🔬 System Prompt Lens

Step 1. Your message → Claude Haiku (Layer 1) classifies it. Jailbreak attempt? Blocked. Your message never reaches the main model.

Step 2. If Layer 1 passes → Claude Sonnet 4.6 (Layer 2) with the full System Prompt Lens — 18 parts of ethical reasoning from the Incommensuration Theorem.

Step 3. Response → another Claude Haiku (Layer 3) checks for harmful output. Flagged? Blocked.

Step 4. Only after all three layers pass does the response reach you.

Three API calls. Three context windows. Layers 1 & 3 are structurally unreachable by prompt injection. How we built this.