diff --git a/README.md b/README.md
index 666ad7a..d8a906b 100644
--- a/README.md
+++ b/README.md
@@ -67,14 +67,42 @@ sparse-only modes are exposed too, so the benchmark can compare all three.
 python -m venv .venv && source .venv/bin/activate
 pip install -e ".[dev]"
 
-# Run the real benchmark (downloads the embedding models on first run).
-tenantq benchmark --embedder fastembed
+# Fastest first check (no model downloads; used by CI):
+pytest -q
+```
+
+Expected output (counts may grow as tests are added):
 
-# Fully offline (deterministic hash embedder, no downloads) — used by CI/tests:
+```
+........                                                                 [100%]
+```
+
+```bash
+# Fully offline benchmark (deterministic hash embedder, no downloads):
 tenantq benchmark --embedder hash
+```
+
+Expected shape (absolute numbers vary by machine; modes and columns do not):
+
+```
+[ingest] 450 points in 0.42s (1071 pts/s)
+### ingest throughput: 1071 pts/s
+
+HNSW m=16 ef_construct=100 hnsw_ef=64 multitenant=True payload_m=16 docs=450 queries=30
+
+| mode | recall@5 | recall@10 | index_recall@10 | p50 ms | p95 ms | p99 ms | qps |
+| --- | --- | --- | --- | --- | --- | --- | --- |
+| dense | 0.xxx | 0.xxx | 0.xxx | ... | ... | ... | ... |
+| sparse | 0.xxx | 0.xxx | n/a | ... | ... | ... | ... |
+| hybrid | 0.xxx | 0.xxx | n/a | ... | ... | ... | ... |
+```
+
+```bash
+# Full embedder (downloads models on first run):
+tenantq benchmark --embedder fastembed
 
 # One-off tenant-scoped search against an ingested collection:
-tenantq search "vector similarity ranking" --tenant acme --mode hybrid
+tenantq search "neural networks" --tenant acme --mode hybrid --limit 5 --embedder hash
 ```
 
 By default everything runs against an in-process Qdrant (`:memory:`), which
@@ -82,6 +110,7 @@ supports sparse vectors and Query API fusion — so the benchmark produces genui
 numbers with zero infrastructure. Point at a real server by setting `QDRANT_URL`
 (see Docker below).
 
+
 ## Configuration
 
 Everything is env-overridable (`Settings.from_env()`) or a CLI flag: