diff --git a/docs/decisions/020-firewall.md b/docs/decisions/020-firewall.md
new file mode 100644
index 0000000..c014bcc
--- /dev/null
+++ b/docs/decisions/020-firewall.md
@@ -0,0 +1,133 @@
+# ADR-020 — Firewall strategy: two-layer model with a shared service catalog
+
+## Status
+
+Accepted (2026-06-06). Resolves TODO 3.5 ("Decide the firewall strategy — which
+firewall, ruleset, per-host vs central").
+
+**Strategy ADR.** It pins the architecture and each layer's responsibilities; the
+detailed builds are separate follow-up efforts (see *Scope*).
+
+## Context
+
+boma needs a firewall strategy that is predictable, declarative, and defends the stated
+threat model — opportunistic external, lateral movement / blast radius, operator/agent
+error (ADR-002). The pieces were already committed across other ADRs (`nftables`
+default-deny on hosts — ADR-002; OPNsense at the perimeter — ADR-007; Docker with
+`iptables: false` — ADR-004), but nothing tied them together: which layer owns what,
+where firewall intent is declared, and how the layers stay consistent. Without that,
+ports drift open ad-hoc and "per-host vs central" stays unanswered.
+
+## Decision
+
+### Two layers, distinct jobs
+
+**OPNsense — perimeter + inter-VLAN.** Owns the WAN edge and all policy *between zones*:
+`lan`/`iot`/`guest` → `srv`, `mgmt` access, and the per-VLAN egress rules (ADR-007). It
+is **structurally blind to intra-`srv` traffic** — services share the switched `srv`
+subnet (VLAN 20), which never reaches the gateway.
+
+**Host nftables — host-local + east-west within `srv`** (in the `base` role, every VM):
+
+- **Default-deny inbound**; allow loopback + established/related.
+- **East-west allowlist**: a service host accepts a connection only from declared
+  sources (e.g. the reverse proxy, a named peer) — the lateral-movement control OPNsense
+  cannot provide.
+- **Permissive egress**: allow outbound + established/related; per-VLAN egress
+  restriction stays at OPNsense (ADR-007). Host-level egress allowlisting is
+  high-friction (every DNS/NTP/update/registry/webhook must be enumerated) for limited
+  added benefit once the VLAN already bounds where a host can go.
+- **Docker**: daemon runs with `"iptables": false`; nftables owns all filtering,
+  including container traffic (ADR-004).
+- **Guaranteed management plane**: loopback, established/related, and `wt0` (NetBird,
+  ADR-016) for SSH + Ansible are always allowed, independent of the catalog, applied
+  atomically — a malformed or empty catalog can never lock out management. (ADR-016: SSH
+  is allowed only on `wt0`.)
+
+So "per-host vs central" is answered: **both**, with clear ownership.
+
+### Single source of truth — a shared service catalog
+
+A central, declarative **service catalog** in `group_vars/` is the one source of truth
+for firewall intent (aligning with ADR-002's "port definitions live in `group_vars/`",
+and keeping connectivity *topology* in inventory rather than in any one self-contained
+service role — ADR-004). Each entry describes a service's **ingress**:
+
+```yaml
+photoprism:
+  ingress:
+    - { from: reverse_proxy, port: 2342, proto: tcp }
+reverse_proxy:
+  ingress:
+    - { from: lan, port: 443, proto: tcp }
+```
+
+`from` is **symbolic**, resolved at render time: a host/group → IP(s) from inventory; a
+role (`reverse_proxy`) → the host(s) filling it; a VLAN/zone (`lan`) → the subnet from
+the ADR-007 table. This keeps the catalog readable and resilient to IP changes.
+
+### Each layer renders only its own slice
+
+| Ingress rule | Host nftables | OPNsense |
+|---|---|---|
+| `from: reverse_proxy` (a `srv` peer) | allow proxy IP → port | — (intra-`srv`, invisible) |
+| `from: lan` (cross-VLAN) | allow `lan` subnet → port | allow `lan` → host:port |
+
+The dominant pattern falls out naturally: most services are **proxied** — their only
+ingress is `from: reverse_proxy`, and users reach them through the reverse proxy, which
+alone carries `from: lan, port: 443` (matches "services sit behind the reverse proxy
+with authentication", ADR-002).
+
+This was chosen over a single connectivity-model-generates-both (too much machinery,
+tight coupling of two very different rule domains) and over fully independent per-layer
+declarations (real drift risk).
+
+### OPNsense automation — owned here, mechanism deferred
+
+OPNsense is Ansible-managed (CLAUDE.md: "OPNsense is entirely Ansible; no Terraform
+OPNsense provider"). It renders the cross-VLAN slice of the catalog plus the static
+ADR-007 facts. The **how** — config-XML templating vs the OPNsense API vs a plugin — is
+deferred to the OPNsense-as-code follow-up spec. Recorded as an explicit open
+sub-decision.
+
+## Guardrails
+
+- **The catalog is authoritative.** If a port is not in the catalog, it does not exist —
+  hardening the existing rule "never open a firewall port ad-hoc on a host" (ADR-002).
+- **The `firewall` tag** (ADR-019) marks firewall tasks; `--tags firewall` re-renders
+  rules.
+- **Drift detection (aspiration).** A deterministic check — in the spirit of
+  `scripts/check-tags.py` — comparing each host's live `nft` ruleset / listening ports
+  against the catalog and flagging anything undeclared. Ties to TODO 8.5
+  (`/security-review`). Not necessarily built first.
+
+## Consequences
+
+- Lateral movement within `srv` is constrained — the gap OPNsense structurally can't
+  close.
+- One declarative catalog → no ad-hoc ports and no cross-layer drift on shared facts
+  (ports, IPs, sources).
+- Cost: the catalog + render-per-layer machinery must be built and maintained; east-west
+  allowlisting adds per-service ingress declarations (mitigated by proxied-by-default,
+  which keeps most entries to a single line).
+
+## Scope
+
+**Decided here:** the two-layer model and responsibilities; host nftables = default-deny
+inbound + east-west allowlist + permissive egress + guaranteed management plane + Docker
+`iptables:false`; the shared `group_vars` catalog as single source of truth with
+symbolic sources; each layer renders its own slice; the no-ad-hoc-ports guardrail.
+
+**Deferred to follow-up specs (each its own brainstorm → plan):**
+
+1. **Host nftables implementation** in `base` — catalog schema, nftables template,
+   Docker `iptables:false` integration, fail-safe ordering, Molecule tests. The natural
+   next spec.
+2. **OPNsense-as-code** — tooling mechanism + cross-VLAN rule rendering.
+3. **Drift-detection check** — if/when built.
+
+## Related
+
+ADR-002 (security baseline: nftables default-deny, fail2ban, blast radius),
+ADR-004 (Docker model: `iptables:false`), ADR-007 (network topology, VLANs, OPNsense,
+per-VLAN egress), ADR-016 (NetBird mesh: SSH on `wt0` only), ADR-019 (`firewall` tag).