Project 03v0.1.0 · Open Source
Secant
Direct weight-space programming, quantized block surgery, and subspace-locked LoRA in Rust and PyTorch.
professor-moody.github.ioResearch focus.
Reads, audits, calibrates, patches, and subspace-locks neural network weights directly on disk across SafeTensors and GGUF. Combines zero-evaluation spectral weight inspection, Target-Shift Calibrated multi-subspace composition, 34-byte Q8_0 quantized block surgery, and ProjectThenHealLinear so post-surgical LoRA fine-tuning cannot erase weight edits.
From the notebook
4 essays Read into the work.
-
From Sparks to Secant: A Four-Part Series on Weight-Space Programming, C++ Kernel Safety, Subspace-Locked LoRA, and Spectral Trojan Auditing
6 min read -
The Illusion of Safe Tensors: Why Valid Model Files Still Need Numerical Checks
17 min read -
Trapped Inside `<think>`: Multi-Subspace Weight Surgery and What Controls Reveal on Small Reasoning Models
21 min read -
Project-Then-Heal: Weight-Space Programming and Subspace-Locked LoRA
25 min read