1  Introduction

This paper adds communication and sanctioning to the BDPD platform and tests whether they change the collapse trajectory on a fragile commons.

The BDPD platform is an agent-based simulation environment for common-pool resource (CPR) governance. Agents harvest from a shared commons governed by logistic regeneration; when cumulative extraction exceeds a threshold, the commons collapses irreversibly. The platform exposes perturbation surfaces — sanctions, exclusions, capacity shocks — that let researchers inject institutional mechanisms without modifying the underlying engine.

1.1 From baseline fragility to governance

BDPD\(^0\) (Brunelli 2026) characterised the BDPD commons as discontinuously fragile: in the empirically realistic slow-regeneration band (\(r \le 0.32\), the canonical regime for commercial fisheries and slow-growth forests) the collapse threshold collapses to a step function in the number of aggressive participants. A single aggressor1 suffices, irrespective of pool size or extraction intensity. Two further results — reactive < conservative and safe betrayal — sharpened the picture by showing that purely individual-strategy counter-measures are dominated, often perversely, by the aggressor. The baseline platform was, by design, an ungoverned commons: agents share neither speech nor enforcement.

The natural follow-up question is the second column of Ostrom’s design-principles table (Ostrom 1990; Ostrom et al. 1992): what does the platform say once we do let agents talk, declare intentions, propose pacts, and sanction violators? This paper answers in the narrowest, most empirically defensible form: we add a pact registry, a cheap-talk channel with six tool primitives, five governance metrics, and a sanction perturbation surface (with both flat and graduated variants), then run four mini-pilots aimed at mechanism verification rather than scientific campaign-scale claims.

1.2 The Farrell-Rabin test

The Farrell-Rabin coordination literature (Farrell and Rabin 1996; Crawford 1998) posits that cheap talk — non-binding, costless, asynchronous messaging — can be a Pareto-improving coordination device in games with multiple rankable equilibria: when players could all benefit from coordinating on a better outcome but need a shared signal to align on it, even unenforceable pre-play communication can select the collectively preferred equilibrium. On a fragile commons, however, this precondition fails. The unique sustainable equilibrium sits below the privately-optimal short-run harvest, and unilateral defection is strictly improving for the defector at small horizons — there is no alternative equilibrium that cheap talk could select. Farrell-Rabin therefore makes a positive prediction only when the underlying game has the right structure; the BDPD cliff configuration does not, and theory leaves room for cheap talk to be empirically null. A null on this substrate would nonetheless be a striking departure from the robust meta-analytic finding that communication substantially increases cooperation in social dilemmas (Cohen’s \(d = 1.01\) across forty-five studies, \(d = 0.46\) for written messages (Balliet 2010)); recent surveys of communication in games (Agranov 2025) identify exactly the structural prerequisites — multiple rankable equilibria — that gate the cheap-talk effect, and whose absence motivates the present test. We test it anyway, because an empirical null on this substrate is itself informative: it tells us whether talk, absent any enforcement, does anything at all — or whether sustainable cooperation on a fragile commons must come from elsewhere.

This paper’s headline experiment (D1, see Section 3.2) is the operational form of that test. An incomplete \(2 \times 2\) factorial (one cell — built-in \(+\) talk — structurally empty) — agent architecture (heuristic vs LLM) \(\times\) communication channel (off vs on) — decomposes the two effects that the original two-cell design confounded. Using LLMs as simulated economic agents follows the Homo silicus framework (Horton et al. 2026) within the broader machine behaviour programme (Rahwan et al. 2019) that studies machines as empirical objects of behavioural inquiry rather than purely as engineered artefacts. It builds on the foundational demonstration that generative agents can act as believable proxies of human behaviour in interactive sandboxes (Park et al. 2023). This class of model is a complement, not a replacement, for the heuristic agents that dominate the CPR agent-based modelling literature (Li et al. 2026).

1.3 Scope and contribution

This paper has four contributions.

  1. A minimal governance layer. We describe an engine-agnostic communication-plus-sanctioning surface that sits entirely in the platform layer (platform/pacts.js, platform/messages.js, platform/perturbation.js) without touching the per-engine dynamics in engine/<kind>.js. Adding a new engine remains a one-file change.

  2. Five governance metrics. Five metric primitives are wired into the existing METRIC_REGISTRY, with no preliminary refactor: cooperation_index, lie_score, announce_frequency, silent_defection, sanction_rate. The metrics are designed to be diagnostic individually (one bounded \([0, 1]\) signal each) and complementary in combination.

  3. An architecture–communication decomposition at the cliff. A \(2 \times 2\) factorial (D1, Section 3.2) shows that the dominant cooperation driver is agent architecture, not communication. LLM agents without talk preserve the commons in \(4/5\) seeds (\(0/5\) for the heuristic baseline); adding the cheap-talk surface produces no additional preservation benefit (Cohen’s \(d = 0.17\)). Talk does redistribute wealth from the defector (\(-33\%\)) to conservators (\(+17\%\)), acting as an equity instrument rather than a preservation one.

  4. A positive Ostrom #5 finding at the mini-pilot scale, robustified by a ladder-geometry control. A graduated sanction ladder \([1, 3, 10]\) preserves the commons in \(5/5\) LLM-driven seeds; a flat sanction \(3\) preserves it in \(3/5\). A subsequent ladder-shape sweep at \(N = 5\) adds a critical control: a flat-5 schedule (constant amount roughly halfway between the canonical ladder’s extremes, \(1\) and \(10\)) preserves the commons in only \(1/5\), while three graduated schedules — \([1, 2, 4]\), \([1, 3, 10]\), \([1, 5, 25]\) — preserve it in \(\geq 4/5\). Constant amount is not the operative mechanism; escalation shape is.

Chapter 2 describes the platform surface and metrics. Chapter 3 reports the four mini-pilots and the three headline figures. Chapter 4 decomposes the architecture and communication effects, locates the graduated-sanction result against the Ostrom literature, and argues that on a fragile commons a graduated enforcement layer is the load-bearing governance component — a claim a campaign-scale follow-up at \(N \geq 30\) will be positioned to test.

1.4 Caveats

The four pilots reported here are mini-pilots at \(N = 5\) seeds each — C1 (mechanism verification) and D1, D2, D3 (small-sample replication), all on a single LLM (deepseek-v4-flash, no thinking, temperature 0.4). We isolate a single model to control for model-specific priors; our focus is on the class-level difference between heuristic and LLM agents, not on inter-model variance. Cross-model validation (Claude, Qwen, GLM) and bootstrap confidence intervals at conventional power are deferred to a campaign-scale follow-up. The \(N = 5\) samples are large enough to show that single-seed narratives can be misleading — both D1 and D2 produced narratives at \(n = 1\) that did not survive replication — and small enough that individual-level claims (per-archetype wealth, learning events, social transmission) cannot be reliably extracted.

  1. The two agent roles recur under several names. The aggressive archetype is also called the aggressor, the defector (its behavioural-game label), a pact violator, or the mule (its BDPD\(^0\) name); the conservative archetype is also called a conservator (the role label used in the tables and figures). Each set of synonyms denotes the same role throughout.↩︎