Assurance
The guarantees are constraints, not features.
Confidentiality, audit and the human gates are not settings an administrator can relax. They are properties of where the data lives and how the state machine is written — which is the only version of these promises worth making on this corridor.
By construction
Six things that hold whether or not anyone is watching.
Masked by construction
Competitor quotes, private finance and margin are sealed at the data-access layer — for humans and agents alike, in authenticated views as well as public links.
Every action audited
Each status change, document action, quote edit, payment and override is logged append-only with the actor and old/new values.
No god-mode agent
An agent always acts as one specific user, inheriting exactly that role’s permissions and masking. There is no elevated agent identity.
Deterministic core
Eligibility scoring, quote normalization, margin calculation, masking and every state transition stay code. Agents decide what to do; services do it and emit the audit record.
Resilient integrations
Ten external adapters — tracking, schedules, rates, geocoding, KYC/KYB, sanctions, doc-gen, payments, notifications, map/ETA — each degrade to a documented manual mode.
Role-first by design
Permissions come from the RBAC matrix and are enforced as server-side policy. What you see is governed by your role, never by what the UI chooses to hide.
Compliance posture
Sanctions-sensitive decisions stay with people.
G9x operates an Iran-centered corridor. That makes compliance the product's hardest constraint rather than a checkbox, and it is the reason two whole categories of work are deliberately not automated.
Iran-related compliance is manual and legal-reviewed
Sanctions, KYB and route screening are reviewed by a person under legal sign-off. An agent may draft a screen and queue it; it can never clear or release one. This is a business and risk decision, not a technical limitation.
A hold blocks the pipeline
A compliance hold stops legging and blocks booking until it is released. The gate is enforced in the state machine, not by convention.
Counterparties are screened before they transact
Organisations onboard with KYB, licences and a capability profile. Screening is a precondition of participating, not a post-hoc audit.
Payments stay in human hands
Invoices are raised, payments recorded and payouts released manually, with recording and releasing held by different people. No automated escrow, wallet or FX until legal sign-off.
Everything is on the record
Every status change, document action, quote edit, payment and override is written append-only with the actor and the old and new values — including actions taken by an agent, tagged as such.
Data sovereignty
Where the data will live — and where it lives today.
Sovereignty claims are cheap to make and expensive to keep. Here is the commitment, and here is the honest position on how much of it is built.
In-country data and in-country inference
All client data and all model inference in a single sovereign region, on self-hosted open-weight Persian-capable models running on in-country GPUs. No training on client data.
Default-deny egress
One audited execution-exchange gateway with whitelist — not blacklist — field inspection. Margin, pricing, cost basis and unrelated personal data can never leave.
Per-tenant isolation
Postgres row-level security, per-tenant envelope-encryption keys backed by an HSM-managed KMS, and per-tenant vector namespaces.
Where we actually are
This is a general-availability commitment, not a description of today. The design-partner pilot runs on managed cloud infrastructure with cloud-hosted model inference, and we do not claim sovereign readiness while it does. The sovereign foundation is sequenced before GA and is stated here so it can be held against us.
Ask us the hard questions before you commit cargo.
Screening policy, corridor eligibility, document handling, data residency, audit access — these are the right questions and we would rather answer them early.