Security one-pager

A condensed reference for security review. The full treatment is on the security page; this is the file-and-forward version.

Trust boundaries by deployment mode

On-prem (Mode 1) — runs on your own hardware and needs no internet once installed. As installed, with the standard on-prem engines, DocMark sends nothing off the box: no model call, no DocMark telemetry. You author; DocMark renders and verifies locally (the deck gate is on by default). An operator can enable AI engines; a job that uses one sends content to the model provider and is never zero-egress (see point 3 of the security page).
Client cloud (Mode 2) — runs in your own cloud tenant. Everything stays inside your cloud boundary; the only external call is to your own LLM account, under your contract.
DocMark cloud (Mode 3) — runs on DocMark’s managed Azure. DocMark processes content, but user identity and PII never reach the rendering plane.

On-prem zero-egress = as installed, DocMark sends nothing off the box. The render path needs no network, so the box can run fully air-gapped once installed; that isolation is your network’s control. When the render sandbox is on (an operator setting; ask for it to be confirmed at handover), the render step itself also runs with no network access.
DocMark-cloud zero-egress = no PII to the execution plane. The control plane holds identity; the worker that renders never receives the user’s email or principal — only the job content it must render.
Client cloud is deliberately not called zero-egress — content reaches your chosen LLM, but only within your own tenant and account. Its AI-authored watched-folder decks are labelled contract-protected-api.

Synthetic presenter media (a presenter likeness or a cloned voice) — when it ships, it is consent-gated and disclosed per client brand. No likeness or voice is generated without explicit, recorded enablement, and outputs carry a disclosure. (Branded narrated video is already live; it uses a generic synthetic voice with no human likeness or voice cloning.)

The verification gate withholds. A deliverable that fails the gate is not delivered — it is withheld with findings. What the gate checks for each format, and where it does not apply, is set out in point 7 of the security page: on a self-managed install the deck gate is on by default and turning it off takes a second, deliberate override as well (named at handover), a partly blank video can ship, and the optional agentic engine’s own visual review is not verified by DocMark — which is why its decks are gated like every other engine’s.
Pixels come from deterministic code. Rendering is tested code; the model authors content. The exception is the optional agentic engine (self-managed deployments only, off unless an operator turns it on), where an AI agent writes and runs each deck’s build script with shell access.
No PII to the execution plane (DocMark cloud) is enforced by construction, not by policy.
On-prem needs no internet with the standard engines. The render path makes no network calls; when the render sandbox is on, the render step has no network access. A public bind is refused unless the sandbox is explicitly configured. An AI engine, if an operator enables one, calls its model provider.
Decks sent in through the watched folder are labelled with their egress class, written by the pipeline itself. Web-app, /v1 API and MCP jobs are not labelled.

Last updated: 22 September 2026 (revised the same day: verification-gate scope and on-prem AI engines).