diff --git a/mbo/hash/measurements/hash_benchmark_report.py b/mbo/hash/measurements/hash_benchmark_report.py
index 9c80185..961b044 100644
--- a/mbo/hash/measurements/hash_benchmark_report.py
+++ b/mbo/hash/measurements/hash_benchmark_report.py
@@ -55,9 +55,10 @@
 import tarfile
 import tempfile
 
-# README tables show this (dense) set; the stored FULL dataset holds ~3x more
-# (a slow exponential) for the curve. Mirror kReadmeSizes in hash_benchmark.cc.
-_README_SIZES = [1, 3, 7, 8, 11, 15, 16, 19, 22, 27, 32, 47, 48, 63, 64, 256, 1024, 4096]
+# README tables read their sizes from the data itself (see _throughput_table),
+# so there is no size list here to drift. The curated README set = kReadmeSizes
+# in hash_benchmark.cc (the FAST mode); the tables are rendered from a fast-mode
+# dataset, the ns-vs-length chart from the dense FULL dataset.
 
 # Ordering uses the benchmark's algorithm keys (data keys); _LABEL_128 only
 # renames "mumbo" to "jumbo" for display in the 128-bit table.
@@ -394,12 +395,13 @@ def line(cells):
     return "\n".join([line(headers), "| " + " | ".join(sep) + " |"] + [line(r) for r in rows])
 
 
-def _throughput_table(data, preferred, relabel, sizes):
-    # Length-per-row, algorithm-per-column: with ~18 README sizes this reads far
-    # better than 18 columns, and each row's bold marks the fastest algorithm at
-    # that length.
+def _throughput_table(data, preferred, relabel):
+    # Length-per-row, algorithm-per-column; each row's bold marks the fastest
+    # algorithm at that length. The sizes are read from the data's own buckets
+    # (the single source of truth), so there is no second size list to drift from
+    # the C++ kReadmeSizes - feed this the README (fast-mode) dataset.
     algos = _order(list(data), preferred)
-    sizes = [s for s in sizes if any(str(s) in data[a] for a in algos)]
+    sizes = sorted({int(s) for a in data.values() for s in a})
     headers = ["Length"] + [relabel.get(a, a) for a in algos]
     rows = []
     for size in sizes:
@@ -447,11 +449,11 @@ def render_tables(results):
         "",
         f"### 64-bit one-shot throughput (ns/op, {agg}; lower is better)",
         "",
-        _throughput_table(results["throughput64"], _ORDER_64, {}, _README_SIZES),
+        _throughput_table(results["throughput64"], _ORDER_64, {}),
         "",
         f"### 128-bit one-shot throughput (ns/op, {agg}; native-128 algorithms only)",
         "",
-        _throughput_table(results["throughput128"], _ORDER_128, _LABEL_128, _README_SIZES),
+        _throughput_table(results["throughput128"], _ORDER_128, _LABEL_128),
         "",
         f"### Mixed-length latency (ns/hash, {agg}; lower is better)",
         "",
diff --git a/mbo/hash/measurements/run_measurements.py b/mbo/hash/measurements/run_measurements.py
index a7bd80e..664a16c 100755
--- a/mbo/hash/measurements/run_measurements.py
+++ b/mbo/hash/measurements/run_measurements.py
@@ -98,17 +98,26 @@ def main(argv):
     extras = []  # extra files packed alongside the canonical in the per-machine bundle
 
     if not args.skip_perf:
-        print(">>> [perf] full performance sweep (runs solo for clean numbers)", file=sys.stderr)
+        cfg = ["--config", args.config] if args.config else []
+        print(">>> [perf] full sweep (curve) + fast sweep (README tables), solo for clean numbers", file=sys.stderr)
+        # FULL sweep: the dense ns-vs-length curve + the authoritative raw (compare.py).
+        subprocess.run(
+            [*report, "run", "--mode", "full", "--reps", str(args.reps),
+             "--raw", os.path.join(data, "raw.json.gz"), "--out", os.path.join(data, "results.json")] + cfg,
+            check=True,
+        )
+        canonical = newest(os.path.join(data, "*_results.json"))
+        extras.append(newest(os.path.join(data, "*_raw.json.gz")))
+        # FAST sweep: the README tables. Its buckets ARE kReadmeSizes, and the tool
+        # reads the sizes from THIS data (no hardcoded list to drift from the C++).
         with open(os.path.join(data, "README_tables.md"), "w") as tables:
             subprocess.run(
-                [*report, "run", "--mode", "full", "--reps", str(args.reps),
-                 "--raw", os.path.join(data, "raw.json.gz"), "--out", os.path.join(data, "results.json"), "--tables"]
-                + (["--config", args.config] if args.config else []),
+                [*report, "run", "--mode", "fast", "--reps", str(args.reps),
+                 "--out", os.path.join(data, "results_readme.json"), "--tables"] + cfg,
                 stdout=tables, check=True,
             )
-        canonical = newest(os.path.join(data, "*_results.json"))
-        extras.append(newest(os.path.join(data, "*_raw.json.gz")))
-        print(">>> [perf] rendering ns-vs-length charts (64-bit + 128-bit, log-log)", file=sys.stderr)
+        extras.append(newest(os.path.join(data, "*_results_readme.json")))
+        print(">>> [perf] rendering ns-vs-length charts from the full sweep (64/128, log-log)", file=sys.stderr)
         subprocess.run([*report, "plot", "--results", canonical, "--out", os.path.join(data, "hash_throughput.svg")], check=True)
         for width in ("64", "128"):
             shutil.copy(