diff --git a/README.md b/README.md
index 2459e98..757e564 100644
--- a/README.md
+++ b/README.md
@@ -248,17 +248,18 @@ dfbench setup \
 ```bash
 dfbench run-commands --id percmd \
   --engines dragonfly,redis,valkey \
-  --command SET \
+  --command SET --dragonfly_num_shards -1 \
   --memtier-threads 32 --memtier-clients 5 --memtier-pipeline 20
 ```
 
 Repeat with the next command whenever you're ready - the setup stays
-provisioned between runs:
+provisioned between runs (`--dragonfly_num_shards` is required whenever
+dragonfly is among the engines; `-1` auto-derives the tuned value):
 
 ```bash
-dfbench run-commands --id percmd --command GET
-dfbench run-commands --id percmd --command LPUSH
-dfbench run-commands --id percmd --command LRANGE
+dfbench run-commands --id percmd --command GET --dragonfly_num_shards -1
+dfbench run-commands --id percmd --command LPUSH --dragonfly_num_shards -1
+dfbench run-commands --id percmd --command LRANGE --dragonfly_num_shards -1
 ```
 
 For each engine the scenario:
@@ -329,15 +330,16 @@ dfbench setup --id maxqps --tune-network --ubuntu-version 24.04 \
   --engines dragonfly
 ```
 
-Then benchmark. `--dragonfly_num_shards` (auto),
-`--dragonfly_conn_use_incoming_cpu`, and `--redis_pin_network` /
-`--valkey_pin_network` default to the tuned configuration, so a plain
-`run-commands` already reproduces it:
+Then benchmark. `--dragonfly_num_shards` must be stated explicitly when
+benchmarking dragonfly (pass `-1` for the auto-derived tuned value);
+`--dragonfly_conn_use_incoming_cpu` and `--redis_pin_network` /
+`--valkey_pin_network` default to the tuned configuration:
 
 ```bash
 dfbench run-commands --id maxqps --engines dragonfly \
   --command SET --memtier-threads 32 --memtier-clients 5 \
   --memtier-pipeline 35 --memtier-data-size 128 \
+  --dragonfly_num_shards -1 \
   --key-maximum 100000000 --test-time 300 \
   --trials 1 --warmup-time 0 --distinct-client-seed
 ```
@@ -367,12 +369,12 @@ Notes and caveats:
 ```bash
 # Baseline: pin an older release.
 dfbench run-commands --id maxqps --engines dragonfly --command SET \
-  --dragonfly_version v1.31.0 \
+  --dragonfly_version v1.31.0 --dragonfly_num_shards -1 \
   --memtier-pipeline 35 --memtier-data-size 128 --test-time 300 --trials 1
 
 # Candidate: latest release (or a nightly/custom tarball URL).
 dfbench run-commands --id maxqps --engines dragonfly --command SET \
-  --dragonfly_version latest \
+  --dragonfly_version latest --dragonfly_num_shards -1 \
   --memtier-pipeline 35 --memtier-data-size 128 --test-time 300 --trials 1
 ```
 
@@ -385,7 +387,7 @@ dfbench run-commands --id maxqps --engines dragonfly --command SET \
 ```bash
 # Local binary dropped into the repo (path is relative to your CWD).
 dfbench run-commands --id maxqps --engines dragonfly --command SET \
-  --dragonfly_version ./bin/dragonfly \
+  --dragonfly_version ./bin/dragonfly --dragonfly_num_shards -1 \
   --memtier-pipeline 35 --memtier-data-size 128 --test-time 300 --trials 1
 ```
 
@@ -596,7 +598,7 @@ Run it once per command you want to profile.
 | `--trials` | `3` | Measured trials per engine. Aggregated as median with [min..max]. |
 | `--distinct-client-seed` | `true` | Each memtier client gets its own RNG seed so the 160 clients don't replay the same key sequence. Disable with `=false`. |
 | `--redis_io_threads`, `--valkey_io_threads` | `0` (one per vCPU) | Override `--io-threads` for Redis / Valkey respectively; their docs suggest ~8 can outperform higher counts on many-core boxes. Dragonfly is unaffected (proactor threads always match vCPUs). Recorded in `results.json` as `io_threads` (per engine). Deprecated shared alias: `--io-threads` (sets both). |
-| `--dragonfly_num_shards` | `-1` (auto) | Dragonfly `--num_shards`. `-1` auto-derives `VCPUS - 2*NUM_RX_QUEUES` on the server so shards stay off the network-I/O CPUs that `setup --tune-network` reserves; `0` uses Dragonfly's default (one shard per proactor); `>0` sets it explicitly. Ignored by Redis/Valkey. Recorded as `num_shards`. Deprecated alias: `--num-shards`. |
+| `--dragonfly_num_shards` | none (required) | Dragonfly `--num_shards`. Required when benchmarking dragonfly so the config always states the shard count explicitly (no silent default). `-1` auto-derives `VCPUS - 2*NUM_RX_QUEUES` on the server so shards stay off the network-I/O CPUs that `setup --tune-network` reserves; `0` uses Dragonfly's default (one shard per proactor); `>0` sets it explicitly. Ignored by Redis/Valkey. Recorded as `num_shards`. Deprecated alias: `--num-shards`. |
 | `--dragonfly_conn_use_incoming_cpu` | `true` | Run Dragonfly with `--conn_use_incoming_cpu` so each connection's I/O runs on the CPU that received its NIC interrupt. Pair with `dfbench setup --tune-network`. Disable with `=false`. Ignored by Redis/Valkey. Recorded as `conn_use_incoming_cpu`. Deprecated alias: `--conn-use-incoming-cpu`. |
 | `--redis_pin_network`, `--valkey_pin_network` | `true` | Best-effort `taskset` pinning for Redis / Valkey respectively onto CPU 0 plus the network-I/O CPUs (falls back to no pinning if NIC-queue detection fails). Disable with `=false`. Ignored by Dragonfly. Recorded as `pin_network` (per engine). Deprecated shared alias: `--pin-network` (sets both). |
 | `--dragonfly_version` | `""` (keep installed) | Reinstall Dragonfly at this build before benchmarking so version deltas are measured on identical tuning. Accepts a release tag (`v1.31.0`), `latest`, a full `https://` URL to a `dragonfly-<arch>.tar.gz` tarball, or a path to a local file (a raw binary or `.tar.gz` in the repo) that is uploaded to the server. Empty keeps whatever `setup` installed. Ignored by Redis/Valkey. The installed version is recorded per-engine in `results.json`. Deprecated alias: `--dfly-version`. |
@@ -609,7 +611,7 @@ measurements):
 ```bash
 dfbench run-commands --id percmd \
   --engines dragonfly \
-  --command GET \
+  --command GET --dragonfly_num_shards -1 \
   --key-maximum 1000000 --test-time 15 --trials 1
 ```
 
@@ -672,8 +674,8 @@ server; they are unrelated.
 | `--dfly-bench-pipeline` | `20` | `dfly_bench --pipeline`: max pending requests per connection. |
 | `--dfly-bench-data-size` | `256` | `dfly_bench -d`: value size in bytes (`__data__` placeholders). |
 | `--key-maximum` | `100000000` | Key range (`--key_maximum`). Structured preloads use a built-in 1M-key range. |
-| `--key-dist` | `U` | `dfly_bench --key_dist` for the measured pass: `U` (uniform, the memtier random `R` equivalent), `N`, `Z`, `S`. |
-| `--qps` | `0` | `dfly_bench --qps` per connection; `0` = closed-loop max load (coordinated omission). |
+| `--dfly-bench-key-dist` | `U` | `dfly_bench --key_dist` for the measured pass: `U` (uniform, the memtier random `R` equivalent), `N`, `Z`, `S`. |
+| `--dfly-bench-qps` | `0` | `dfly_bench --qps` per connection; `0` = closed-loop max load (coordinated omission). |
 | `--test-time` | `300` | Seconds of measured load per trial. |
 | `--warmup-time` | `10` | Seconds of discarded warmup load per engine. `0` disables. |
 | `--trials` | `3` | Measured trials per engine (median with [min..max]). |
@@ -689,7 +691,7 @@ avg columns for both tools.
 ```bash
 dfbench run-commands-dfly --id percmd \
   --engines dragonfly \
-  --command GET \
+  --command GET --dragonfly_num_shards -1 \
   --key-maximum 1000000 --test-time 15 --trials 1
 ```
 
diff --git a/benchmarks/GET/dfly_bench/GET_reproduce.md b/benchmarks/GET/dfly_bench/GET_reproduce.md
index 2341288..69cea79 100644
--- a/benchmarks/GET/dfly_bench/GET_reproduce.md
+++ b/benchmarks/GET/dfly_bench/GET_reproduce.md
@@ -52,14 +52,14 @@ in the chosen AZ.
 dfbench run-commands-dfly --id maxqps2 --engines dragonfly,redis,valkey --command GET \
   --dragonfly_version s3://df-use1-pub/v2/dragonfly-aarch64.tar.gz \
   --dragonfly_num_shards 16 \
-  --key-maximum 100000000 --key-dist U \
+  --key-maximum 100000000 --test-time 60 --warmup-time 10 --trials 1 \
   --dfly-bench-threads 32 --dfly-bench-conns 5 --dfly-bench-pipeline 30 --dfly-bench-data-size 128 \
-  --qps 0 --test-time 60 --warmup-time 10 --trials 1
+  --dfly-bench-key-dist U --dfly-bench-qps 0
 ```
 
-`--qps 0` is closed-loop max load (coordinated omission). To pin throughput near a
-target rate instead, set `--qps` per connection = target_total / 160 (e.g.
-`--qps 68750` ≈ 11M ops/s total). Redis/Valkey are the preinstalled builds started
+`--dfly-bench-qps 0` is closed-loop max load (coordinated omission). To pin throughput near a
+target rate instead, set `--dfly-bench-qps` per connection = target_total / 160 (e.g.
+`--dfly-bench-qps 68750` ≈ 11M ops/s total). Redis/Valkey are the preinstalled builds started
 with `io_threads=32` (auto); only Dragonfly is reinstalled from `--dragonfly_version`.
 
 ## Engine comparison (dragonfly vs redis vs valkey)
@@ -89,16 +89,16 @@ Server: m7g.8xlarge (arm64) | Client: c6gn.8xlarge (arm64) | value size 128B, 16
 
 | Pipeline | Rate       | Throughput | p50      | p99      | p99.9      | Avg latency |
 | -------- | ---------- | ---------- | -------- | -------- | ---------- | ----------- |
-| 120      | `--qps 0`  | 13.47M ops/s | 1.064 ms | 7.435 ms | 33.307 ms | 1.344 ms |
-| 80       | `--qps 0`  | 12.98M ops/s | 0.780 ms | 3.971 ms | 15.049 ms | 0.928 ms |
-| 30       | `--qps 68750` (~11M) | 11.82M ops/s | 1.064 ms | 3.669 ms | 7.666 ms | 1.173 ms |
-| **30**   | **`--qps 0`** | **10.31M ops/s** | **0.431 ms** | **0.765 ms** | **0.996 ms** | **0.442 ms** |
+| 120      | `--dfly-bench-qps 0`  | 13.47M ops/s | 1.064 ms | 7.435 ms | 33.307 ms | 1.344 ms |
+| 80       | `--dfly-bench-qps 0`  | 12.98M ops/s | 0.780 ms | 3.971 ms | 15.049 ms | 0.928 ms |
+| 30       | `--dfly-bench-qps 68750` (~11M) | 11.82M ops/s | 1.064 ms | 3.669 ms | 7.666 ms | 1.173 ms |
+| **30**   | **`--dfly-bench-qps 0`** | **10.31M ops/s** | **0.431 ms** | **0.765 ms** | **0.996 ms** | **0.442 ms** |
 
 **Reading the sweep:** with 160 connections at max load, tail latency scales with
 in-flight depth (`conns x pipeline`). Dropping the pipeline from 120 to 30 cuts
 p99.9 by ~33x (33.3 ms -> 1.0 ms) while only trimming throughput ~24% (13.5M ->
 10.3M), because Dragonfly is CPU-bound and 160x30 still keeps every shard busy.
-Throttling pipeline-30 up toward 11M (`--qps 68750`) recovers throughput but
+Throttling pipeline-30 up toward 11M (`--dfly-bench-qps 68750`) recovers throughput but
 re-inflates the tail to ~7.7 ms — still under 10 ms, if you want the extra ops/s.
 
 ## Comparison with memtier
@@ -108,7 +108,7 @@ measured Dragonfly at **11.63M ops/s, p99.9 = 7.97 ms** with pipeline 120. dfly_
 at pipeline 120 pushes harder (13.5M ops/s) but with a much longer tail, so the two
 tools are not directly comparable at identical pipeline settings — dfly_bench's
 closed-loop driver keeps deeper queues. For a memtier-like tail, use a lower
-pipeline (30) or throttle with `--qps`.
+pipeline (30) or throttle with `--dfly-bench-qps`.
 
 Both tools' results share the same schema (this run is tagged `tool="dfly_bench"`),
 so a memtier GET file and a dfly_bench GET file diff with `dfbench compare`.
diff --git a/benchmarks/GET/memtier/GET_reproduce.md b/benchmarks/GET/memtier/GET_reproduce.md
index 6fd1b76..9c5e1de 100644
--- a/benchmarks/GET/memtier/GET_reproduce.md
+++ b/benchmarks/GET/memtier/GET_reproduce.md
@@ -62,9 +62,9 @@ load generator instead of memtier. dfly_bench is installed on the client at
 dfbench run-commands-dfly --id maxqps2 --engines dragonfly --command GET \
   --dragonfly_version s3://df-use1-pub/v2/dragonfly-aarch64.tar.gz \
   --dragonfly_num_shards 16 \
+  --key-maximum 100000000 --test-time 60 --warmup-time 10 --trials 1 \
   --dfly-bench-threads 32 --dfly-bench-conns 5 --dfly-bench-pipeline 120 --dfly-bench-data-size 128 \
-  --key-maximum 100000000 --key-dist U --qps 0 \
-  --test-time 60 --warmup-time 10 --trials 1
+  --dfly-bench-key-dist U --dfly-bench-qps 0
 ```
 
 Notes:
@@ -74,8 +74,8 @@ Notes:
   thread covering the whole `--key_maximum` range) rather than reused from a
   memtier preload.
 - **Flag mapping:** `--dfly-bench-threads` -> `--proactor_threads`, `--dfly-bench-conns` ->
-  `-c` (32x5 = 160 connections), `--key-dist U` approximates memtier's random
-  key pattern `R`, `--qps 0` = closed-loop max load.
+  `-c` (32x5 = 160 connections), `--dfly-bench-key-dist U` approximates memtier's random
+  key pattern `R`, `--dfly-bench-qps 0` = closed-loop max load.
 - Results use the same schema (tagged `tool="dfly_bench"`), so
   `dfbench compare <memtier-get.json> <dflybench-get.json>` diffs the two tools.
 - dfly_bench's `--json_out_file` writes a memtier-compatible latency report, so
diff --git a/benchmarks/SET/dfly_bench/SET_reproduce.md b/benchmarks/SET/dfly_bench/SET_reproduce.md
index c99bdd9..b58cc6b 100644
--- a/benchmarks/SET/dfly_bench/SET_reproduce.md
+++ b/benchmarks/SET/dfly_bench/SET_reproduce.md
@@ -47,15 +47,15 @@ instance types must have capacity in the chosen AZ.
 dfbench run-commands-dfly --id maxqps2 --engines dragonfly,redis,valkey --command SET \
   --dragonfly_version s3://df-use1-pub/v2/dragonfly-aarch64.tar.gz \
   --dragonfly_num_shards 16 --dragonfly_conn_use_incoming_cpu=true \
-  --key-maximum 100000000 --key-dist U \
+  --redis_io_threads 8 --valkey_io_threads 8 \
+  --key-maximum 100000000 --test-time 300 --warmup-time 10 --trials 1 \
   --dfly-bench-threads 32 --dfly-bench-conns 5 --dfly-bench-pipeline 30 --dfly-bench-data-size 128 \
-  --qps 0 --test-time 300 --warmup-time 10 --trials 1 \
-  --redis_io_threads 8 --valkey_io_threads 8
+  --dfly-bench-key-dist U --dfly-bench-qps 0
 ```
 
 ## Results
 
-Latest 3-engine run (benchmark `20260721-115805`, pipeline 30, `--qps 0`, 160
+Latest 3-engine run (benchmark `20260721-115805`, pipeline 30, `--dfly-bench-qps 0`, 160
 connections, 1 x 300s trial, value size 128B).
 
 | Engine | Throughput (median) | p50 | p99 | p99.9 | Avg latency |
diff --git a/benchmarks/SET/memtier/SET_reproduce.md b/benchmarks/SET/memtier/SET_reproduce.md
index cc78f31..0340d7b 100644
--- a/benchmarks/SET/memtier/SET_reproduce.md
+++ b/benchmarks/SET/memtier/SET_reproduce.md
@@ -35,13 +35,14 @@ load generator instead of memtier. dfly_bench is installed on the client at
 
 ```
 dfbench run-commands-dfly --id test --engines dragonfly --command SET \
+  --dragonfly_version s3://df-use1-pub/v2/dragonfly-aarch64.tar.gz \
   --dragonfly_num_shards 16 \
+  --key-maximum 100000000 --test-time 300 --trials 1 \
   --dfly-bench-threads 32 --dfly-bench-conns 5 --dfly-bench-pipeline 30 --dfly-bench-data-size 128 \
-  --key-maximum 100000000 --qps 0 \
-  --test-time 300 --trials 1 --dragonfly_version s3://df-use1-pub/v2/dragonfly-aarch64.tar.gz
+  --dfly-bench-qps 0
 ```
 
 Flag mapping vs memtier: `--dfly-bench-threads` -> `--proactor_threads`, `--dfly-bench-conns`
--> `-c` (so 32x5 = 160 connections, same as memtier threadsxclients), `--qps 0`
+-> `-c` (so 32x5 = 160 connections, same as memtier threadsxclients), `--dfly-bench-qps 0`
 = closed-loop max load. Results use the same schema (tagged `tool="dfly_bench"`),
 so `dfbench compare <memtier-set.json> <dflybench-set.json>` diffs the two tools.
diff --git a/cmd/run_commands.go b/cmd/run_commands.go
index 2d4f3d7..a640d29 100644
--- a/cmd/run_commands.go
+++ b/cmd/run_commands.go
@@ -6,6 +6,7 @@ import (
 	"os"
 	"os/signal"
 	"path/filepath"
+	"slices"
 	"strconv"
 	"strings"
 	"syscall"
@@ -102,6 +103,14 @@ marked approximate in the output.`,
 			if keyMaximum <= 0 {
 				return fmt.Errorf("--key-maximum must be > 0, got %d", keyMaximum)
 			}
+			// num_shards must be stated explicitly (via the flag or its
+			// deprecated alias) so a config never hides the shard count
+			// behind a silent auto default.
+			if slices.Contains(engines, "dragonfly") &&
+				!cmd.Flags().Changed("dragonfly_num_shards") && !cmd.Flags().Changed("num-shards") {
+				return fmt.Errorf("--dragonfly_num_shards is required when benchmarking dragonfly; " +
+					"pass an explicit value (-1 auto-derives VCPUS - 2*NUM_RX_QUEUES on the server, 0 uses Dragonfly's default, >0 sets it explicitly)")
+			}
 			// Resolve engine-specific tuning: the engine-prefixed flag wins,
 			// else the deprecated shared alias, else the default.
 			tuning, err := resolveEngineTuning(cmd.Flags())
@@ -157,7 +166,7 @@ marked approximate in the output.`,
 
 	// Dragonfly-specific tuning (ignored by Redis/Valkey).
 	f.IntVar(&dflyNumShards, "dragonfly_num_shards", -1,
-		"Dragonfly --num_shards: -1 auto-derives VCPUS - 2*NUM_RX_QUEUES on the server (keeps shards off the network-I/O CPUs), 0 uses Dragonfly's default (one shard per proactor), >0 sets it explicitly.")
+		"Dragonfly --num_shards (required when benchmarking dragonfly): -1 auto-derives VCPUS - 2*NUM_RX_QUEUES on the server (keeps shards off the network-I/O CPUs), 0 uses Dragonfly's default (one shard per proactor), >0 sets it explicitly.")
 	f.BoolVar(&dflyConnIncoming, "dragonfly_conn_use_incoming_cpu", true,
 		"Run Dragonfly with --conn_use_incoming_cpu so each connection's I/O runs on the CPU that received its NIC interrupt. Pair with 'dfbench setup --tune-network'. Disable with =false.")
 	f.StringVar(&dflyVersion, "dragonfly_version", "",
diff --git a/cmd/run_commands_dfly.go b/cmd/run_commands_dfly.go
index 8264ff8..f55c40c 100644
--- a/cmd/run_commands_dfly.go
+++ b/cmd/run_commands_dfly.go
@@ -6,6 +6,7 @@ import (
 	"os"
 	"os/signal"
 	"path/filepath"
+	"slices"
 	"strings"
 	"syscall"
 	"time"
@@ -93,6 +94,10 @@ compare'.`,
 			if keyMaximum <= 0 {
 				return fmt.Errorf("--key-maximum must be > 0, got %d", keyMaximum)
 			}
+			if slices.Contains(engines, "dragonfly") && !cmd.Flags().Changed("dragonfly_num_shards") {
+				return fmt.Errorf("--dragonfly_num_shards is required when benchmarking dragonfly; " +
+					"pass an explicit value (-1 auto-derives VCPUS - 2*NUM_RX_QUEUES on the server, 0 uses Dragonfly's default, >0 sets it explicitly)")
+			}
 			tuning, err := resolveEngineTuning(cmd.Flags())
 			if err != nil {
 				return err
@@ -152,9 +157,16 @@ compare'.`,
 	_ = f.MarkDeprecated("dfly-data-size", "use --dfly-bench-data-size")
 	f.IntVar(&keyMaximum, "key-maximum", 100_000_000,
 		"Key range (--key_maximum). Size it to 30-90% of server memory. Commands with multi-item structured preloads use a built-in 1M-key range instead.")
-	f.StringVar(&keyDist, "key-dist", "U",
+	f.StringVar(&keyDist, "dfly-bench-key-dist", "U",
 		"dfly_bench --key_dist for the measured pass: U (uniform, the memtier random equivalent), N (normal), Z (zipfian), S (sequential)")
-	f.IntVar(&qps, "qps", 0, "dfly_bench --qps per connection; 0 means closed-loop max load (coordinated omission)")
+	f.IntVar(&qps, "dfly-bench-qps", 0, "dfly_bench --qps per connection; 0 means closed-loop max load (coordinated omission)")
+
+	// Deprecated pre-prefix aliases: keep old commands working. Hidden from
+	// help and print a deprecation notice when used.
+	f.StringVar(&keyDist, "key-dist", "U", "Deprecated alias for --dfly-bench-key-dist")
+	f.IntVar(&qps, "qps", 0, "Deprecated alias for --dfly-bench-qps")
+	_ = f.MarkDeprecated("key-dist", "use --dfly-bench-key-dist")
+	_ = f.MarkDeprecated("qps", "use --dfly-bench-qps")
 	f.IntVar(&testTime, "test-time", 300, "Seconds of measured load per trial")
 	f.IntVar(&warmupTime, "warmup-time", 10, "Seconds of discarded warmup load per (engine, command); 0 disables")
 	f.IntVar(&trials, "trials", 3, "Number of measured trials per (engine, command) (>=1)")
@@ -163,7 +175,7 @@ compare'.`,
 
 	// Dragonfly-specific server tuning (ignored by Redis/Valkey).
 	f.IntVar(&dflyNumShards, "dragonfly_num_shards", -1,
-		"Dragonfly --num_shards: -1 auto-derives VCPUS - 2*NUM_RX_QUEUES on the server, 0 uses Dragonfly's default, >0 sets it explicitly.")
+		"Dragonfly --num_shards (required when benchmarking dragonfly): -1 auto-derives VCPUS - 2*NUM_RX_QUEUES on the server, 0 uses Dragonfly's default, >0 sets it explicitly.")
 	f.BoolVar(&dflyConnIncoming, "dragonfly_conn_use_incoming_cpu", true,
 		"Run Dragonfly with --conn_use_incoming_cpu. Pair with 'dfbench setup --tune-network'. Disable with =false.")
 	f.StringVar(&dflyVersion, "dragonfly_version", "",