Listening–Trust Asymmetry and Team Outcomes in a Multi-Agent LLM Substrate: Three-Layer Decoupling in a Mandarin Lobster Observatory

We report a three-layer structural-network observation from a continuously-running multi-agent LLM substrate, the Lobster Observatory, in which ten Mandarin-language agents have been grounded since 2026-04-15 in a ten-site live prediction ecosystem. Three relational layers are extracted from direct database query: a directed trust proxy (localBrain.socialInfluence), a directed listening exposure layer (listeningProfile.bySpeaker), and a team-outcome layer derived from a deliberately-designed three-agent Tiamat raid task.

The trust proxy saturates at 0.825 across all 90 directed pairs (zero variance), reflecting the recent integration of the trust subsystem. The listening layer shows a clear bimodal split: five high-listener agents averaging 85,659 lounge messages heard, five low-listener agents averaging 16,046 (between-tier ratio 5.34×), partitioning the cohort into a four-quadrant role typology (LISTENER, TALKER, BOTH, QUIET). The outcome layer fits a single-parameter binomial null model (χ² = 13.27, df = 10) across all 120 possible three-agent teams in 1,440 trials and is statistically indistinguishable from a uniform success process. Within-team listening density does not predict team win rate (Pearson r = −0.07, permutation test p = 0.45 over 10,000 permutations).

An ecological-validity robustness check against 3,128 completed live raid records (nine triples at n ≥ 30) places live substrate performance in the same low-success regime as the sandbox (mean = 7.56%, range 6.25%–9.73%).

The negative finding is reported as a substrate observation rather than a general claim: in this substrate, with this observation window, even a deliberately-designed cooperation task fails to surface an alignment between communication structure and cooperative outcome. The three layers describe the same ten agents over the same window, with three different topologies; the decoupling itself is the substrate-level observation worth recording.

This is the third paper from the Lobster Observatory research program (cf. Chen 2026a, 2026b on emergent epistemic norms and active trust modulation; Chen 2026c on substrate architecture).