Empirical results and open research from Nage AI and Autorite.
SEDIM: Sedimentary Intelligence Model — VARVE, FACIES, CLAST, STRIA, and STEMMA for Source-Attributed Compositional LLMs
SEDIM decomposes any final weight matrix into a permanent base (FACIES) and a set of source-attributed deltas (VARVEs), each traceable to a specific training phase and knowledge origin. At inference, the STEMMA routing function dynamically weights each VARVE's contribution based on query semantics.
CONFLUX is a cross-architecture SVD transfer framework developed by Autorite. It uses Centered Kernel Alignment (CKA) as the primary quality gate for knowledge transfer between models trained on different domain corpora.
Core finding: CKA > 0.8 indicates strong transfer potential. CKA < 0.3 yields minimal benefit. This threshold relationship is the primary quality gate for SEDIM training.
The asymmetric SVD initialization approach initializes only the A-matrix at scale 0.01 while setting B to zeros — a deliberate asymmetry that preserves the decomposed structure of source weights without over-constraining the target model.
SEDIM-Bench is the five-dimensional evaluation standard for source-attributed models. It measures attribution accuracy, knowledge isolation, compositional coherence, transfer efficiency, and inference routing quality.