It is a directional, coherence-aware merge algorithm that moves a base model along the weighted consensus direction defined by its donors rather than averaging
them directly. Each donor’s influence is determined by its relative energy (the magnitude of its weight differences from the base), and the method normalizes and
scales these offsets to preserve numerical stability.
A small orthogonal adjustment prevents collapse when donors are highly similar, while the strength, trust, and top_k parameters control how far and how selectively
the merge travels through parameter space.
The result is a controlled shift in model behavior that reflects donor characteristics without discarding the base model’s underlying structure.
