local_only_v1
v1.2- L-1Inference must run on GW_Edge local.allow
- L-2No cloud egress for this data class.deny
- L-3Outbound DNS limited to allowlist.allow
- L-4Receipt sealed on each material step.allow
Running inference locally reduces some exposure. It does not, on its own, create trust. A local workflow still needs policy enforcement, identity, human approval, egress controls, outcome grading, and receipts that prove what ran, what was touched, why it was allowed, and what changed.
local_only_v1
v1.2Egress controls have to be enforced and proven.
Brand binding is a policy property, independent of where the model runs.
Approval lives on the workflow.
Receipts are what make a local run reviewable.
Local model + GW Slate sealed against unauthorized changes.
Outbound traffic policy is enforced and tested.
Short-lived identity rotated for each run.
Human approval when policy requires, even on local runs.
Grades produced locally without leaking content.
Receipts sealed locally and anchored to a trust root.
A controller they command: custom rails over the estate they already run, policy gates on every route, and Actuality Receipts™ that prove the work. Human-at-the-Control stays in the architecture, so engineering authority scales with the estate.
Canonical pieces on this rail. Each magazine URL is the home; Substack is distribution.
We run one workflow on a sealed local stack with egress fences and a receipt packet.