Receiver-Anchored Tests for Latent Communication: Constant-Capacity Causal Localization and a Negative Cross-Model Result

Abstract

This preprint presents a receiver-anchored evaluation protocol for latent communication between language models. It separates receiver-carrier viability from learned cross-model transport using receiver-native oracle replay, constant-capacity matched-versus-shuffled identity controls, sealed task-level functional confirmation, and hardened code execution. In Study A, receiver-native residual packets produced 79/96 functional passes for the terminal matched packet versus 0/96 for the same-capacity shuffled control. A registered 2 x 2 x 2 factorial localized positive contributions to the terminal instruction core and function-name states under the tested Llama-3-8B-Instruct receiver. In Study B, the receiver-native oracle anchor again succeeded (84/96 matched versus 0/96 shuffled), but the registered DeepSeek-Coder-to-Llama learned bridge produced 1/288 functional passes in both matched and shuffled conditions. The result is therefore negative for this complete registered cross-model system despite successful held-out geometric gates. It does not establish that heterogeneous latent communication is impossible. The manuscript's priority claim is limited to the documented LIP evaluation protocol, its constant-capacity causal localization experiment, and its receiver-anchored negative cross-model finding. Source code, protocol specifications, claim ledger, and machine-readable result snapshot are linked through the public repository. Repository: https://github.com/zigfreud/latent-llm-agent-communicationCorrespondence: ziwehdafe@gmail.com