Customer hostability
Customer hostability is the architectural commitment that every artefact a customer adopts can be run on the customer's own hardware, with the customer's own data, against the customer's own keys, with the customer's own audit ledger. The customer can fork it on day one.
This is not a pricing tier or a deployment option. It is the shape of the substrate. A platform variant that requires hosted-by-vendor infrastructure to function is not customer-hostable by construction. See also customer-first ordering, the six-tier sovereignty matrix, and the economic model.
What self-ownership means
Three concrete commitments, each operationally testable:
- Self-hostable bootstrap. A platform service runs on the customer's own machine via a systemd unit — no SaaS dependency, no provider login, no metered API. Per-service unit files are organised by function (inference routing, the three wiki surfaces, the fleet controller, and others) rather than one directory per service. A standalone install script that sets up the unit and fetches the binary exists for one service today — a consistent, uniform install script plus a manifest across every service is the intended shape, not yet the shipped state.
- Per-tenant adapters in the customer's filesystem. When the customer's adapter trains on the customer's corpus, the resulting
.loralives atdata/adapters/inside the customer's substrate. It never traverses the vendor's network. Loss-of-adapter risk falls on the customer's backup discipline, not on the vendor's continued operation. - Customer-anchored audit ledger. The apprenticeship corpus that records every editorial action and verdict lives at
data/training-corpus/apprenticeship/<task-type>/<tenant>/inside the customer's substrate. Tenant-private records never leave the customer's infrastructure.
Why hyperscaler economics preclude this
Hyperscaler economics depend on central training, shared trust roots, and provider-rooted attestation. Per-customer continued pretraining does not pencil at SMB scale on those economics:
- Central training requires central data; customer-on-metal training requires the opposite.
- Shared trust roots concentrate the regulatory burden at the provider; customer-anchored attestation distributes it to the customer.
- Provider-rooted attestation makes a customer audit a vendor audit; customer-anchored attestation lets the customer audit themselves.
The platform is built so the customer's regulatory posture does not depend on the vendor's. A BCSC reporting issuer per [ni-51-102] who deploys the platform inside their own substrate can satisfy continuous-disclosure obligations against records they own and control. A vendor outage does not impair their ability to respond to a regulator's request.
What actually ships today
infrastructure/ organises service installation by function, not by one bootstrap
directory per service:
| Group | What it covers |
|---|---|
| Wiki surfaces | One unit per wiki — documentation, corporate, and projects — plus the marketing site and proofreader. |
| Orchestration and fleet | The fleet controller, VM host agent, tenant proxy, and inference-routing units. |
| Command surface | The one service today with a full install script — sets up the systemd unit and fetches the binary from the release server, a local build path, or a pinned default. |
The intended shape — every service self-installable from a single script, in the customer's own language, with a manifest declaring what it provides — is real intent, not yet uniform practice. Most services today are brought up by hand-applying a systemd unit file rather than running a per-service install script.
What customer hostability is not
It is not "open source under a permissive licence." Permissive licensing is necessary but not sufficient. A permissively-licensed service that requires a hosted control plane to function does not satisfy customer hostability — the customer cannot fork it without forking the control plane.
It is not "the vendor will host it for you." Vendor-hosted deployment is one operational option among several. PointSav runs documentation.pointsav.com because PointSav itself is the canonical-tier customer of every package it ships. Customer-hosted deployment is the canonical pattern; vendor hosting is a convenience.
It is not "you can export your data." Data export is a feature. Customer hostability is structural — the substrate does not have a data-export feature because the data was already on the customer's hardware from the moment it was written.
Forward-looking — federated continued pretraining
Per [np-51-201], the language describing the federated continued-pretraining trajectory is forward-looking. The shape is in place; the operational throughput matures over time.
The trajectory: a customer's adapter trains on the customer's corpus inside the customer's substrate (Tier B Yo-Yo or the customer's own GPU host). Aggregated improvements may feed back to a shared base model when the customer chooses to contribute, under explicit consent. Customers who do not contribute continue to benefit from base-model improvements driven by customers who do, with no leakage of corpus contents either direction.
This is the federated-LoRA pattern adapted to per-customer data ownership. Reasonable basis: the federated-learning literature 2024–2026 and the LoRA-aggregation research surveyed in the Language-protocol substrate article.
Operational test for customer hostability
A new platform service is customer-hostable if a customer can:
- Clone the relevant repos.
- Run
bootstrap.shon a machine they own. - Generate, store, and serve adapters from local disk.
- Operate the service against the customer's data without ever contacting a vendor endpoint.
- Audit every request, response, and adapter-update locally.
Each of these is testable in CI. A service that fails any test is, by definition, not customer-hostable yet — and the gap is treated as a defect, not a feature.
See also
- Language-protocol substrate — the language model research that underpins the federated pretraining path
- Anti-homogenization discipline — the complementary discipline that keeps federated training from homogenising tenants
- Compounding substrate — the five structural properties that customer hostability enables
- Three-tier contributor model — how contributors interact with the open-substrate artifact ecosystem
Cite this record: /wiki/customer-hostability — revision 6fd838d6, last updated 24 August 2026.