Skip to content
/ binling Public

Deterministic Agent Control via Geometric Latent Space. An LLM-driven Lattice VM (Rust/WASM) that compiles natural language intent into verifiable assembly, eliminating execution hallucination.

License

Notifications You must be signed in to change notification settings

eddlev/binling

Repository files navigation

BinLing (BLE): Deterministic Agent Control via Geometric Latent Space

"English is too fuzzy for physics. We need an Agent that thinks in Assembly."

[Green Swarm Visualization]

(Self-Replicating Swarm Probe generated by Qwen 2.5 and executed in BinLing VM)

🔬 The Experiment

Standard AI Agents rely on "Chain of Thought" (CoT) in natural language to execute tasks. This works for emails but fails for autonomous control because English is high-entropy and prone to hallucination.

BinLing proves that by constraining an LLM's latent space to a low-fertility instruction set, we can achieve deterministic, physics-compliant behavior.

🧠 Theoretical Basis

BinLing implements a Von Neumann Universal Constructor architecture. The REPL opcode functions as a kinematic self-replicator, treating the execution lattice as a resource substrate. This makes the Swarm Probe a true Von Neumann Probe—an autonomous agent capable of exponential replication via in-situ resource utilization.

The Stack:

  • Intelligence: Qwen 2.5 7B (acting as a Compiler).
  • Bridge: Python Driver enforcing strict schema constraints.
  • Physics: Custom Rust Virtual Machine (binling_cli) managing a 3D Polytope Lattice.
  • Visualization: WebAssembly (WASM) Real-time Renderer.

📊 The Findings: "The Tokenizer Tax"

During our v0.1 benchmarks, we observed a critical distinction between Token Efficiency and Execution Density.

Metric Natural Language Input BinLing ASM Output
Content "Create a self-replicating swarm..." LOAD 0 0 0 80 REPL 1 0 0...
Length 287 Bytes 134 Bytes
Token Count 80 Tokens 86 Tokens
Reliability Probabilistic (Might fail) 100% Deterministic

Insight: While modern tokenizers are biased toward English (making ASM appear "expensive" in token count), the Byte Density of the executed logic is 6x higher. We trade token efficiency for execution safety.

🛠️ Components

  • core/: The Rust Reference VM (Lattice Scheduler, Levin Method).
  • cli/: The Assembler and Runtime Environment.
  • interface/: The "Oracle" text files used to bridge Python and Rust.
  • spec/: Authoritative definitions of the Encoding and Scheduling.

🚀 How to Replicate

  1. Boot the VM:
    cargo run -p binling_cli --bin binling_cli
  2. Run the Benchmark:
    # Runs the physics engine benchmark (approx 600µs per genesis cycle)
    cargo bench -p binling_core --features cli-mode
  3. Inject Intent: Use the Python driver (or edit interface/oracle_in.txt manually) to inject BinLing Assembly.

📜 Roadmap & Status

  • Reference VM: Rust implementation of the Lattice Scheduler.
  • Levin Method: Active/Next queue scheduling implemented.
  • Assembly Bridge: Text-to-Bytecode compiler (CLI).
  • Visual Tokenization: Mapping opcodes to single pixels (v2.0).

License

MIT License - Open Source for the advancement of Deterministic AI.

About

Deterministic Agent Control via Geometric Latent Space. An LLM-driven Lattice VM (Rust/WASM) that compiles natural language intent into verifiable assembly, eliminating execution hallucination.

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published