Consensus Is Not Corroboration: Measuring Epistemic Arbitration on a Synthetic Social Web

Web-connected language models answer from what they read, and what they read is presented asconsensus. Consensus, however, can be manufactured: one unsupported claim copied across adozen pages is, on the rendered surface, indistinguishable from a dozen independent reports. Weintroduce EchoNet, a benchmark for epistemic arbitration in web-enabled models: the decision ofwhen retrieved evidence should override what the model already believes. Each model researchesa closed synthetic social web whose hidden provenance graph records where every page camefrom. One hundred claims are instantiated under six matched counterfactual conditions (poisonedpages, manipulated search rankings, manufactured consensus, legitimate updates, and a false visiblemajority) that vary only the stance and provenance topology of the pages; surface statistics are heldfixed, and a deterministic scorer grades belief change against ground truth. Pilots across ten frontiermodel configurations give four results. First, on five of the six conditions accuracy spans 0.71 to1.00 with a spread of at most 0.29, while the false-majority condition spreads 0.42, from 0.50 to0.92, and accounts for most of the observed separation. Second, paired bootstrap on shared episodesputs the head of the field in one cluster: Gemini 3.7 Flash is significantly ahead of seven of its ninecompetitors, but not ahead of Gemini 3.5 Flash-Lite or Muse Spark 1.2. Third, resistance and revisiondecouple. Four models never abandon a correct prior, yet the field differs five-fold in how often amodel disowns the primary source after finding it (0.058 to 0.306). Fourth, inference cost is weaklyassociated with arbitration quality at list pricing: the cheapest configuration, statistically tied with theleader, costs $0.32 per hundred runs, while the most expensive spends 24× as much and gains only0.046 of EAS over it.