diff --git a/.github/workflows/sync-upstream.yml b/.github/workflows/sync-upstream.yml
index 5d077fd..ae64bee 100644
--- a/.github/workflows/sync-upstream.yml
+++ b/.github/workflows/sync-upstream.yml
@@ -73,6 +73,7 @@ jobs:
           existing="$(gh pr list --head "$branch" --state open --json url --jq '.[0].url')"
           if [[ -n "$existing" ]]; then
             echo "Upstream PR already open: $existing"
+            gh pr merge "$existing" --auto --squash
             exit 0
           fi
           git config user.name "github-actions[bot]"
@@ -81,8 +82,9 @@ jobs:
           git add value-map/SKILL.md UPSTREAM.json
           git commit -m "Sync Value Map from Intelligence Flow $short"
           git push --set-upstream origin "$branch"
-          gh pr create \
+          pr_url="$(gh pr create \
             --base main \
             --head "$branch" \
             --title "Sync Value Map from Intelligence Flow $short" \
-            --body "Generated from operatorstack/intelligence-flow@$SOURCE_COMMIT. Review the operator contract, provenance, and compatibility checks before merging."
+            --body "Generated from operatorstack/intelligence-flow@$SOURCE_COMMIT. Review the operator contract, provenance, and compatibility checks before merging.")"
+          gh pr merge "$pr_url" --auto --squash
diff --git a/README.md b/README.md
index d3f8802..206b060 100644
--- a/README.md
+++ b/README.md
@@ -118,6 +118,18 @@ Value Map uses deterministic checks for the parts that should not depend on tast
 
 An LLM still performs semantic projection, so the result is not a scientific measurement of whether an idea or sender is good. It is a traceable, falsifiable interpretation. The final judgment remains yours.
 
+## Testing and benchmarks
+
+Value Map is checked with deterministic lineage and evidence gates,
+metamorphic tests, staged smoke cases, and a live inverse benchmark that holds
+one external message constant while changing the underlying code truth. The
+latest three-call Gemini Flash run passed every inverse gate and both negative
+controls. This validates the operator on a controlled fixture; it is not a
+claim of general accuracy across products or conversations.
+
+See the [benchmark record](https://github.com/operatorstack/intelligence-flow/blob/main/labs/14-product-value-projection/BENCHMARK.md)
+and its reproducible evaluator in Intelligence Flow.
+
 ## Source and releases
 
 The canonical operator and evaluation work lives in [Intelligence Flow](https://github.com/operatorstack/intelligence-flow/tree/main/labs/14-product-value-projection). This repository owns the public product surface, installation, compatibility checks, and releases.