diff --git a/docs/health.md b/docs/health.md
new file mode 100644
index 000000000..1ea37c4de
--- /dev/null
+++ b/docs/health.md
@@ -0,0 +1,30 @@
+
+# Instance tracking for health checks
+
+To facilitate health checks to named workers through Hyperbahn, workers may
+advertise both the service name (e.g., `tcollector`) and a instance name (e.g.,
+`tcollector01-1-dc1`).
+The advertisement is broadcast to every affine Hyperbahn relay for that
+service.
+
+With partial affinity, each relay is responsible for a proportional subset of
+the advertising service workers.
+Relays determine the relay responsible for a worker by sorting the known relay
+and worker lists and projecting an index from one to the other.
+
+Every affine relay tracks an additional mapping from worker host:port to the
+last known worker identifier.
+
+Requests may have an additional `in` (instance name) transport header.
+Egress relays respect the `in` by forwarding to a relay responsible for
+maintaining an open connection.
+When a relay receives a request:
+
+1. If the relay is not an exit node for `sn` (service name), forwards to one
+that is.
+2. Looks up the host:port for the instance name (instance address).
+3. Using the known relays and known workers, discerns the set of host:port for
+relays responsible for connections to the instance address (instance relays).
+4. If the relay is not among instance relays, forwards to one that is.
+5. Forwards the request to the instance address.
+
diff --git a/docs/protocol.md b/docs/protocol.md
index c0f2dfe4c..66a109478 100644
--- a/docs/protocol.md
+++ b/docs/protocol.md
@@ -624,7 +624,8 @@ valid or not in a call req or res.  Following sections will elaborate on details
 | `re`  | Y   | N   | Retry Flags
 | `se`  | Y   | N   | Speculative Execution
 | `fd`  | Y   | Y   | Failure Domain
-| `sk`  | Y   | N   | Shard key
+| `sk`  | Y   | N   | Shard Key
+| `in`  | Y   | N   | Instance Name
 
 ### Transport Header `as` -- Arg Scheme
 
@@ -718,6 +719,16 @@ For example you may want to keep some in memory state, i.e. cache, aggregation.
 You can use read the `sk` and forward the call request to a specific process
 that has ownership for the shard key.
 
+### Transport Header `in` -- Instance Name
+
+An instance name is a unique identifier for an instance that may be more
+consistent than a host:port address, particularly if the instance delegates to
+the operating system to choose a port.
+Instances may advertise their instance name in addition to their service name.
+TChannel relays may respect the instance name header by forwarding to the exact
+host:port of the instance instead of forwarding to any instance with the
+request service name.
+
 ### A note on `host:port` header values
 
 While these `host:port` fields are indeed strings, the intention is to provide