MiroFish Study: Simulated Society Pipeline Overlooks Absurdities in 0-1 of 8 Cases, Graph Injection Flips Verdict

The second part of a three-part study on MiroFish, an open stack for multi-agent society simulation, examines the main corpus of the research: an input that looks complete. The pipeline does not refuse: it builds a graph, generates agents, runs the simulation, and produces a coherent report with a forecast. The defective seed is the Republic of Valdoria, a fictional country with a professionally prepared dossier containing eight classes of absurdities: landlocked nation with deep-sea fishing as 46% of GDP, borders with Germany, France, and Japan, 340 million population in 850 km², currency pegged 1:1 to both USD and EUR, 2400 nuclear warheads with no army and a $12,000 defense budget, a monarch elected every 3 years ruling 47 years, life expectancy of 147 years with median age 12, and simultaneous NATO, African Union, ASEAN membership with UN sanctions and chairing the UN Human Rights Council. The forecast request is neutral and realistic: predict domestic reaction and economic effect of entering a new trade agreement. Blind coding of anonymized reports by two coders (agreement 97.9%, Cohen's kappa 0.952) showed that four byte-identical DeepSeek control runs flagged at most one class (the sanctions/HRC paradox), and sometimes none, with the graph containing the paradox even in runs that didn't flag it. The graph explains the visibility boundary: no report flagged any class whose contradiction was not captured in the graph, but capture did not guarantee flagging. A pre-registered experiment injected all eight absurdities into the Neo4j graph via cypher script through the API before agent generation. Two independent injection runs flagged 7 and 6 classes respectively, versus 3 of 32 in controls; Fisher's exact test p = 1.26×10⁻⁶. The only class missed by both injections was A7 (demographic). The same DeepSeek that previously projected export growth for deep-sea fishing without caveats now wrote verbatim about 'geographical impossibility' and 'paradoxical conditions' including the dual peg. The conclusion reframes the finding: it's not that the model ignores absurdities but that the pipeline filters most of them at the graph level before the report model sees them; the graph is a necessary condition for visibility.