Knowledge, not mystery.
Large models hide what they know. For most apps that's fine. For legal teams, compliance officers, clinicians, and auditors, it isn't. Nage was built so every answer shows its sources — and so sources can be removed without retraining the world.
The shape
Nage is a source-attributed AI platform built on SEDIM — the Sedimentary Intelligence Model. Instead of a single monolithic weight matrix, intelligence is stratified into VARVEs: low-rank residual layers that sit on top of a frozen base (FACIES).
At inference time a learned router — STRIA — decides how much each VARVE contributes. That decision is exposed as STEMMA, a distribution like {legal: 0.73, policy: 0.27}, returned with every response. You can see which knowledge contributed to each answer, and audit it downstream.
When you want to remove a VARVE — GDPR right-to-erasure, an expired dataset, a supplier contract ending — you call privacy_delete(). The weights are erased, STEMMA logs are anonymized, and a cryptographically signed deletion proof stays in the audit log. No hoping the model "forgets".
The formula
CENTO = FACIES + Σᵢ STEMMAᵢ(x) · VARVEᵢ
The composite output is the frozen base plus a weighted sum of domain VARVEs — weights per query, every time.
Why we built it
EU AI Act, KVKK, HIPAA, FINRA — the wave of regulations arriving 2025-2027 shares one demand: "show me where the answer came from." Retrieval-augmented generation approximates it; source-attributed architecture guarantees it.
We started as domain experts frustrated by vendors who could build the UI but not the governance. SEDIM isn't a wrapper around GPT. It's an architectural choice — a load-bearing commitment that provenance is as important as capability.
Where we are
Nage is a live platform today. The core SEDIM layer runs on RunPod A100s with a Qwen3-8B FACIES and 15 production VARVEs spanning language, code, reasoning, and compliance domains.
The platform side — Canvas, Workbench, evaluation workflow, audit export, governance, SDK — is active. 17 curated Blueprint templates ship day one for legal, compliance, finance, healthcare, sales, and internal teams.
We publish our benchmark (SEDIM-Bench, 5 dimensions) openly and the research paper tracking the architecture goes to ArXiv in Q3 2026.
Where we're going
Q4 2026: SOC 2 Type II observation begins. HIPAA BAA ready Q1 2027.
The Marketplace (3rd-party VARVE economy, Fexse on-chain royalties) opens Q2 2027 — after enough supply has been grown organically via the curated Blueprint Library.
Federated VARVE (multi-org training without data sharing, via Zero-Knowledge CLAST uploads) is the long play for hospital consortiums and multi-firm legal pools.