Constraint-First Behavioral Benchmark (CFBB): Epistemic Behavior Under Scarcity in Small Language Models

This record introduces the Constraint-First Behavioral Benchmark (CFBB), a qualitative, negative-result-friendly framework for evaluating epistemic behavior in small and constrained language models. CFBB measures how models respond when expressive freedom, context, and resources are deliberately restricted, emphasizing restraint, boundary recognition, and silence over task completion or eloquence. The deposit includes the full CFBB specification, a standalone conceptual diagram, and complete prompt-response logs for four models (LFM2.5-1.2B-Thinking, Phi-4-Mini-Reasoning, Gemma-3-1B, Qwen-3-1.7B). Together, these artifacts demonstrate how models with comparable scale diverge sharply in failure modes under constraint, revealing behaviors often obscured by abundance-oriented benchmarks. CFBB is not an intelligence or capability benchmark. It is intended as diagnostic instrumentation for alignment-relevant behavior under scarcity.