BlackMind

Publications

Research

BlackMind studies machine intelligence at its true substrate — compiled bytecode, Boolean logic, and standard-cell silicon. Every capability claim is discharged by machine proof, and every number re-derives from a committed, machine-checked receipt.

Latest

Alphamon: Bit-Exact GPU Self-Play, a Training-Free Value Bound, and Search-Dominated Play
September 2026Reinforcement learningGPU systemsEvaluation

Alphamon: Bit-Exact GPU Self-Play, a Training-Free Value Bound, and Search-Dominated Play

Alphamon is a complete AlphaGo-style agent for the Pokémon Trading Card Game — and three transferable lessons from building it. It ports a 1,267-card commercial engine to CUDA at one game per thread, certified bit-exact over 70,000 games and 7.9M decisions; uses that engine for a training-free bound showing the AlphaZero recipe sits near an aleatoric ceiling the game itself imposes; and finds that search, not learning, carries the agent's strength.

Jovonni L. Pharr · Georgia Cyber Warfare Range / BlackMind
Read paper

All papers

Chasing the Small-Frontier on a Fixed Rig
September 2026Small language modelsEvaluation

Chasing the Small-Frontier on a Fixed Rig

A measurement-driven account of training a small language model to frontier quality on two fixed GPUs — with adversarial self-audit of the lab's own instruments.

Read paper
Large Binary Models
September 2026Language modelsProgram synthesis

Large Binary Models

A model that learns to read and write compiled bytecode directly — no natural language in the loop.

Read paper
Binary Circuit Models
August 2026HardwareFormal verification

Binary Circuit Models

Compiling a trained binary model into SAT-verified logic gates — and taping it out in 130nm silicon.

Read paper
The General Purpose Neural Computer
August 2026Neural computersFormal verification

The General Purpose Neural Computer

A differentiable machine that learns WebAssembly's semantics from execution alone — deterministic, length-generalizing, machine-proven.

Read paper