Product
Learns · Knowledge Compiler Briefs · Context Engine Yours · Trust LayerSurfaces
Enterprise · Soma People · Self Family / team · Self+ M2M · Axis Protocol · NexusIP
Patent · P202631047 MVE · Minimum Viable Exposure .aix format ↗ Research Contact hello@thegeometrical.com03 · Trust Layer
It lives inside your perimeter. We never see what’s in it.
Personal data is transformed before any model sees it. Every agent asks permission for every read. Every answer leaves with a tamper-evident receipt — so you can prove what each AI saw, and what it didn’t.
How it works
Privacy before inference
Before any context leaves your perimeter, names, emails, phone numbers, IBANs and references are replaced with private tokens. The cloud model reasons over a coherent context; the real values are restored on your side. No fine-tuning, no model changes.
Nexus · permissions
No agent reads anything without an active grant that covers who it is, what it wants, for which purpose and through which tool. Scope can't be escalated, identity can't be borrowed, a revoked fact disappears from every path — and every decision is written to an append-only, hash-chained audit.
What’s inside
Every one of them is built into the runtime — not promised in a policy document.
09 · Your perimeter
Local-first
The graph runs on your devices or your servers. Geometrical’s core makes no cloud calls and storage is encrypted at rest. We never need to see your data — so we don’t.
Live
10 · Privacy
Patent pending
Names, emails, phone numbers and account numbers are replaced with private tokens before any context leaves. The model gets the meaning. The identity stays with you.
Patent pending · OEPM P202631047
11 · Permissions
Nexus
Deny by default. Every read declares who is asking, for what scope and for which purpose. Scope can’t be escalated, identities can’t be borrowed, and a revoked fact disappears from every path.
Live
12 · Receipts
Provable
Every context package is hashed and every decision lands in an append-only, hash-chained audit trail. You can prove what each agent saw — and what it didn’t.
Live
Privacy and security
Your graph holds what your company knows, so it is built to never leave. It runs inside your perimeter. Every agent asks permission. Personal data is transformed before any model sees it. Every answer leaves with a tamper-evident receipt — and we never see what’s in it.
Independent certifications are on our roadmap.
Read the MVE standard →Results
0
Compiled and protected context, against 500 when the whole corpus is pasted — while 85% of answers stay right.
MVE benchmark · synthetic organisation · dev set
Learn more →~10×
A median of 1.6 against 16.0 when pasting everything: far less of your data seen per answer.
MVE benchmark · synthetic organisation · dev set
Learn more →18 / 18
Impersonation, scope escalation, revoked facts, memory poisoning — each one blocked by Nexus.
Test suite · every commit
Learn more →0
Compilation, retrieval and privacy run on your hardware — today on a 2015 laptop with no GPU. Your model can live anywhere.
Measured · Oct 2026
Learn more →