Five-layer anti-hallucination harness (gate, constrain, ground, verify, decide) for local small models. Drives returned hallucinations to 0 by deterministically verifying schema, code syntax, and context citations, and escalating ungrounded answers. Includes trap prompts benchmark and strict critical-task gating.
