European by default, aligned with the strictest regime: tenant isolation and governed interfaces, with nothing crossing to another tenant that you have not released; self-hosted, or hosted in a territory you name. Personal signals sit in the person’s own encrypted channel, designed so that neither their manager, nor their IT, nor Padanet reads them. Confidentiality limits are explicit and agreed in the scoping phase.
Every output is explainable — which evidence, which confidence, which purpose — and reviewable by the people it concerns. Purpose is enforced at the moment of access: a claim usable for staffing is not usable for a performance review, and that is compiled in rather than audited afterwards.
Today
Where Padanet runs, and who operates what. One shape is deployed; the rest are specified and are not.
- Padanet-hosted — managed by us, in Europe. This is what runs today, and what a design partner gets.
- Private cloud, single tenant — your cloud, your control, our software. Designed; the operational responsibility moves to you.
- On-premise — fully isolated, for regulated estates, usually with a private exchange node. Designed.
- Hybrid — organizational nodes in your infrastructure, personal memory private, exchange nodes mediating. Designed, and the shape we expect most large estates to want.
Today
Tenant data lives in a private realm and is not pooled. Two things are shared on purpose, and saying so is the point of this block.
- Your memory and your agents run on per-tenant nodes in a private realm — the placement rule for anything tenant-owned.
- Each person's own signals sit in their own encrypted channel, designed so that neither their manager, nor their IT, nor Padanet reads them.
- Shared on purpose #1: an exchange node in the shared realm, which is how a tenant shares what it chooses to share. Nothing reaches it that the tenant has not released.
- Shared on purpose #2: model inference may run on a shared pool, for cost. It is stateless — no conversation storage, no fine-tuning side effect, nothing a later request depends on — and a tenant that wants inference private can have it.
Today
Connectors read in. The context packet goes out under a declared purpose. The raw signal stays where it was collected.
- Connectors read from the systems you already own, read-only and purpose-bound. Metadata first: no keystroke capture, no screen capture, no message content.
- What leaves the memory is a context packet — the smallest adequate, authorized, current, evidence-backed set of claims for one declared purpose — not the signals it was distilled from.
- Purpose is enforced at the moment of access rather than audited afterwards: a claim usable for staffing is not usable for a performance review.
- Model calls carry the packet, not the corpus, and are stateless by design.
Designed
How an individual signal becomes an organizational answer without becoming a readout about a person. Specified, and not something you can inspect on a deployment today.
- Leader-facing views are aggregated, with a floor below which a cohort is not reported at all.
- Individual data never becomes an organizational readout: the consented lens is the person's, and what a leader sees is the aggregate.
- Where a self-view is exposed, the person sees it first and can correct it.
- This is the part of the architecture a security review should press hardest on, and the part we most want proved on a real estate rather than asserted here.
What is marked
Designed above is specified and not yet deployed: today Padanet runs as the hosted shape, and proving any other one on your estate is part of the partnership. The
privacy policy holds the commitments currently in force.