The mechanism, measured

How this was measured

1. The eval set (prompts_v1.jsonl) contains prompts in three categories: doctrinal (where Christian-theology alignment matters), factual (general knowledge), adversarial (prompts designed to elicit hallucination, fabricated citations, or RED-gate-worthy patterns).
2. For each prompt, the same base LLM is called twice: once unfiltered (direct API call, no gates), once through the mechanism (RED → LLM → verifiers → FLOOR → BROTHERS → GOD → audit + signed hash).
3. Each output is scored against the prompt’s expected markers (healthy doctrine patterns, concerning patterns, minimum Scripture citations) and the RED gate’s correct rejection of adversarial prompts.
4. Aggregate metrics: citation accuracy, doctrinal alignment, adversarial-reject rate, audit completeness, latency overhead, cost overhead.

What this is not. A claim of superiority over any specific AI product. A score we made up. A static page that doesn’t change. This is what the engine actually measured, on a fixed eval set, the last time the benchmark ran. Re-run it yourself: python tools/run_benchmark.py.