diff --git a/docs/decisions.md b/docs/decisions.md
index 5d442ec..a965ecc 100644
--- a/docs/decisions.md
+++ b/docs/decisions.md
@@ -75,3 +75,11 @@ entry that supersedes it, never by editing an old one.
 - Why: psql substitutes a variable into the statement text, so ALTER ROLE ... PASSWORD with the plaintext puts the password in the server's log under log_statement ddl or all, or on a failing statement; PostgreSQL stores a pre-hashed verifier as it is
 - Applies to: engine/ansible/roles/cloudfall_postgresql/tasks/owner_password.yml, sdk/src/cloudfall/secrets.py, engine/src/cloudfall_engine/cli.py
 - Enforced by: a regression test that the staged file holds a verifier and never the password
+
+## D-10: Components are isolated by the systemd sandbox, and separate servers are recommended, not required
+
+- Decided: 2026-11-02, by the maintainer; reasoning in docs/decisions/2026-11-02-isolation-without-containers.md
+- Rule: On every server, components are kept apart by each unit's systemd sandbox (own user, ProtectSystem, ProtectHome, PrivateTmp, ProtectProc, SystemCallFilter, capability and address-family limits, cgroup limits), a database role per component, secrets through systemd credentials, keys scoped by environment and a non-root deploy user; runners are ephemeral sandboxed units, denied production directories, network peers and keys when they share a server with production; separate servers for different trust levels (production and staging, build or CI and production, different owners) are recommended, and a team may place everything on one server, several independent projects included, by declaring the accepted risk in the plan; Cloudfall warns about an undeclared placement of two trust levels on one server, and warns more strongly when they belong to different owners; Cloudfall adds no container runtime and no hypervisor of its own
+- Why: A Unix user alone keeps only files and signals apart; containers use the same kernel mechanisms systemd exposes and add no boundary, only a runtime to operate; a separate kernel is the one stronger boundary and comes from a separate server, but requiring it would rule out the one-server teams Cloudfall serves, and a team can accept that risk for its own projects, not for other owners
+- Applies to: engine/ansible/roles/cloudfall_deploy/templates/component.service.j2, engine/ansible/roles/cloudfall_access, server placement in plans, runner and CI setup
+- Enforced by: review
diff --git a/docs/decisions/2026-11-02-isolation-without-containers.md b/docs/decisions/2026-11-02-isolation-without-containers.md
new file mode 100644
index 0000000..7e4795a
--- /dev/null
+++ b/docs/decisions/2026-11-02-isolation-without-containers.md
@@ -0,0 +1,155 @@
+# Decision: isolation without containers or self-run VMs
+
+Date: 2026-11-02. Status: decided, revisit on the triggers listed below.
+
+## Question
+
+How does Cloudfall keep components, runners and environments apart, and does
+that need containers or virtual machines it runs itself?
+
+## Decision
+
+No containers and no hypervisor of Cloudfall's own. Isolation has two
+boundaries:
+
+- **Inside a server**, the boundary is the systemd sandbox of each unit, a
+  database role per component, and cgroup limits. This is required on every
+  server
+- **Between trust levels**, the stronger boundary is a separate server: its
+  own kernel. A VPS counts as a server; its hypervisor belongs to the
+  provider, not to Cloudfall. This is recommended, not required
+
+A team may run everything on one server, several independent projects
+included, when it accepts the shared-kernel risk. The acceptance is declared
+in the plan, never implied: Cloudfall warns about a placement that puts two
+trust levels on one server unless the plan declares it.
+
+A separate Unix user per component is necessary but is not a boundary on its
+own.
+
+## Why a Unix user is not enough
+
+A user keeps files and signals apart. It does not stop:
+
+- a local privilege escalation in the shared kernel, which reaches every
+  component on the host
+- a database, cache or queue on localhost that trusts every local user, or
+  that components share one account of
+- membership in the `docker` group or a broad sudo rule, either of which is
+  root
+- a runner that executes code from a pull request or a compromised
+  dependency, which gets every key the runner holds; the keys, not the uid,
+  set the blast radius
+- reading other processes' command lines in `/proc`, world-readable home
+  directories, and one component exhausting memory, disk or processes for
+  the rest
+
+## Why not containers
+
+Docker and Podman isolate with the same kernel mechanisms systemd exposes
+directly: namespaces, cgroups, seccomp and capabilities. A container shares
+the kernel exactly as a sandboxed unit does, so it adds no boundary. What it
+adds is packaging (images, independent library versions) and a runtime to
+operate. Cloudfall's base is Debian and systemd
+([`2026-09-14-no-nixos-base.md`](2026-09-14-no-nixos-base.md)), container
+orchestration is a non-goal ([`../product.md`](../product.md), N-1), and the
+exit test is that any Linux admin can take over a server with Debian
+knowledge.
+
+## Why not self-run VMs
+
+A hypervisor on Cloudfall's hosts (Proxmox, KVM, Firecracker) is a second
+infrastructure layer to provision, patch, back up and explain to whoever
+takes over. A separate small server gives the same kernel boundary for a
+few euros a month and fits the model Cloudfall already manages.
+
+## The target model
+
+The first four parts are required on every server, one server or many. The
+last is a recommendation.
+
+**Each unit is sandboxed by systemd.** The component unit gets, by default:
+
+- its own user (or `DynamicUser=`), `NoNewPrivileges=` and an empty
+  `CapabilityBoundingSet=`
+- `ProtectSystem=strict`, `ProtectHome=`, `PrivateTmp=`, `PrivateDevices=`,
+  writes only to `ReadWritePaths=` and `StateDirectory=`
+- `ProtectProc=invisible`, `SystemCallFilter=@system-service`,
+  `RestrictNamespaces=`, `RestrictAddressFamilies=`
+- `IPAddressAllow=` and `IPAddressDeny=` where the component's peers are
+  known
+- `MemoryMax=`, `CPUQuota=`, `TasksMax=` from the component's declared
+  envelope ([`../fleet-goals.md`](../fleet-goals.md), goal 8)
+- secrets through `LoadCredentialEncrypted=` rather than `EnvironmentFile=`,
+  so a value is readable by that unit only and never sits in the process
+  environment
+
+`systemd-analyze security` scores a unit; a score threshold can become a
+deploy gate.
+
+**Shared services authenticate per component.** Each component gets its own
+PostgreSQL role, preferably through `peer` authentication over the Unix
+socket, so the unit's user is the role and no password exists. A service
+without authentication is never shared between components of different
+owners.
+
+**Runners are ephemeral and sandboxed.** A runner executes untrusted code.
+It runs one job per transient unit started with `systemd-run` under the same
+sandbox and `DynamicUser=`, its state removed after the job, and its network
+open only for the dependency step. On a server that also runs production
+components, the runner's unit is also denied the production components'
+directories and network peers (`InaccessiblePaths=`, `IPAddressDeny=`) and
+holds no production keys. This is the most exposed point of a one-server
+placement.
+
+**Access is scoped by environment.** A runner holds the SSH and age keys of
+one environment ([`../ci-deploy-guide.md`](../ci-deploy-guide.md)). Deploys
+log in as a dedicated user with sudo limited to the commands a deploy runs,
+not as root.
+
+**Separate servers separate trust levels (recommended).** Where the budget
+allows, these do not share a kernel:
+
+- production and staging
+- build and CI, and production; production servers then receive built
+  artifacts only
+- components of different customers or owners
+
+A team can only accept risk for its own projects. Placing components of
+different owners on one server puts the risk on owners who did not accept
+it, so Cloudfall keeps a separate, stronger warning for it. Cloudfall cannot
+verify ownership, so this stays a warning, not a refusal.
+
+## Where Cloudfall is today
+
+- Each component already runs as its own application user with
+  `NoNewPrivileges=true` and `PrivateTmp=true`
+  (`engine/ansible/roles/cloudfall_deploy/templates/component.service.j2`);
+  the rest of the sandbox is missing
+- SSH keys are scoped by environment, but the baseline installs them into
+  root's `authorized_keys`, so a runner's key is root on every server of its
+  environment
+- Secrets reach units through `EnvironmentFile=`
+- Resource envelopes are not declared or enforced yet
+
+## Triggers to revisit
+
+- **A customer needs untrusted code from several tenants on one server**,
+  where a separate server per tenant does not pay; then a stronger sandbox
+  (gVisor, microVMs) is the question, not containers
+- **A required workload ships only as a container image** with no reasonable
+  Debian install path
+- **systemd sandboxing proves insufficient** for a concrete attack that a
+  container runtime would have stopped
+
+## Consequences
+
+- The component unit template gains the sandbox directives above as
+  defaults, with per-component opt-outs that are declared, not silent
+- No container runtime, image build or registry appears in Cloudfall's
+  engine or docs as the way to isolate components
+- One server is a supported placement, not a degraded one; the required
+  parts above carry the whole isolation there
+- A plan that puts two trust levels on one server declares it as an
+  accepted risk; Cloudfall warns when it is undeclared, and warns again,
+  more strongly, when the trust levels belong to different owners