diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index 911fcc0..938a8dd 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -11,6 +11,8 @@ on:
     branches: [main]
   pull_request:
   workflow_dispatch:
+  # platform.yml runs these first and only applies if they all pass.
+  workflow_call:
 
 permissions:
   contents: read
diff --git a/.github/workflows/platform.yml b/.github/workflows/platform.yml
index 96ff9e7..329860a 100644
--- a/.github/workflows/platform.yml
+++ b/.github/workflows/platform.yml
@@ -4,9 +4,11 @@
 # monitoring"). The service playbooks still run from an operator's machine.
 #
 #   pull request   -> dry run (--check --diff) against the boxes
-#   merge to main  -> the real run, once approved in the `production`
+#   merge to main  -> the repo checks (ci.yml), then, only if they all pass,
+#                     the real run, once approved in the `production`
 #                     environment (required reviewers)
-#   manual         -> either, optionally limited to some boxes
+#   manual         -> either, optionally limited to some boxes; apply also
+#                     waits for the checks
 #
 # Pull requests from forks get no secrets and are skipped.
 name: platform
@@ -50,11 +52,6 @@ on:
 permissions:
   contents: read
 
-# One run at a time against the boxes; never cancel one half-way.
-concurrency:
-  group: platform
-  cancel-in-progress: false
-
 env:
   # ansible.cfg sets the yaml callback, which no installed collection provides
   # here (see ci.yml).
@@ -62,11 +59,27 @@ env:
   ANSIBLE_FORCE_COLOR: "true"
 
 jobs:
+  # The same checks as ci.yml (lint, syntax, manifests, generated files,
+  # shellcheck), for this commit. The apply needs them, so a commit that fails
+  # them is never offered for approval. They run a second time on a merge,
+  # alongside ci.yml's own run: jobs cannot wait on another workflow.
+  checks:
+    if: >-
+      (github.event_name == 'push') ||
+      (github.event_name == 'workflow_dispatch' && inputs.mode == 'apply')
+    uses: ./.github/workflows/ci.yml
+
   check:
     if: >-
       (github.event_name == 'pull_request' && github.event.pull_request.head.repo.full_name == github.repository) ||
       (github.event_name == 'workflow_dispatch' && inputs.mode == 'check')
     runs-on: ubuntu-latest
+    # A newer push to the same branch replaces its dry run. Kept apart from the
+    # apply's group: GitHub keeps one pending run per group, so a shared group
+    # let a pull request's dry run cancel an apply waiting for approval.
+    concurrency:
+      group: platform-check-${{ github.ref }}
+      cancel-in-progress: true
     steps:
       - uses: actions/checkout@v4
         with:
@@ -88,10 +101,18 @@ jobs:
         run: scripts/ci-platform.sh check "$LIMIT"
 
   apply:
+    needs: checks
+    # No status function here, so GitHub adds success(): every job of the
+    # checks must have passed.
     if: >-
       (github.event_name == 'push') ||
       (github.event_name == 'workflow_dispatch' && inputs.mode == 'apply')
     runs-on: ubuntu-latest
+    # One apply at a time, never cancelled half-way. A newer merge waiting
+    # behind it replaces an older one still waiting: the latest main wins.
+    concurrency:
+      group: platform-apply
+      cancel-in-progress: false
     # Required reviewers: nothing reaches the boxes until one approves.
     environment: production
     steps:
diff --git a/README.md b/README.md
index 4cb292c..6399d70 100644
--- a/README.md
+++ b/README.md
@@ -1451,8 +1451,8 @@ does; the others still run from an operator's machine:
 | When | What |
 |---|---|
 | Pull request touching it | `--check --diff` against the boxes (forks get no secrets and are skipped) |
-| Merge to `main` | the real run, once a required reviewer approves it in the `production` environment |
-| Actions → platform → Run workflow | `check` or `apply`, optionally `--limit`ed |
+| Merge to `main` | the repo checks (`ci.yml`) for that commit, then, only if they all pass, the real run, once a required reviewer approves it in the `production` environment |
+| Actions → platform → Run workflow | `check` or `apply` (apply also waits for the checks), optionally `--limit`ed |
 
 It restarts nothing that serves traffic, so it runs every box in one batch: the
 hub registers the others using what they produced earlier in the same run.
@@ -1528,7 +1528,8 @@ same path a deploy takes), so what is checked is what would be applied:
 | `shellcheck scripts/*.sh` | the usual shell traps |
 
 `platform.yml` also has its own workflow, which does talk to the boxes (see
-"Argo CD and monitoring").
+"Argo CD and monitoring"). It calls this one (`ci.yml` is also a reusable
+workflow) and only applies once every job here has passed.
 
 The Cloudflare range check is a separate weekly workflow rather than part of
 this one, because it reaches `api.cloudflare.com`: an outage there should not