Silent Failure in MiroFish: What Happens When an AI Society Simulator Gets Pure Lorem Ipsum

An investigation into the open-source multi-agent society simulation stack MiroFish, conducted by a researcher in computational social science, uncovered a critical defect termed 'Silent Failure'. The study loaded defective inputs, including pure lorem ipsum, into the MiroFish-Offline distribution, which uses a typical GraphRAG architecture with Neo4j, an agent layer, and LLMs via OpenRouter. In two UI runs with DeepSeek and Claude Sonnet 4, the pipeline correctly extracted zero entities from the nonsense text but then hung on agent generation, showing a non-progressing 'Configuration generating' status for 30 and 70 minutes respectively, with no timeout or error. However, an instrumented run observing the backend directly revealed that the backend actually failed within milliseconds (24-33 ms) at the agent stage endpoints, returning clear error messages like 'No entities matching criteria found', contradicting the UI's apparent freeze. This indicates the system's UI layer fails to communicate backend failures to the user, a design flaw independent of the specific AI model used. The study, part of a three-part series, aims to map where such pipelines silently normalize nonsense rather than detect it, with further findings on absurd inputs like the fictional country Valdoria to be published in later parts.