Source-Attributed Intelligence

Know where your AI's knowledge comes from

SEDIM decomposes model weights into traceable knowledge layers. Every response comes with a STEMMA — a real-time attribution map showing which sources contributed.

CENTO = FACIES + Σ STEMMA · VARVE
15
VARVEs
8 languages + code + reasoning
94.8%
Routing Accuracy
STRIA domain classification
90.6
SEDIM-Bench
5-dimension evaluation score
0.0265
Val Loss
Beats FFT (0.0281) & LoRA (0.0278)
Built for regulated teams

Where “my model said so” isn't an answer. Source attribution becomes a compliance artifact.

Legal
Discovery, memo drafting
Compliance
Regulatory mapping
Finance
Audit trail, FINRA
Healthcare
HIPAA, clinical notes
Sales
Revenue ops
Internal
Policy + ops

Architecture

15 VARVEs on Qwen3-8B FACIES, per-block STRIA routing

FEHM8
TR, EN, AR, DE, FR, ES, JA, ZH
MING2
Python, TypeScript
CORTEX5
ML, Data, Reasoning, Planning, Eval

SEDIM-Bench

Five-dimensional evaluation — overall score 90.6 / 100

RoutingIQSTRIA accuracy + calibration
94.8
FusionIQMulti-VARVE synergy
100
IsolationIQCross-VARVE contamination
100
AttributionIQSTEMMA faithfulness
85
VARVEiqDomain-specific accuracy
73.3

vs. Baselines

SEDIM outperforms while providing source attribution

MethodVal LossSource Attr.Modular
Full Fine-Tune0.0281✗✗
LoRA r=1280.0278✗✗
SEDIM (K=2)0.0265✓✓
SEDIM (K=15)2.28*✓✓
* Best single-VARVE eval loss (fehm-en). K=15 trained with domain-separated data.

Production Ready

Deployed on serverless GPU with predictive warm-up

~$57
per month
1K requests/day
2.66 GB
VARVE weights
15 VARVEs, ~16% of base
<30s
cold start
Predictive warm-up on login

Ready to build with attributed AI?

Create your first VARVE in minutes