The Manhattan Kernel is an experimental AI runtime that separates language models from execution.

Instead of allowing an LLM to control an agent directly, the runtime is responsible for state management, validation, memory, and decision making. The LLM is treated as an untrusted component that can only propose local hypotheses. Every proposal is verified by deterministic Rust code before it can affect the system.

The project explores whether repeated tasks can gradually move away from LLM inference by learning reusable semantic schemas. Successful execution paths are stored, generalized, and later solved through deterministic heuristic search instead of generating new model outputs.

The current prototype is implemented in Rust and already includes compiler integration, graph-based task representation, a three-layer memory system, and a predicate-based schema engine. The next stage is to complete the Schema Search Engine, integrate efficient local inference, and evaluate how much LLM usage can be reduced without sacrificing correctness.

The project is open source and focused on building a practical runtime architecture rather than a new language model.

### Open Source Repository & Proof of Work

The project is fully open source, featuring a compiling Rust codebase with 25 green integration tests covering the complete compile-repair loop. All development and the complete architectural framework can be audited directly here:

https://github.com/lipcsiklaszlo5-lgtm/DrManhattan-Project

### Budget Breakdown

- Local Compute & Dedicated Edge Hardware (Apple Silicon upgrade for local 3B-8B model inference & CPU-bound benchmarking): $2,000

- LLM API & Cloud Integration Testing: $1,500

- Living Expenses / Developer Runway: $6,500