diff --git a/examples/retrieval/README.md b/examples/retrieval/README.md
new file mode 100644
index 0000000..b09e4f4
--- /dev/null
+++ b/examples/retrieval/README.md
@@ -0,0 +1,31 @@
+# Retrieval example
+
+Offline example that exercises the retrieval scorers (`recall_at_k`, `mrr`, `ndcg`) end to
+end on a small labelled corpus and two deterministic retrievers. No models or network.
+
+## Files
+
+- `cases.jsonl`: 20 queries with graded relevance judgements (0 to 3).
+- `retrievers.py`: two systems.
+  - `good_retriever` ranks the corpus by term overlap with the query.
+  - `noisy_retriever` is the same ranking with a deliberate regression: on queries with an
+    even number of characters, the top hit drops to rank 5.
+- `suite.yaml`: recall@5, MRR@10 and nDCG@5, gating on nDCG@5.
+
+## Run
+
+```bash
+for v in good noisy; do
+  uv run evalproof run examples/retrieval/suite.yaml \
+    --system "python:retrievers:${v}_retriever" --out "runs/retrieval-$v.json" --quiet
+done
+uv run evalproof compare runs/retrieval-good.json runs/retrieval-noisy.json \
+  --suite examples/retrieval/suite.yaml --gate; echo "exit $?"   # 1: real regression
+```
+
+## What it shows
+
+Moving the top hit to rank 5 leaves recall@5 unchanged at 0.775, because recall ignores
+order inside the cut-off. MRR falls from 1.000 to 0.642 and the nDCG@5 pass rate from
+100% to 55%: nine of twenty queries regress, none improve, and the exact McNemar test gives
+p = 0.0039, so the gate fails the change with exit 1.
diff --git a/examples/retrieval/cases.jsonl b/examples/retrieval/cases.jsonl
new file mode 100644
index 0000000..ac600c7
--- /dev/null
+++ b/examples/retrieval/cases.jsonl
@@ -0,0 +1,20 @@
+{"id": "q1", "input": "machine learning tutorial", "relevant": {"doc1": 3, "doc2": 2, "doc5": 1}}
+{"id": "q2", "input": "python data science", "relevant": {"doc3": 3, "doc4": 1}}
+{"id": "q3", "input": "neural networks deep learning", "relevant": {"doc1": 2, "doc6": 3}}
+{"id": "q4", "input": "web development javascript", "relevant": {"doc7": 3, "doc8": 2}}
+{"id": "q5", "input": "database sql query", "relevant": {"doc9": 3, "doc10": 1}}
+{"id": "q6", "input": "cloud computing aws", "relevant": {"doc11": 3, "doc12": 2}}
+{"id": "q7", "input": "mobile app android", "relevant": {"doc13": 3}}
+{"id": "q8", "input": "security encryption", "relevant": {"doc14": 3, "doc15": 2}}
+{"id": "q9", "input": "devops ci cd", "relevant": {"doc16": 3, "doc17": 1}}
+{"id": "q10", "input": "natural language processing", "relevant": {"doc18": 3, "doc1": 1}}
+{"id": "q11", "input": "computer vision opencv", "relevant": {"doc19": 3}}
+{"id": "q12", "input": "reinforcement learning", "relevant": {"doc20": 3, "doc6": 2}}
+{"id": "q13", "input": "frontend react", "relevant": {"doc7": 2, "doc21": 3}}
+{"id": "q14", "input": "backend nodejs", "relevant": {"doc22": 3, "doc8": 1}}
+{"id": "q15", "input": "data visualization matplotlib", "relevant": {"doc23": 3, "doc3": 1}}
+{"id": "q16", "input": "kubernetes docker", "relevant": {"doc24": 3, "doc16": 2}}
+{"id": "q17", "input": "git version control", "relevant": {"doc25": 3}}
+{"id": "q18", "input": "api design rest", "relevant": {"doc26": 3, "doc22": 1}}
+{"id": "q19", "input": "testing pytest", "relevant": {"doc27": 3, "doc17": 2}}
+{"id": "q20", "input": "open source contribution", "relevant": {"doc28": 3, "doc25": 1}}
diff --git a/examples/retrieval/retrievers.py b/examples/retrieval/retrievers.py
new file mode 100644
index 0000000..37d9292
--- /dev/null
+++ b/examples/retrieval/retrievers.py
@@ -0,0 +1,62 @@
+"""Simple deterministic retrievers for the retrieval example.
+
+Documents are hardcoded for reproducibility. Relevance is graded 0-3.
+"""
+
+from __future__ import annotations
+
+# Tiny corpus: id -> text
+CORPUS = {
+    "doc1": "introduction to machine learning and neural networks",
+    "doc2": "machine learning basics tutorial for beginners",
+    "doc3": "python for data science and pandas",
+    "doc4": "data analysis with python",
+    "doc5": "supervised learning overview",
+    "doc6": "deep learning neural networks tutorial",
+    "doc7": "javascript web development fundamentals",
+    "doc8": "frontend javascript and html",
+    "doc9": "sql database queries tutorial",
+    "doc10": "relational databases explained",
+    "doc11": "aws cloud computing services",
+    "doc12": "introduction to cloud infrastructure",
+    "doc13": "android mobile app development",
+    "doc14": "encryption and security best practices",
+    "doc15": "cryptography basics",
+    "doc16": "devops ci cd pipelines",
+    "doc17": "continuous integration tools",
+    "doc18": "natural language processing techniques",
+    "doc19": "computer vision with opencv",
+    "doc20": "reinforcement learning algorithms",
+    "doc21": "react frontend framework",
+    "doc22": "nodejs backend development",
+    "doc23": "data visualization with matplotlib",
+    "doc24": "kubernetes container orchestration",
+    "doc25": "git version control tutorial",
+    "doc26": "rest api design principles",
+    "doc27": "pytest testing framework",
+    "doc28": "contributing to open source projects",
+}
+
+
+def _score(query: str, doc: str) -> int:
+    q_terms = set(query.lower().split())
+    d_terms = set(doc.lower().split())
+    return len(q_terms & d_terms)
+
+
+def good_retriever(query: str) -> dict:
+    """Rank all documents by term overlap with the query."""
+    ranked = sorted(CORPUS.items(), key=lambda item: -_score(query, item[1]))
+    retrieved = [doc_id for doc_id, _ in ranked[:10]]
+    return {"output": f"Retrieved {len(retrieved)} docs", "retrieved": retrieved}
+
+
+def noisy_retriever(query: str) -> dict:
+    """good_retriever with a regression: on even-length queries the top hit drops to rank 5."""
+    ranked = sorted(CORPUS.items(), key=lambda item: -_score(query, item[1]))
+    retrieved = [doc_id for doc_id, _ in ranked[:10]]
+    # Deterministic "noise": if query length even, move first doc to position 5
+    if len(query) % 2 == 0 and len(retrieved) > 5:
+        top = retrieved.pop(0)
+        retrieved.insert(4, top)
+    return {"output": f"Retrieved {len(retrieved)} docs (noisy)", "retrieved": retrieved}
diff --git a/examples/retrieval/suite.yaml b/examples/retrieval/suite.yaml
new file mode 100644
index 0000000..ede8baf
--- /dev/null
+++ b/examples/retrieval/suite.yaml
@@ -0,0 +1,26 @@
+# Retrieval example: exercises recall@k, MRR and nDCG with a tiny corpus.
+# Offline, deterministic, suitable for CI.
+name: retrieval-example
+dataset: cases.jsonl
+
+system:
+  target: python:retrievers:good_retriever
+
+scorers:
+  - name: recall
+    type: recall_at_k
+    params: {k: 5}
+  - name: mrr
+    type: mrr
+    params: {k: 10}
+  - name: ndcg
+    type: ndcg
+    params: {k: 5, pass_threshold: 0.5}
+
+gate:
+  metric: ndcg
+  max_drop: 0.1
+  alpha: 0.05
+  min_cases: 15
+
+cache_dir: .evalproof/cache