diff --git a/docs/4-implementation/swarm/review-classifier.md b/docs/4-implementation/swarm/review-classifier.md
index 8a73b4b89..95a526f45 100644
--- a/docs/4-implementation/swarm/review-classifier.md
+++ b/docs/4-implementation/swarm/review-classifier.md
@@ -735,11 +735,13 @@ The classifier logs each run's result to its journal, which ships to
 VictoriaLogs like every swarm service container's
 ([`services.md`](../../2-handbook/swarm/services.md#logs-victorialogs)).
 
-A run that ends without a verdict logs `review classification run failed` at
-`warn`, with the cause in `F_ERROR`. The Grafana alert
-`hh-review-classifier-run-failed` fires on that line and keeps firing for 30
-minutes after the last one. It fires on a failed run, so a dead key with no PRs
-arriving stays quiet until the next PR.
+A run that ends without a verdict logs a `warn` line: `review classification
+run failed` with the cause in `F_ERROR`. A request that doesn't decode, and a
+reply the classifier can't encode or publish, each log a line of their own
+that names the error.
+The Grafana alert `hh-review-classifier-run-failed` fires on any of these and
+keeps firing for 30 minutes after the last one. It fires on a failed run, so a
+dead key with no PRs arriving stays quiet until the next PR.
 
 ## Still open
 
diff --git a/nix/host-modules/swarm-grafana.nix b/nix/host-modules/swarm-grafana.nix
index b6af90bd4..d872f05c8 100644
--- a/nix/host-modules/swarm-grafana.nix
+++ b/nix/host-modules/swarm-grafana.nix
@@ -1627,18 +1627,20 @@ in
                       for = "0s";
                       summary = "{{ $labels.hive }}/{{ $labels.agent }} could not connect to the swarm queue and its harness exited; systemd starts it again 30s later";
                     })
-                    # `swarm-review-classifier` logs this at `warn` when a run
-                    # ends without a verdict (a rejected or unreadable API key,
-                    # a failed model call), so that PR has no classifier
-                    # review. The cause is in `F_ERROR`. The window is how long
-                    # the rule stays firing after the last failure.
+                    # A run that ends without a verdict leaves that PR with no
+                    # classifier review. `swarm-review-classifier` logs four
+                    # `warn` lines for it: `review classification run failed`
+                    # (a rejected or unreadable API key, a failed model call;
+                    # the cause is in `F_ERROR`), an undecodable request, and a
+                    # reply it could not encode or publish. The window is how
+                    # long the rule stays firing after the last one.
                     (logCountRule {
                       uid = "hh-review-classifier-run-failed";
                       title = "review classifier run failed";
                       windowSeconds = 1800;
-                      expr = ''_time:30m _msg:"review classification run failed" | stats count() as failures'';
+                      expr = ''_time:30m (_msg:"review classification run failed" OR _msg:"an undecodable try-out; answering with the error" OR _msg:"encoding a try-out's reply failed" OR _msg:"publishing a try-out's reply failed") | stats count() as failures'';
                       for = "0s";
-                      summary = "a review-classifier run ended without a verdict, so that PR has no classifier review; the cause is in the F_ERROR field of the log line";
+                      summary = "a review-classifier run ended without a verdict, so that PR has no classifier review; the log line names the cause (F_ERROR on a failed run)";
                     })
                     (logCountRule {
                       uid = "hh-swarm-agent-state-publish-skipped";
diff --git a/nix/module-eval/grafana.nix b/nix/module-eval/grafana.nix
index a205d0daa..97b7a99eb 100644
--- a/nix/module-eval/grafana.nix
+++ b/nix/module-eval/grafana.nix
@@ -860,9 +860,9 @@ let
         && lib.hasInfix "stats by (hive, agent)" expr;
     }
     {
-      # A failed classifier run is visible only as this log line, so a
-      # misspelled `_msg` provisions cleanly and never fires. The literal is
-      # checked against the logging call's source.
+      # A classifier run without a verdict is visible only as these log
+      # lines, so a misspelled `_msg` provisions cleanly and never fires.
+      # Each literal is checked against the logging calls' source.
       name = "the review-classifier failed-run alert rule matches the classifier's log message";
       ok =
         let
@@ -872,10 +872,15 @@ let
           rule = lib.findFirst (r: r.uid == "hh-review-classifier-run-failed") null rules;
           expr = (lib.head rule.data).model.expr;
           source = builtins.readFile ../../swarm-review-classifier/src/main.rs;
+          messages = [
+            "review classification run failed"
+            "an undecodable try-out; answering with the error"
+            "encoding a try-out's reply failed"
+            "publishing a try-out's reply failed"
+          ];
         in
         rule != null
-        && lib.hasInfix ''_msg:"review classification run failed"'' expr
-        && lib.hasInfix ''"review classification run failed"'' source;
+        && lib.all (m: lib.hasInfix "_msg:\"${m}\"" expr && lib.hasInfix "\"${m}\"" source) messages;
     }
     {
       # The hive-health alerts query metric names that hive-c0re spells in