In an end‑to‑end game, a sender and a receiver learn a discrete protocol and routing policy without any communication supervision. Across two independent batches of 20 seeds each, protocols formed in 20/20 and 19/20 runs respectively. Online partial coverage increased the share of shadow routes over no monitoring by $0.701\;(95\%\;CI\;[0.574,0.826])$ and $0.575\;(CI\;[0.412,0.737])$, while task accuracy stayed at $0.905$ and $0.913$, and the signed frozen‑message contribution remained $0.592$ and $0.563$. Full coverage removed most selective migration but did not eliminate communication. Frozen interventions, including a fresh pre‑outcome confirmation, showed that routing is semantic: shuffling joint message‑route pairs altered $66.8\%$ of receiver predictions, and routing alone retained $38.8\%$ accuracy versus a $25\%$ chance level. A corrected explicit‑silence cost made $18/20$ policies fully silent, lowering mean sending to $0.050$, yet the strict complete‑suppression gate failed because two seeds kept communicating. Failed semantic‑evasion and operational gates further bound the result. Overall evidence supports adaptive reallocation of joint message‑route codes under partial monitoring, rather than open‑world route discovery, emergent encryption, or a universal defense.
Review: The study provides a rigorous experimental demonstration of how monitoring subtly reshapes communication strategies, offering valuable insight into adaptive information encoding.