diff --git a/.env.example b/.env.example
index 14f8b8f..9ab55f2 100644
--- a/.env.example
+++ b/.env.example
@@ -37,10 +37,12 @@ TIMEZONE=America/New_York
 # =============================================================================
 
 # Healthchecks.io push URL for runner liveness. If set, runner-entrypoint.sh
-# pings <URL>/start at boot and <URL> every 60s while the container is alive.
-# Configure the HC check with period=1m and grace=5m so brief gaps between
-# ephemeral job cycles don't false-alert. Wire HC's integrations (Gotify,
-# email, etc.) to alert on missed pings. Leave unset to disable.
+# pings <URL> every 60s while the container is alive. Configure the HC check
+# with period=1m and grace=20m: the runner is ephemeral (restarts between every
+# job), so a tight grace false-alerts on normal restart gaps and trains you to
+# ignore it. A generous grace makes a DOWN mean a sustained outage. Wire HC's
+# integrations (Gotify, email, etc.) to alert on missed pings. Leave unset to
+# disable.
 #
 # Why this exists: a deprecated runner binary, a crash loop, or a hung
 # container won't emit pings — HC.io alerts after the grace period.
diff --git a/README.md b/README.md
index 42509fa..361c2e1 100644
--- a/README.md
+++ b/README.md
@@ -295,7 +295,7 @@ The runner has three layers of failure detection, each catching a different clas
 ### Setting up the Healthchecks.io heartbeat
 
 1. Create a check at <https://healthchecks.io/> (or your self-hosted HC instance)
-2. Configure: **period** = `1 minute`, **grace** = `5 minutes`. The grace covers brief gaps between ephemeral job cycles.
+2. Configure: **period** = `1 minute`, **grace** = `20 minutes`. The runner is ephemeral — the container restarts between every job — so the heartbeat has a gap on each cycle. A generous grace keeps those normal restart gaps (and brief self-recovering hiccups) from false-alerting, so a DOWN means a *sustained* outage worth paging on. (A too-tight grace causes alert fatigue: this check flipped 20× in 50 days on `grace=5m`, mostly self-recovered blips, which trained the alert to be ignored right before a real 18h outage.)
 3. Add HC's **Gotify** (or email, Slack, etc.) integration to the check
 4. Copy the check's ping URL into `.env`:
 
@@ -305,7 +305,7 @@ The runner has three layers of failure detection, each catching a different clas
 
 5. `sudo systemctl restart <service-name>` to pick up the change
 
-The runner pings `<URL>/start` at container boot and `<URL>` every 60 s while it's alive. A deprecated runner that crashes within 10 s of startup never gets to the periodic ping, so HC.io alerts after the 5-min grace.
+The runner pings `<URL>` every 60 s while it's alive. A runner that crashes within seconds of startup (deprecated binary, bad config, missing token) never gets to the periodic ping, so once the gap exceeds period + grace HC.io alerts. It does **not** ping `<URL>/start`: the container is ephemeral and dies before a run "completes", so a start signal never gets a matching success — it only adds noise.
 
 ## Troubleshooting
 
diff --git a/runner-entrypoint.sh b/runner-entrypoint.sh
index 03061d7..1379d3a 100755
--- a/runner-entrypoint.sh
+++ b/runner-entrypoint.sh
@@ -22,13 +22,17 @@ fi
 export RUNNER_TOKEN
 
 # Optional liveness heartbeat to Healthchecks.io.
-# When HEALTHCHECKS_URL is set, ping <URL>/start at boot and <URL> every 60 s
-# while this container is alive. Configure the HC check with period=1m and
-# grace=5m so brief gaps between ephemeral job cycles don't false-alert.
+# When HEALTHCHECKS_URL is set, ping <URL> every 60 s while this container is
+# alive. Configure the HC check with a grace period well above one job-cycle
+# restart (~20m) so brief gaps between ephemeral job cycles don't false-alert.
 # The background subshell dies with PID 1 when /entrypoint.sh exits, so we
 # don't need explicit cleanup.
+#
+# We intentionally do NOT ping <URL>/start: the container is ephemeral and dies
+# before a run "completes", so a start signal never gets a matching success and
+# only adds noise (it never registered on the HC side anyway). The plain 60s
+# success ping is the whole signal.
 if [ -n "${HEALTHCHECKS_URL:-}" ]; then
-  curl -fsS -m 10 -o /dev/null "$HEALTHCHECKS_URL/start" || true
   (
     while sleep 60; do
       curl -fsS -m 10 -o /dev/null "$HEALTHCHECKS_URL" || true