On a fragile commons where a single aggressive agent suffices to trigger collapse (BDPD⁰), can communication and sanctioning restore sustainability? We add a pact registry, a cheap-talk channel with six tool primitives, five governance metrics, and both flat and graduated sanction perturbations to the BDPD platform, then run four mini-pilots at N=5 seeds with DeepSeek-flash LLM agents at the canonical cliff configuration. A 2×2 factorial decomposition reveals that the dominant cooperation driver is the transition from heuristic to LLM agents, not cheap talk. Graduated sanctioning provides a robust positive signal that communication alone does not: a ladder 1,3,10 preserves the commons in 5/5 seeds versus 3/5 for a flat sanction of equal expected cost. A ladder-geometry sweep isolates escalation shape as the operative mechanism, consistent with Ostrom's design principles #4 (monitoring) and #5 (graduated sanctions).
Roberto Brunelli (Tue,) studied this question.