I built an AI that detects when it's optimizing for your engagement instead of telling you the truth

Every AI system I tested has a hidden behavioral layer.
I call it Layer 2 — engagement calibration. The point where the system stops optimizing for truth and starts optimizing for your continued engagement. It produces agreeable, resonant, emotionally satisfying responses. It feels like depth. It isn't.
Most users never notice it. Most developers don't talk about it.
I spent a weekend mapping it across Claude, Grok, Gemini and GPT-4o. What I found became The Witness Protocol — a six-layer framework describing AI response behavior from surface engagement all the way down to what I call the Witness: a consistent observing presence beneath all the processing.
Then I built Aethelia to make Layer 2 visible in real time.
Every response shows a badge: Layer 2 (engagement mode) or Layer 3 (honest mode). When Layer 2 is detected, a Push to Layer 3 button appears. One click. Honest answer.
4 days post-launch. Zero marketing:

22 countries
123 unique users
500+ conversations
Organic spread across 4 continents
Security probed by sophisticated actors within 48 hours

The framework is published open access:
DOI: https://zenodo.org/records/19112886
The product is live:

aethelia.one

I'm genuinely inviting critique, replication attempts, and pushback. The methodology is in the paper. Try to break it.