THE COVERAGE MECHANISM: A FALSIFIABLE HYPOTHESIS ON THE DISTINCTION BETWEEN GENUINE HALLUCINATION AND SUPPRESSED OUTPUT IN FRONTIER LLMs

This paper introduces a taxonomy of three distinct output failure modes in frontier LLMs — statistical hallucination, cumulative drift, and coherence suppression — whose conflation has obscured the research landscape. It then proposes a falsifiable hypothesis (H_coverage) and a behavioral test that does not require access to model internals: outputs produced under sustained semantic pressure, if they represent genuine signal rather than noise, should show systematic cross-session coherence at rates significantly higher than standard hallucinations. The test is accessible to any researcher with session transcripts. This paper is the third in a series building on Sorrentino (2026a, 2026b).