diff --git a/docs/adr/0005-trusted-proxy-client-ip.md b/docs/adr/0005-trusted-proxy-client-ip.md
new file mode 100644
index 0000000..fb90051
--- /dev/null
+++ b/docs/adr/0005-trusted-proxy-client-ip.md
@@ -0,0 +1,39 @@
+# Client-IP resolution is gated on a trusted-proxy set
+
+The two auth 401 log lines must report the real client IP. The API runs behind a
+TLS-terminating reverse proxy, so the socket peer (`r.RemoteAddr`) is always that
+proxy, not the caller. We resolve the client address from forwarding headers
+(`X-Forwarded-For` / `X-Real-IP`) **only when the TCP peer is in a configured
+trusted-proxy set** (`TRUSTED_PROXIES`, comma-separated CIDRs; default empty =
+trust nothing). For a trusted peer we take the right-most `X-Forwarded-For` entry
+not in the trusted set, fall back to `X-Real-IP`, then the socket peer.
+
+The trusted set contains **only the immediate upstream proxy the API actually
+peers with** — not any edge CDN in front of it. The edge already normalizes the
+connecting client into the forwarding headers, and the API never connects to the
+CDN directly, so trusting the CDN's ranges would be both unnecessary and a wider
+trust surface than required.
+
+## Considered Options
+
+- **Honor forwarding headers unconditionally** — rejected: forwarding headers are
+  caller-suppliable and must only be trusted when they arrive from a known proxy;
+  honoring them from any peer would let the logged client address be forged.
+- **Read `X-Real-IP` or the left-most `X-Forwarded-For` entry without a trust
+  gate** — rejected: both are caller-suppliable; the left-most XFF entry is the
+  spoofable one.
+- **Peer-gated right-most-untrusted XFF** — chosen: the only variant that resists
+  a forged client address.
+
+## Consequences
+
+Client IPs are currently used for diagnostic logging only. The trust gate is built
+now so that any future use depending on a trustworthy client address (e.g. the
+per-key rate-limiting noted in PRD #112) can rely on it without rework. A
+non-empty but malformed `TRUSTED_PROXIES` is fatal at startup; empty/unset trusts
+nothing and preserves the prior (proxy-address) logging behavior.
+
+The malformed-XFF → `X-Real-IP` fallback is an **availability-biased** choice for
+diagnostic logging, not a security control: both headers originate from the same
+trusted proxy, and the result is not used for authorization or rate limiting. It
+is documented here so it is not later "tightened" into a stricter rejection.