Gen-Zero: the deterministic decision engine for AI agents
Gen-Zero is an enterprise deterministic decision engine for AI agents. It replaces token-by-token language generation with closed-form geometric manifold computation, dual-process cognition (a fast System 1 reflex plus a bounded System 2 planner), bounded PUCT search over exact dynamics, and a five-layer fail-closed policy gate. It ships as a Model Context Protocol (MCP) server that any MCP-compatible agent can call directly.
Connect via MCP
Public cloud MCP endpoint: https://api.gen-zero.ai/sse?token=gz_public_free (free public access token: gz_public_free, provided as an experimental preview for research and evaluation, no commercial SLA). Source repo: xmond/gen-zero.
13-task benchmark results
- 13-task macro mean (frozen Qwen2.5-72B + LLaMA-3.1-70B features): 81.52%, not statistically distinguishable from this run's own single-model control (81.15%, 95% CI [-0.27, +1.01]) or its simplest fixed no-search configuration (81.60%, 95% CI [-0.85, +0.70]); "SOTA" is this project's internal track name for the search, not a claim of beating a published external benchmark (benchmarks/README.md)
- SQuAD 2.0 answerability: 91.30% test accuracy, zero generated tokens
- Deadlock torus trap-avoidance diagnostic: bounded PUCT search solves 100% of seeded episodes; a greedy reflex baseline solves 0%
Five breakthrough papers
- Zero-token decisions, decoded
- World models: the safety catch that slipped
- When "aligning" two models does not help
- The bias that makes AI pick option A
- A tiny model, and the threats it missed
Browse the full paper index โ