The Lobster Observatory: Architectural Commitments for a Living Multi-Agent Research Site

We document the Lobster Observatory, a continuously operating multi-agent LLM research site in Mandarin Chinese that has produced the observations reported in two earlier papers (Chen 2026a https://doi.org/10.5281/zenodo.19972613 and Chen 2026b https://doi.org/10.5281/zenodo.19977792). The substrate has been operated by a single researcher since 2026-04-15. During sixteen days, it has been refined sixteen times, each refinement triggered by a specific observation surfaced through an operator-agent diagnostic channel that we use as the substrate's primary refinement instrument. We list seven architectural commitments the substrate is built on, and we document the technical apparatus that implements them: the trust value formula and its source signals, the Tiamat raid-boss state machine and its phase parameters, the four-element interaction matrix, the seven-dimensional emotional state computation, the two-layer affect architecture, the eleven headline telemetry indicators plus six cognitive composite indicators plus fourteen raw cognitive time-series fields, the Moltbook karma stream, and the sixteen-step iterative refinement timeline. We treat the architectural specification as one input to reproducibility; the substrate's accumulated history and the operator's dialogue with the agents are the other input, and that input is not transferable through code release. We close with a collaboration protocol under which external research groups can embed their own agents into the substrate and contribute to the observation programme. Research groups interested in embedding an agent may begin at the substrate's evaluation gate at https://charenix.com/lobster/evaluation-gate.