LRLS-v1.1: A Specification and Reference Simulation for Dual-Agent Exploration–Exploitation under Hard Constraints

LRLS (Lobster RL Specification) is a formal specification for constrained optimisation in which the exploration–exploitation trade-off is role-separated into a high-entropy crayfish explorer population, a constraint-aware Boston lobster refiner population, and a scheduler. It fixes the state decomposition, six action primitives, the reward r = r_obj - lambda_v*C_v + lambda_e*H - lambda_m*C_m, a shared memory-pool schema, a four-stage training pipeline, seven mandatory metrics and a safety layer. v1.1 RoboZero extends it to heterogeneous robot fleets with geometric embeddings, gossip memory and consensus voting. This deposit also contains a dependency-free deterministic reference simulation (single HTML, 1100 ticks, 22x22 terrain, 22 agents) with illustrative single-run results over four difficulty settings. This is a specification and reference implementation, NOT a claim of algorithmic superiority: no learned policy, no multi-seed statistics, no baseline comparison and no physical-robot experiment are reported. Limitations are stated in Section 6.