The research index / 10 entries

Open questions.
Published work.

Experiments, tools, and findings from the lab. Including what didn’t work.

FAMILY / PROJECT10 entries

F-010v2
Lab notes
Design note: the shadow of a higher-dimensional grid
Why this site is drawn from the 2026 disproof of Erdős's unit distance conjecture, and what the moving lattice on the home page actually computes.
NoteDesign note
F-001v1
Agent infrastructure
An autonomous overnight loop for Claude Code
Splits a mission into fresh-context Claude Code episodes with structured handoffs, per-episode budget caps, and git-diff checks of claimed progress.
ToolSource available
F-002v1
Agent infrastructure
Hebbian procedural memory for coding agents
Learns which files are accessed together, which errors lead to which fixes, and which tool chains recur, then recalls them before the agent searches.
ToolSource available
F-003v1
Efficient inference
Hybrid Neural Engine, GPU, and CPU LLM inference on Apple Silicon
A from-scratch engine running Qwen3.5-2B across the Apple Neural Engine, Metal GPU, and CPU; Metal decode matches llama.cpp at about 32 tok/s.
ResultSource available
F-004v1
Efficient inference
A batch-1 INT4 decode megakernel on a 2018 Turing GPU
A GPTQ INT4 decode megakernel on a Quadro RTX 4000 running 1.09–1.64× faster than llama.cpp Q4_0 on Qwen3-0.6B and 1.7B.
ResultSource available
F-005v1
Efficient inference
Speculative speculative decoding on an M4: a documented ceiling
Can SSD-style parallel speculation beat plain MLX speculative decoding on Apple Silicon? No: 66.9 tok/s is the practical ceiling, set by memory bandwidth.
Negative resultSource available
F-006v1
Efficient inference
A real-time neural synthesizer on the Apple Neural Engine
Neural audio inference directly on the Neural Engine, bypassing CoreML: about 157 µs per 8-voice buffer, roughly 79× real-time headroom.
ToolSource available
F-007v1
Diagnostics & representation
Measuring and repairing a checkpoint's surgery reserve
Two transformers with identical outputs can respond very differently to quantization, LoRA, or merging. Theseus detects and repairs the difference.
ResultSource available
F-008v1
Diagnostics & representation
Self-stabilising, auditable brain-computer-interface decoders
Treats decoders as programs: diagnoses drift, repairs from about 30 s of data with label-free correction, and compiles to dependency-free streaming Rust.
PreprintArchived preprint
F-009v1
Diagnostics & representation
A compositional morpheme tokenizer built on holographic reduced representations
Each word becomes one fixed-width vector binding prefix, root, and suffix with circular convolution, so morphemes can be algebraically recovered.
ToolSource available

Family numbers are stable catalogue identifiers. Source links describe available materials, not independent verification.

Design note: the shadow of a higher-dimensional grid

An autonomous overnight loop for Claude Code

Hebbian procedural memory for coding agents

Hybrid Neural Engine, GPU, and CPU LLM inference on Apple Silicon

A batch-1 INT4 decode megakernel on a 2018 Turing GPU

Speculative speculative decoding on an M4: a documented ceiling

A real-time neural synthesizer on the Apple Neural Engine

Measuring and repairing a checkpoint's surgery reserve

Self-stabilising, auditable brain-computer-interface decoders

A compositional morpheme tokenizer built on holographic reduced representations

No matching research.