diff --git a/.github/scripts/replay-ui-summary-test.sh b/.github/scripts/replay-ui-summary-test.sh
new file mode 100755
index 0000000..b75e988
--- /dev/null
+++ b/.github/scripts/replay-ui-summary-test.sh
@@ -0,0 +1,125 @@
+#!/usr/bin/env bash
+# Drives replay-ui-summary.sh over the traces in testdata/ and checks the
+# summary it renders. Run under the flags GitHub Actions uses for a `run:`
+# block, because that is where a swallowed failure hides.
+#
+# testdata/replay-ui-real-run.jsonl is the first 10 steps of the dogfood run in
+# actions run 31873049857 on master, with the per-step `hierarchy` dumps and the
+# rowElements/tabElements extractors removed so the file stays readable. Nothing
+# else was touched. That run was green, and badgeCountMatchesThePanel judged
+# nothing in all 80 of its steps. The other two traces are written by hand.
+set -euo pipefail
+
+here="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+script="$here/replay-ui-summary.sh"
+testdata="$here/testdata"
+work="$(mktemp -d)"
+trap 'rm -rf "$work"' EXIT
+
+failed=0
+rendered=""
+status=0
+
+summarise() { # <case name> <output dir>
+  rendered="$work/$1.md"
+  : > "$rendered"
+  status=0
+  GITHUB_STEP_SUMMARY="$rendered" SEED=3 MAX_STEPS=80 \
+    bash -eo pipefail "$script" "$2" >/dev/null 2>"$work/$1.err" || status=$?
+}
+
+fail() {
+  echo "FAIL: $*" >&2
+  failed=1
+}
+
+expect_status() { # <want> <case>
+  [ "$status" = "$1" ] || fail "$2: exit $status, want $1"
+}
+
+expect_line() { # <line> <case>
+  grep -qxF -- "$1" "$rendered" || fail "$2: summary has no line '$1'"
+}
+
+expect_absent() { # <pattern> <case>
+  if grep -q -- "$1" "$rendered"; then fail "$2: summary should not mention '$1'"; fi
+}
+
+plant() { # <case name> <fixture> -> echoes the output dir
+  local dir="$work/$1/runs/20260815-075347"
+  mkdir -p "$dir"
+  cp "$testdata/$2" "$dir/trace.jsonl"
+  echo "$work/$1/runs"
+}
+
+# A green run of the real thing. Every count here was measured, not chosen.
+summarise real "$(plant real replay-ui-real-run.jsonl)"
+expect_status 0 real
+expect_line "- 10 steps recorded, 10 verified, 0 with violations" real
+expect_line "| selectedStepIsInRange | 10 | 0 |" real
+expect_line "| exactlyOneStepIsSelected | 10 | 0 |" real
+expect_line "| stepCountMatchesTheList | 10 | 0 |" real
+expect_line "| screenshotShowsTheSelectedStep | 10 | 0 |" real
+expect_line "| switchingTabsKeepsTheStep | 2 | 8 |" real
+expect_line "| badgeCountMatchesThePanel | **0** | 10 |" real
+expect_line "- \`badgeCountMatchesThePanel\` judged nothing on this run: no step had a violations badge on screen" real
+expect_absent "checked nothing" real
+
+# Every property judged at least once, including the one the real run never
+# reached. Without this the counts above are consistent with a guard that can
+# only ever return zero.
+summarise every "$(plant every replay-ui-every-property.jsonl)"
+expect_status 0 every
+expect_line "- 4 steps recorded, 3 verified, 0 with violations" every
+expect_line "| noUncaughtExceptions | 3 | 0 |" every
+expect_line "| screenshotShowsTheSelectedStep | 3 | 0 |" every
+expect_line "| switchingTabsKeepsTheStep | 2 | 1 |" every
+expect_line "| badgeCountMatchesThePanel | 1 | 2 |" every
+
+# The fuzzer sat on the run list: the step page never rendered, so nothing was
+# ever compared. This is the run that used to pass.
+summarise blind "$(plant blind replay-ui-nothing-rendered.jsonl)"
+expect_status 1 blind
+expect_line "| selectedStepIsInRange | **0** | 4 |" blind
+expect_line "| badgeCountMatchesThePanel | **0** | 4 |" blind
+expect_line "- \`exactlyOneStepIsSelected\` judged nothing on this run: no step had a row in the step list" blind
+grep -q "this run checked nothing" "$rendered" || fail "blind: no checked-nothing verdict"
+grep -q "the step page never rendered" "$work/blind.err" || fail "blind: nothing on stderr"
+
+# The run directory exists but the trace does not, and the glob matches nothing
+# at all. Both are the harness dying before it checked anything.
+mkdir -p "$work/empty-run/runs/20260815-075347"
+summarise empty-run "$work/empty-run/runs"
+expect_status 1 empty-run
+grep -q "no trace at" "$rendered" || fail "empty-run: no missing-trace line"
+
+summarise no-glob "$work/no-glob/runs"
+expect_status 1 no-glob
+grep -q "nothing under" "$rendered" || fail "no-glob: no missing-directory line"
+
+# A property this summary does not know about is a summary that silently counts
+# six of seven properties, which is the bug one level up.
+drifted="$(plant drift replay-ui-every-property.jsonl)"
+sed 's/badgeCountMatchesThePanel/badgeAgreesWithThePanel/g' \
+  "$testdata/replay-ui-every-property.jsonl" > "$drifted/20260815-075347/trace.jsonl"
+summarise drift "$drifted"
+expect_status 1 drift
+grep -q "these counts cannot be trusted" "$rendered" || fail "drift: no untrusted verdict"
+
+# Same for a reading the spec no longer declares: the count would quietly go to
+# zero and read as a UI that stopped rendering.
+sed 's/extract("violationBadges"/extract("violationCounters"/' \
+  "$here/../../replay-ui/sanderling/spec.ts" > "$work/renamed-spec.ts"
+rendered="$work/renamed.md"
+: > "$rendered"
+status=0
+GITHUB_STEP_SUMMARY="$rendered" SPEC="$work/renamed-spec.ts" \
+  bash -eo pipefail "$script" "$(plant renamed replay-ui-real-run.jsonl)" \
+  >/dev/null 2>&1 || status=$?
+expect_status 1 renamed
+grep -q "no longer declares violationBadges" "$rendered" || fail "renamed: no drift verdict"
+
+if [ "$failed" = 0 ]; then
+  echo "replay-ui-summary.sh: ok"
+fi
+exit "$failed"
diff --git a/Makefile b/Makefile
index f86a2cf..50532e5 100644
--- a/Makefile
+++ b/Makefile
@@ -29,7 +29,7 @@ WEB_DIST := replay-ui/dist
 
 GOLINES := $(shell $(GO) env GOPATH)/bin/golines
 
-.PHONY: bootstrap proto sidecar sanderling sanderling-web sanderling-android sanderling-ios install test test-go test-browser test-companion test-kotlin test-spec-api spec-typecheck web-test web-typecheck web-build web-dev replay-dev docs clean release-cli release-npm-dry fmt fmt-go fmt-kotlin fmt-ts fmt-swift
+.PHONY: bootstrap proto sidecar sanderling sanderling-web sanderling-android sanderling-ios install test test-go test-browser test-companion test-kotlin test-spec-api test-ci-scripts spec-typecheck web-test web-typecheck web-build web-dev replay-dev docs clean release-cli release-npm-dry fmt fmt-go fmt-kotlin fmt-ts fmt-swift
 
 bootstrap:
 	$(GO) mod download
@@ -117,7 +117,7 @@ fmt-ts:
 fmt-swift:
 	xcrun swift-format format -i -r companion/Sources
 
-test: test-go test-kotlin spec-typecheck test-spec-api web-typecheck web-test
+test: test-go test-kotlin spec-typecheck test-spec-api web-typecheck web-test test-ci-scripts
 
 test-go:
 	$(GO) test $(GO_PACKAGES)
@@ -141,6 +141,11 @@ test-companion: $(COMPANION_EMBED) $(RUNNER_EMBED)
 test-kotlin:
 	ANDROID_HOME=$(ANDROID_HOME) $(GRADLE) :sidecar:test
 
+# The CI scripts that read a trace and decide whether a green leg is
+# evidence. bash and python3 only, which is all a runner has.
+test-ci-scripts:
+	.github/scripts/replay-ui-summary-test.sh
+
 test-spec-api:
 	cd pkg/spec && npm test --silent