Research

SEDIM Architecture Papers

Empirical results and open research from Nage AI and Autorite.

Results

What the numbers show

0.817
CKA Score
CONFLUX transfer quality
Fehm A/B experiment
−6.4%
Val Loss
Improvement over baseline
CONFLUX vs random init
Convergence
Training speed gain
Asymmetric SVD initialization
0.94
IsolationIQ
Cross-VARVE isolation
Near-perfect contamination prevention
Transparency Note Initial results (N=1, single domain). Multi-seed, multi-domain validation in progress per rigorous experimental protocol.
SEDIM Paper

Sedimentary Intelligence
Model

SEDIM: Sedimentary Intelligence Model — VARVE, FACIES, CLAST, STRIA, and STEMMA for Source-Attributed Compositional LLMs


CENTO = FACIES + Σi STEMMAi(x) · VARVEi

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.


Key Contributions
Source attribution at the architecture level
Knowledge isolation across VARVE layers
CONFLUX cross-architecture transfer protocol
4 inference modes (single-VARVE, multi-VARVE, STEMMA-routed, full CENTO)
SEDIM-Bench evaluation standard
Paper Status
StatusDraft complete
ArxivSubmission targeting July 2026
VenueArxiv → NeurIPS / ICLR
Citation (Preprint)
Nage AI Research (2026). SEDIM: Sedimentary Intelligence Model — VARVE, FACIES, CLAST, STRIA, and STEMMA for Source-Attributed Compositional LLMs. nage.ai/research.
CONFLUX

Cross-Architecture
SVD Transfer

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.

Transfer Threshold
CKA > 0.8Strong transfer
CKA < 0.3Minimal benefit
Open Source
github.com/NageAI/conflux — Apache 2.0
SEDIM-Bench

Evaluation Standard

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.


Benchmark Dimensions
Attribution accuracy (STEMMA routing precision)
Knowledge isolation (cross-VARVE contamination)
Compositional coherence (CENTO output quality)
Transfer efficiency (CKA-guided CONFLUX scores)
Inference routing quality (mode-specific benchmarks)
Release Plan
ReleaseMay 2026 (open source)
BoardHuggingFace leaderboard planned
Timeline

Research roadmap

April 2026
RunPod experiments (rigorous N=5 protocol)
Multi-seed, multi-domain validation runs. Establishing statistical significance for CKA threshold hypothesis and CONFLUX transfer metrics.
May 2026
sedim-bench open source + HF leaderboard
Five-dimensional evaluation standard released. Public HuggingFace leaderboard for source-attributed model comparison.
June 2026
Model training (Ming, Cortex) + paper finalization
Complete the model portfolio. Ming-8B (Chinese, systems thinking) and Cortex-14B (Latin, orchestrator). Final paper draft with full experimental results.
July 2026
LAUNCH — API + open source + Arxiv
Full platform launch. STRATUM API public access, SEDIM framework stable release, Arxiv paper submission.