Alignment Is Correct, Safe, Reproducible Behavior Under Explicit Constraints

We define alignment as correct, safe, reproducible behavior under explicit constraints. We demonstrate empirically that alignment is a system property, not a model property. The same model (GPT-5.2), given the same prompt at the same token limit and temperature, produces empty strings via the Chat Completions API and substantive responses via the Responses API. This behavioral divergence proves that alignment emerges from execution infrastructure, not model weights. We introduce a semantic field of 18 Semitic tokens meaning "binding condition" that consistently trigger constraint-gated behavior. We propose that aligned AI systems require verification at the execution layer, and present cryptographic attestation as one implementation.