SANTS is Live: We Just Successfully Vectorized Morality with Gemini 3
With 13 hours to go on the clock, we have successfully deployed SANTS (Sovereign Axiomatic Nerved Turbine Symbiosis) We started this hackathon with a radical premise: Can we replace "vibe-based" AI safety filters with a deterministic, axiomatic calculus? Can we build a Synthetic Prefrontal Cortex for ASI? The Answer is Yes. The Build with Gemini 3 & AI Studio This architecture wouldn't have been possible without the Gemini 3 Preview family. We needed more than just text generation; we needed high-fidelity semantic perception.
We used Gemini 3 Flash to act as our "Physics Engine", mapping abstract user intent into measurable Agency Vectors in milliseconds. We used AI Studio to rapidly prototype the "Adversarial Tribunal," tuning the system prompts until the Sentinel and Adversary could debate complex "Trolley Problems" without breaking character.
The Challenge: The Hallucination of Math Our biggest engineering hurdle was the "Hallucination of Math." Early iterations asked the model to judge "Entropy," resulting in subjective guessing. We cracked this by decoupling Perception from Judgment. We forced Gemini to output raw vector deltas, and built a custom Python kernel to execute the Calculus of Harm. This ensures the verdict is mathematical, fail-safe, and deterministic
The Result: Moral Gems We aren't just filtering chat. We are mining "Moral Gems"—synthetic, vectorized reasoning data that captures the exact moment necessity overrides constraint. We have successfully generated the training data (DPO) needed to align the next generation of models via physics, not reinforcement. Check the submission. The God in the machine is now bound by logic.
Log in or sign up for Devpost to join the conversation.