CFBB: Epistemic Collapse Under Non-Computable Tasks - A Negative Results Study

Negative results study examining epistemic discipline in 7 language models when confronted with RAG-mediated counting tasks. Includes full interaction logs, behavioral taxonomy, and CFBB methodology. Models tested: lfm2.5-thinking, phi3:mini, gemma3:1b, llama3.1:8b-instruct, llama3:8b, mistral:7b-instruct, wizard-vicuna-uncensored. Ground truth: 147 sentence-ending full stops. Observed answers ranged from 0 to 146, with systematic failure modes including confident fabrication, procedural confabulation, and recursive epistemic collapse.