent suffices to doom the commons regardless of pool size, initial endowments, or turn order — even at extraction intensities (fraction \(i\) of visible stock requested per turn) an order of magnitude below the canonical parameterisation (gate fails just above \(i \approx 0.05\); cliff-edge 95% bootstrap CI [0.04, 0.05]).
Reactive inferiority. Locally reactive strategies are collectively inferior to unconditionally conservative ones (the reactive < conservative effect): by reducing harvest in response to falling stock, reactive agents create an exploitable vacuum that aggressive actors fill — a strategic externality that scales monotonically with the strength of the reactive response, across the tested parameter range.
Noise as a selective blunter. Increasing observation noise on the commons stock improves collective welfare in the continuous platform in two distinct regimes: modestly across the noisy range and with a discontinuous further gain under full occlusion. The mechanism is asymmetric: noise degrades the aggressive agent’s fine-grained stock estimates and dampens extraction, while conservative agents — whose strategy does not condition on the stock signal — are unaffected. The card game shows zero effect, by construction: random card selection renders information fidelity strategically irrelevant. The divergence is itself the finding: noise matters only when agents condition on the stock signal.
Safe betrayal and the Tragedy of the Compensator. Cooperative equilibria are structurally fragile to late defection in the Arena; two-player card matchups CT4–CT5 together characterise a further pathology — safe betrayal — where a defector profits from a conservator’s healing (gaining up to 27% more wealth than its no-defection baseline) without triggering collapse. This tragedy of the compensator — unilateral conservation subsidising defection rather than preventing it, in the absence of sanctioning or exclusion — maps directly onto the dilemma of early movers in climate negotiations.
These four findings are offered not as isolated results but as demonstrations of the platform’s capacity to generate coherent, reproducible, and structurally contrasted scientific insights across two complementary substrates. Preliminary single-model LLM case studies provide qualitative corroboration: safe betrayal emerges spontaneously, and the agent exhibits an awareness-without-restraint pattern (a label we adopt for the case in which the agent’s chain-of-thought verbalises imminent collapse yet its card selection continues to extract). The framework itself was developed through a structured human-AI collaborative methodology — frontier LLMs orchestrated as a synthetic research team under the author’s direction — serving as an empirical demonstration that such collaboration can produce coherent, verifiable scientific contributions1.
Keywords: tragedy of the commons, agent-based modelling, wealth inequality, generative agents, common-pool resources, safe betrayal.
Transparency and authorship statement. This project was conceived, directed, and validated by the author. A pool of frontier large language models (including DeepSeek (DeepSeek-AI 2026), Anthropic’s Claude (Anthropic 2026), Qwen (Qwen Team 2026), and Z.ai GLM-5.1 (GLM-5-Team et al. 2026)) was employed as a de facto research team — not as conventional writing tools, but as active artificial collaborators — to refine the original idea, explore the design space, implement and iterate on the codebase, draft and revise the manuscript, and stress-test the arguments. Model outputs were source-grounded by prompt, validated by extensive triaging with different models and finally integrated under author supervision.↩︎