diff --git a/benchmarks/bench.sh b/benchmarks/bench.sh
index 419fd5be3ad2b..df33d2aec7e54 100755
--- a/benchmarks/bench.sh
+++ b/benchmarks/bench.sh
@@ -106,7 +106,7 @@ wide_schema:            Small-projection queries on a wide synthetic dataset (10
                           (runs both 'wide' and 'narrow' subgroups: narrow is an internal baseline; the wide-vs-narrow ratio is the signal)
 predicate_eval:         Conjunctive (AND) filter-evaluation micro-benchmarks; each subgroup is a different predicate pattern, to test how an
                           adaptive predicate-ordering system behaves across them (see https://github.com/apache/datafusion/issues/11262)
-                          (subgroups via BENCH_SUBGROUP: costsel, cost, selectivity, cardinality, width, scale, neutral, correlation, drift)
+                          (subgroups via BENCH_SUBGROUP: costsel, cost, selectivity, cardinality, width, scale, neutral, correlation, drift, nulls)
                           (toggle a system under test with its native DATAFUSION_* env var; size data with PRED_ROWS, string width with PRED_FILL)
 parquet_row_filter_skip: Per-RG fully-matched RowFilter skip on Parquet (apache/datafusion#23696); clustered string key + low-selectivity
                           range filter + pushdown, so most row groups are fully matched and the per-row RowFilter is skipped on them
@@ -876,7 +876,11 @@ run_push_down_topk() {
 # micro-benchmarks where each subgroup is a different predicate pattern, used to
 # test how an adaptive predicate-ordering system behaves across them (see
 # https://github.com/apache/datafusion/issues/11262). Data is generated inline
-# by the suite's load SQL, so there is no data step.
+# by the suite's load SQL, so there is no data step (drift q82 and q83 share 16
+# small Parquet files written into sql_benchmarks/predicate_eval/scratch/, which
+# is gitignored; q82 reads them with target_partitions=1, so the selectivity flip
+# lands halfway through one stream, and q83 with 16, so each stream gets one whole
+# file and a fixed profile).
 #
 # By default the suite measures DataFusion's built-in left-deep AND short-circuit
 # and sets no engine config of its own. To evaluate a system under test, export
@@ -885,7 +889,7 @@ run_push_down_topk() {
 #   DATAFUSION_EXECUTION_ADAPTIVE_FILTER_REORDERING=true ./bench.sh run predicate_eval
 # Suite-specific knobs (string-substituted into the load SQL, not engine config):
 #   BENCH_SUBGROUP   run one subgroup (costsel, cost, selectivity, cardinality,
-#                    width, scale, neutral, correlation, drift)
+#                    width, scale, neutral, correlation, drift, nulls)
 #   PRED_ROWS        synthetic row count (default 1_000_000; the scale subgroup
 #                    overrides this per query)
 #   PRED_FILL        filler chars per marker = string-column width knob
diff --git a/benchmarks/sql_benchmarks/README.md b/benchmarks/sql_benchmarks/README.md
index 1ce1fa488c6b8..b8bf8adedede2 100644
--- a/benchmarks/sql_benchmarks/README.md
+++ b/benchmarks/sql_benchmarks/README.md
@@ -43,7 +43,7 @@ in the community:
 | `tpcds`               | TPC‑DS queries                                                     |
 | `tpch`                | TPC‑H queries                                                      |
 | `wide_schema`         | Small-projection queries on a wide (1024-col, 256-file) synthetic dataset; runs `wide` + `narrow` subgroups for comparison |
-| `predicate_eval`      | Conjunctive (AND) filter-evaluation micro-benchmarks; each subgroup is a different predicate pattern, to test how an adaptive predicate-ordering system behaves across them ([#11262](https://github.com/apache/datafusion/issues/11262)). Subgroups (`--subgroup`): `costsel`, `cost`, `selectivity`, `cardinality`, `width`, `scale`, `neutral`, `correlation`, `drift`. Configure the system under test through its DataFusion settings. |
+| `predicate_eval`      | Conjunctive (AND) filter-evaluation micro-benchmarks; each subgroup is a different predicate pattern, to test how an adaptive predicate-ordering system behaves across them ([#11262](https://github.com/apache/datafusion/issues/11262)). Subgroups (`--subgroup`): `costsel`, `cost`, `selectivity`, `cardinality`, `width`, `scale`, `neutral`, `correlation`, `drift`, `nulls`. The suite sets no engine config of its own, so by default it measures DataFusion's built-in left-deep `AND` short-circuit; point it at a system under test by exporting that system's own DataFusion setting (the harness builds its `SessionConfig` with `SessionConfig::from_env`). Every query is a `count(*)`, and the counts are checked in under `predicate_eval/results/`, so `--result-mode validate` also checks that a reordering under test still returns the same rows; the checked-in counts were persisted at the suite defaults (`PRED_ROWS=1000000`, `PRED_FILL=30`), so validation assumes those (the `scale` and `width` subgroups pin their own values per query and validate at any setting). |
 | `parquet_row_filter_skip` | Micro-benchmark for the per-row-group fully-matched RowFilter skip on Parquet scans ([#23696](https://github.com/apache/datafusion/issues/23696)). Subgroups (`--subgroup`): `skip` (clustered key, most row groups fully matched by statistics so the per-row filter is skipped), `control` (scrambled key, no row group is ever fully matched). Size the data with `PRED_ROWS` and the row-group size with `RG_SIZE`. |
 
 # Running Benchmarks
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/cardinality/q30.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/cardinality/q30.benchmark
index 760ea2ca902a4..d404a193d4dd3 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/cardinality/q30.benchmark
+++ b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/cardinality/q30.benchmark
@@ -1,7 +1,6 @@
 subgroup cardinality
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=cardinality
-QPAD=30
+QUERY=cardinality/q30
 DATASET=ints
 NAME=cardinality_q30_k2
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/cardinality/q31.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/cardinality/q31.benchmark
index 74f22715d1eb6..54a25337bdd65 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/cardinality/q31.benchmark
+++ b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/cardinality/q31.benchmark
@@ -1,7 +1,6 @@
 subgroup cardinality
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=cardinality
-QPAD=31
+QUERY=cardinality/q31
 DATASET=ints
 NAME=cardinality_q31_k4
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/cardinality/q32.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/cardinality/q32.benchmark
index b6b69c3852361..aff0173e3d890 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/cardinality/q32.benchmark
+++ b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/cardinality/q32.benchmark
@@ -1,7 +1,6 @@
 subgroup cardinality
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=cardinality
-QPAD=32
+QUERY=cardinality/q32
 DATASET=ints
 NAME=cardinality_q32_k8
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/cardinality/q33.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/cardinality/q33.benchmark
index 1260e68137860..bb1464d200d54 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/cardinality/q33.benchmark
+++ b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/cardinality/q33.benchmark
@@ -1,7 +1,6 @@
 subgroup cardinality
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=cardinality
-QPAD=33
+QUERY=cardinality/q33
 DATASET=ints
 NAME=cardinality_q33_k16
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/cardinality/q34.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/cardinality/q34.benchmark
new file mode 100644
index 0000000000000..c1851e34ad130
--- /dev/null
+++ b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/cardinality/q34.benchmark
@@ -0,0 +1,6 @@
+subgroup cardinality
+
+template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
+QUERY=cardinality/q34
+DATASET=ints_wide
+NAME=cardinality_q34_k8_wide64
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/correlation/q70.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/correlation/q70.benchmark
index ef20f7dc495b8..2084ddab97cae 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/correlation/q70.benchmark
+++ b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/correlation/q70.benchmark
@@ -1,7 +1,6 @@
 subgroup correlation
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=correlation
-QPAD=70
+QUERY=correlation/q70
 DATASET=corr
 NAME=correlation_q70_independent
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/correlation/q71.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/correlation/q71.benchmark
index 8875f6c44e359..cd5e63aa95fb6 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/correlation/q71.benchmark
+++ b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/correlation/q71.benchmark
@@ -1,7 +1,6 @@
 subgroup correlation
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=correlation
-QPAD=71
+QUERY=correlation/q71
 DATASET=corr
 NAME=correlation_q71_positive
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/correlation/q72.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/correlation/q72.benchmark
index 8109f1439aedb..e1dfc3fdafdaf 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/correlation/q72.benchmark
+++ b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/correlation/q72.benchmark
@@ -1,7 +1,6 @@
 subgroup correlation
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=correlation
-QPAD=72
+QUERY=correlation/q72
 DATASET=corr
 NAME=correlation_q72_anti
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/correlation/q73.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/correlation/q73.benchmark
index cc3f7bcf54901..c9a89b1da801f 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/correlation/q73.benchmark
+++ b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/correlation/q73.benchmark
@@ -1,7 +1,6 @@
 subgroup correlation
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=correlation
-QPAD=73
+QUERY=correlation/q73
 DATASET=corrproxy
 NAME=correlation_q73_redundant_proxy
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/cost/q10.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/cost/q10.benchmark
index 9b864b859457d..57c7034573714 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/cost/q10.benchmark
+++ b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/cost/q10.benchmark
@@ -1,7 +1,6 @@
 subgroup cost
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=cost
-QPAD=10
+QUERY=cost/q10
 DATASET=mixed
 NAME=cost_q10_expensive_first
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/cost/q11.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/cost/q11.benchmark
index 296ea443b3fec..30398925cb740 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/cost/q11.benchmark
+++ b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/cost/q11.benchmark
@@ -1,7 +1,6 @@
 subgroup cost
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=cost
-QPAD=11
+QUERY=cost/q11
 DATASET=mixed
 NAME=cost_q11_cheap_first
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/costsel/q01.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/costsel/q01.benchmark
index abedd1d580831..123839a2defb6 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/costsel/q01.benchmark
+++ b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/costsel/q01.benchmark
@@ -1,7 +1,6 @@
 subgroup costsel
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=costsel
-QPAD=01
+QUERY=costsel/q01
 DATASET=markers
 NAME=costsel_q01_regexp_selective_last
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/costsel/q02.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/costsel/q02.benchmark
index f50aab66427ec..304e67189d1e9 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/costsel/q02.benchmark
+++ b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/costsel/q02.benchmark
@@ -1,7 +1,6 @@
 subgroup costsel
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=costsel
-QPAD=02
+QUERY=costsel/q02
 DATASET=markers
 NAME=costsel_q02_regexp_selective_first
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/costsel/q03.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/costsel/q03.benchmark
index 10c4ce184eb34..d039e6e28dcc0 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/costsel/q03.benchmark
+++ b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/costsel/q03.benchmark
@@ -1,7 +1,6 @@
 subgroup costsel
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=costsel
-QPAD=03
+QUERY=costsel/q03
 DATASET=mixed
 NAME=costsel_q03_cheap_unselective_then_expensive_selective
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/costsel/q04.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/costsel/q04.benchmark
new file mode 100644
index 0000000000000..811e5d642aca1
--- /dev/null
+++ b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/costsel/q04.benchmark
@@ -0,0 +1,6 @@
+subgroup costsel
+
+template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
+QUERY=costsel/q04
+DATASET=mixed
+NAME=costsel_q04_expensive_selective_then_cheap_unselective
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/drift/q80.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/drift/q80.benchmark
index 970adc53f8017..fcbd32b0bfd1f 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/drift/q80.benchmark
+++ b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/drift/q80.benchmark
@@ -1,7 +1,6 @@
 subgroup drift
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=drift
-QPAD=80
+QUERY=drift/q80
 DATASET=drift
 NAME=drift_q80_a_then_b
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/drift/q81.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/drift/q81.benchmark
deleted file mode 100644
index 93cde75ffef87..0000000000000
--- a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/drift/q81.benchmark
+++ /dev/null
@@ -1,7 +0,0 @@
-subgroup drift
-
-template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=drift
-QPAD=81
-DATASET=drift
-NAME=drift_q81_b_then_a
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/drift/q82.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/drift/q82.benchmark
new file mode 100644
index 0000000000000..8ec68bd5c1d91
--- /dev/null
+++ b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/drift/q82.benchmark
@@ -0,0 +1,10 @@
+subgroup drift
+
+template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
+QUERY=drift/q82
+DATASET=drift_split
+NAME=drift_q82_late_flip
+
+# One stream over the whole directory: the flip lands mid-scan.
+init
+set datafusion.execution.target_partitions = 1;
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/drift/q83.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/drift/q83.benchmark
new file mode 100644
index 0000000000000..be25d56612e1a
--- /dev/null
+++ b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/drift/q83.benchmark
@@ -0,0 +1,10 @@
+subgroup drift
+
+template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
+QUERY=drift/q83
+DATASET=drift_split
+NAME=drift_q83_per_partition_skew
+
+# One file per stream: each stream sees a single profile.
+init
+set datafusion.execution.target_partitions = 16;
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/neutral/q60.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/neutral/q60.benchmark
index 039fee622b48b..5d39993e98d09 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/neutral/q60.benchmark
+++ b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/neutral/q60.benchmark
@@ -1,7 +1,6 @@
 subgroup neutral
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=neutral
-QPAD=60
+QUERY=neutral/q60
 DATASET=ints
 NAME=neutral_q60_cheap_uniform
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/neutral/q61.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/neutral/q61.benchmark
index edaf89b471c5f..ec66cdfc39786 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/neutral/q61.benchmark
+++ b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/neutral/q61.benchmark
@@ -1,7 +1,6 @@
 subgroup neutral
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=neutral
-QPAD=61
+QUERY=neutral/q61
 DATASET=markers
 NAME=neutral_q61_expensive_uniform
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/nulls/q90.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/nulls/q90.benchmark
new file mode 100644
index 0000000000000..4e82661a56ac5
--- /dev/null
+++ b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/nulls/q90.benchmark
@@ -0,0 +1,6 @@
+subgroup nulls
+
+template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
+QUERY=nulls/q90
+DATASET=ints
+NAME=nulls_q90_nullable_selective_first
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/nulls/q91.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/nulls/q91.benchmark
new file mode 100644
index 0000000000000..377b139bf0b46
--- /dev/null
+++ b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/nulls/q91.benchmark
@@ -0,0 +1,6 @@
+subgroup nulls
+
+template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
+QUERY=nulls/q91
+DATASET=ints
+NAME=nulls_q91_nullable_selective_last
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/scale/q50.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/scale/q50.benchmark
index 0bef31e14f402..1c8f2c148f34a 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/scale/q50.benchmark
+++ b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/scale/q50.benchmark
@@ -1,8 +1,7 @@
 subgroup scale
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=scale
-QPAD=50
+QUERY=costsel/q03
 DATASET=mixed
 PRED_ROWS=5000
 NAME=scale_q50_5k
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/scale/q51.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/scale/q51.benchmark
index 8f1315fb113b1..52bd2a0fd0a02 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/scale/q51.benchmark
+++ b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/scale/q51.benchmark
@@ -1,8 +1,7 @@
 subgroup scale
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=scale
-QPAD=51
+QUERY=costsel/q03
 DATASET=mixed
 PRED_ROWS=100000
 NAME=scale_q51_100k
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/scale/q52.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/scale/q52.benchmark
index 7ddbfc19b443d..9b37a4862738a 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/scale/q52.benchmark
+++ b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/scale/q52.benchmark
@@ -1,8 +1,7 @@
 subgroup scale
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=scale
-QPAD=52
+QUERY=costsel/q03
 DATASET=mixed
 PRED_ROWS=5000000
 NAME=scale_q52_5m
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/scale/q53.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/scale/q53.benchmark
index 6cea5c44a108b..9839289113b56 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/scale/q53.benchmark
+++ b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/scale/q53.benchmark
@@ -1,8 +1,7 @@
 subgroup scale
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=scale
-QPAD=53
+QUERY=costsel/q03
 DATASET=mixed
 PRED_ROWS=50000000
 NAME=scale_q53_50m
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/selectivity/q20.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/selectivity/q20.benchmark
index 077a62650d2f0..f587dc50d941d 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/selectivity/q20.benchmark
+++ b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/selectivity/q20.benchmark
@@ -1,7 +1,6 @@
 subgroup selectivity
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=selectivity
-QPAD=20
+QUERY=selectivity/q20
 DATASET=ints
 NAME=selectivity_q20_unselective_first
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/selectivity/q21.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/selectivity/q21.benchmark
index 24fc6ef4cd62f..655b067211aed 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/selectivity/q21.benchmark
+++ b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/selectivity/q21.benchmark
@@ -1,7 +1,6 @@
 subgroup selectivity
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=selectivity
-QPAD=21
+QUERY=selectivity/q21
 DATASET=ints
 NAME=selectivity_q21_selective_first
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/width/q40.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/width/q40.benchmark
index df66cf16a37ec..186fafc40bd93 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/width/q40.benchmark
+++ b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/width/q40.benchmark
@@ -1,8 +1,7 @@
 subgroup width
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=width
-QPAD=40
+QUERY=costsel/q01
 DATASET=markers
 PRED_FILL=2
 NAME=width_q40_narrow
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/width/q41.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/width/q41.benchmark
index c260dc9985a0c..16967c4d87d66 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/width/q41.benchmark
+++ b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/width/q41.benchmark
@@ -1,8 +1,7 @@
 subgroup width
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=width
-QPAD=41
+QUERY=costsel/q01
 DATASET=markers
 PRED_FILL=30
 NAME=width_q41_wide
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/width/q42.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/width/q42.benchmark
index 988ff59c70fe5..44ebe3137de13 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/width/q42.benchmark
+++ b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/width/q42.benchmark
@@ -1,8 +1,7 @@
 subgroup width
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=width
-QPAD=42
+QUERY=costsel/q01
 DATASET=markers
 PRED_FILL=170
 NAME=width_q42_xwide
diff --git a/benchmarks/sql_benchmarks/predicate_eval/load/corr.sql b/benchmarks/sql_benchmarks/predicate_eval/load/corr.sql
index 2d7ceb73e608d..946a999ceba28 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/load/corr.sql
+++ b/benchmarks/sql_benchmarks/predicate_eval/load/corr.sql
@@ -1,15 +1,6 @@
--- Correlation dataset: a base column plus derived columns that control the
--- *conditional* selectivity of one predicate given another (its selectivity
--- among the rows that already passed the other).
---
---   x       uniform [0,100)
---   x_pos   = x         (perfectly positively correlated: `x<k AND x_pos<k`
---                        passes ~k%, not the ~k%^2 an independence assumption
---                        predicts)
---   x_anti  = 99 - x    (anti-correlated: `x<k AND x_anti<k` is empty for k<=50)
---   ind     independent control column, uniform [0,100)
---
--- PRED_ROWS sizes the table.
+-- Correlation dataset: `x` uniform on [0,100); `x_pos` = x (perfectly positively
+-- correlated); `x_anti` = 99 - x (anti-correlated); `ind` uniform on [0,100) and
+-- independent of x. PRED_ROWS sizes the table.
 CREATE TABLE t AS
 SELECT
   (value * 7)  % 100        AS x,
diff --git a/benchmarks/sql_benchmarks/predicate_eval/load/corrproxy.sql b/benchmarks/sql_benchmarks/predicate_eval/load/corrproxy.sql
index f06e68d38cee2..e17d0d4925863 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/load/corrproxy.sql
+++ b/benchmarks/sql_benchmarks/predicate_eval/load/corrproxy.sql
@@ -1,30 +1,20 @@
--- Correlated-proxy dataset: a cheap integer predicate that is a perfect proxy
--- for three string predicates, plus one independent string predicate.
---
+-- Correlated-proxy dataset:
 --   c0          = 1 for ~30% of rows (cheap proxy)
---   s1, s2, s3  each contain a marker exactly where c0 = 1     (correlated)
---   s4          contains a marker for an independent ~30%      (independent)
---
--- The four string columns are deliberately *identical in shape*: same width,
--- the same single marker at the same offset, each matched by an equally cheap
--- regex with the same ~30% marginal selectivity. Marginally the four regex
--- predicates are therefore indistinguishable -- same cost, same selectivity, in
--- every position -- so neither a marginal cost/selectivity estimator nor
--- runtime timing can prefer one over another. Only their *conditional*
--- behaviour behind the proxy differs: after `c0 = 1`, the s1/s2/s3 regexes keep
--- every survivor (each re-tests the proxy's condition) while the s4 regex still
--- discards ~70%. Only joint statistics can see that; an independence assumption
--- prices all four regexes identically in every position.
+--   s1, s2, s3  each carry a marker exactly where c0 = 1  (correlated)
+--   s4          carries a marker for an independent ~30%  (independent)
+-- PRED_FILL sets the filler width each side of the marker, PRED_ROWS the row count.
 --
--- PRED_FILL sets the filler width on each side of the marker (a non-matching
--- `regexp_like` must scan the whole value), and PRED_ROWS sizes the table.
+-- The four string columns are identical in shape -- same width, same marker
+-- offset, same regex cost, same ~30% marginal selectivity -- so marginally they
+-- are indistinguishable in every position. Only their joint distribution with the
+-- proxy differs: after `c0 = 1` the s1/s2/s3 regexes keep every survivor while s4
+-- still discards ~70%. Ranking them therefore takes joint statistics; an
+-- independence assumption prices all four regexes identically everywhere.
 CREATE TABLE t AS
 WITH base AS (
   SELECT
-    -- The cheap proxy and the independent control share one definition each, so
-    -- the perfect-proxy / independence invariants can't drift apart silently.
-    (value * 7)  % 100 < 30 AS proxy,   -- ~30%, drives c0 and s1/s2/s3
-    (value * 13) % 100 < 30 AS indep    -- ~30%, independent of proxy, drives s4
+    (value * 7)  % 100 < 30 AS proxy,
+    (value * 13) % 100 < 30 AS indep
   FROM generate_series(1, ${PRED_ROWS:-1000000})
 )
 SELECT
diff --git a/benchmarks/sql_benchmarks/predicate_eval/load/drift.sql b/benchmarks/sql_benchmarks/predicate_eval/load/drift.sql
index d5635d91dd88a..0345abc3f291b 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/load/drift.sql
+++ b/benchmarks/sql_benchmarks/predicate_eval/load/drift.sql
@@ -1,16 +1,11 @@
--- Drift dataset: two predicates whose *relative* selectivity flips partway
--- through the scan, so whole-table selectivity differs from per-batch
--- selectivity. Rows are emitted in `seq` order, so batches observe the drift in
--- order.
---
---   a_sel = 0  is selective (~0.1%) in the first 10% of rows, unselective
---              (~50%) afterwards.
---   b_sel = 0  is the mirror: unselective early, selective late.
---
--- PRED_ROWS sizes the table.
+-- Drift dataset: rows are emitted in `seq` order and which predicate is more
+-- selective flips partway through, so whole-table selectivity differs from
+-- per-batch selectivity. `a_sel = 0` matches ~0.1% over the first 2% of rows and
+-- ~50% after; `b_sel = 0` is the mirror. At the default PRED_ROWS, which sizes the
+-- table, that flip lands inside the pooled 8-batch warm-up (~65k rows).
 CREATE TABLE t AS
 SELECT
   value AS seq,
-  CASE WHEN value < ${PRED_ROWS:-1000000} / 10 THEN value % 1000 ELSE value % 2    END AS a_sel,
-  CASE WHEN value < ${PRED_ROWS:-1000000} / 10 THEN value % 2    ELSE value % 1000 END AS b_sel
+  CASE WHEN value < ${PRED_ROWS:-1000000} / 50 THEN value % 1000 ELSE value % 2    END AS a_sel,
+  CASE WHEN value < ${PRED_ROWS:-1000000} / 50 THEN value % 2    ELSE value % 1000 END AS b_sel
 FROM generate_series(1, ${PRED_ROWS:-1000000});
diff --git a/benchmarks/sql_benchmarks/predicate_eval/load/drift_split.sql b/benchmarks/sql_benchmarks/predicate_eval/load/drift_split.sql
new file mode 100644
index 0000000000000..b36a0d122f444
--- /dev/null
+++ b/benchmarks/sql_benchmarks/predicate_eval/load/drift_split.sql
@@ -0,0 +1,96 @@
+-- Split drift dataset, shared by q82 and q83: 16 Parquet files of PRED_ROWS / 16
+-- rows each, at the q80 rates -- f00..f07 carry profile A (`a_sel = 0` selective
+-- at ~0.1%, `b_sel = 0` unselective at ~50%) and f08..f15 the mirror.
+--
+-- One `COPY` per file, because a directory-target `COPY` names its output with a
+-- random write id and Parquet has no overwrite, so a second load would add files;
+-- the names are zero-padded so they sort numerically. `FileGroup::split_files`
+-- sorts a group's files by path, so `target_partitions = 1` (q82) reads f00..f15
+-- in order and `= 16` (q83) gives each stream one whole file; each query sets it.
+-- `repartition_file_scans` stays off, or both profiles land in one stream again.
+
+COPY (SELECT value AS seq, value % 1000 AS a_sel, value % 2    AS b_sel
+      FROM generate_series(1, ${PRED_ROWS:-1000000} / 16) ORDER BY value)
+TO 'sql_benchmarks/predicate_eval/scratch/drift_split/f00.parquet'
+STORED AS PARQUET;
+
+COPY (SELECT value AS seq, value % 1000 AS a_sel, value % 2    AS b_sel
+      FROM generate_series(1, ${PRED_ROWS:-1000000} / 16) ORDER BY value)
+TO 'sql_benchmarks/predicate_eval/scratch/drift_split/f01.parquet'
+STORED AS PARQUET;
+
+COPY (SELECT value AS seq, value % 1000 AS a_sel, value % 2    AS b_sel
+      FROM generate_series(1, ${PRED_ROWS:-1000000} / 16) ORDER BY value)
+TO 'sql_benchmarks/predicate_eval/scratch/drift_split/f02.parquet'
+STORED AS PARQUET;
+
+COPY (SELECT value AS seq, value % 1000 AS a_sel, value % 2    AS b_sel
+      FROM generate_series(1, ${PRED_ROWS:-1000000} / 16) ORDER BY value)
+TO 'sql_benchmarks/predicate_eval/scratch/drift_split/f03.parquet'
+STORED AS PARQUET;
+
+COPY (SELECT value AS seq, value % 1000 AS a_sel, value % 2    AS b_sel
+      FROM generate_series(1, ${PRED_ROWS:-1000000} / 16) ORDER BY value)
+TO 'sql_benchmarks/predicate_eval/scratch/drift_split/f04.parquet'
+STORED AS PARQUET;
+
+COPY (SELECT value AS seq, value % 1000 AS a_sel, value % 2    AS b_sel
+      FROM generate_series(1, ${PRED_ROWS:-1000000} / 16) ORDER BY value)
+TO 'sql_benchmarks/predicate_eval/scratch/drift_split/f05.parquet'
+STORED AS PARQUET;
+
+COPY (SELECT value AS seq, value % 1000 AS a_sel, value % 2    AS b_sel
+      FROM generate_series(1, ${PRED_ROWS:-1000000} / 16) ORDER BY value)
+TO 'sql_benchmarks/predicate_eval/scratch/drift_split/f06.parquet'
+STORED AS PARQUET;
+
+COPY (SELECT value AS seq, value % 1000 AS a_sel, value % 2    AS b_sel
+      FROM generate_series(1, ${PRED_ROWS:-1000000} / 16) ORDER BY value)
+TO 'sql_benchmarks/predicate_eval/scratch/drift_split/f07.parquet'
+STORED AS PARQUET;
+
+COPY (SELECT value AS seq, value % 2    AS a_sel, value % 1000 AS b_sel
+      FROM generate_series(1, ${PRED_ROWS:-1000000} / 16) ORDER BY value)
+TO 'sql_benchmarks/predicate_eval/scratch/drift_split/f08.parquet'
+STORED AS PARQUET;
+
+COPY (SELECT value AS seq, value % 2    AS a_sel, value % 1000 AS b_sel
+      FROM generate_series(1, ${PRED_ROWS:-1000000} / 16) ORDER BY value)
+TO 'sql_benchmarks/predicate_eval/scratch/drift_split/f09.parquet'
+STORED AS PARQUET;
+
+COPY (SELECT value AS seq, value % 2    AS a_sel, value % 1000 AS b_sel
+      FROM generate_series(1, ${PRED_ROWS:-1000000} / 16) ORDER BY value)
+TO 'sql_benchmarks/predicate_eval/scratch/drift_split/f10.parquet'
+STORED AS PARQUET;
+
+COPY (SELECT value AS seq, value % 2    AS a_sel, value % 1000 AS b_sel
+      FROM generate_series(1, ${PRED_ROWS:-1000000} / 16) ORDER BY value)
+TO 'sql_benchmarks/predicate_eval/scratch/drift_split/f11.parquet'
+STORED AS PARQUET;
+
+COPY (SELECT value AS seq, value % 2    AS a_sel, value % 1000 AS b_sel
+      FROM generate_series(1, ${PRED_ROWS:-1000000} / 16) ORDER BY value)
+TO 'sql_benchmarks/predicate_eval/scratch/drift_split/f12.parquet'
+STORED AS PARQUET;
+
+COPY (SELECT value AS seq, value % 2    AS a_sel, value % 1000 AS b_sel
+      FROM generate_series(1, ${PRED_ROWS:-1000000} / 16) ORDER BY value)
+TO 'sql_benchmarks/predicate_eval/scratch/drift_split/f13.parquet'
+STORED AS PARQUET;
+
+COPY (SELECT value AS seq, value % 2    AS a_sel, value % 1000 AS b_sel
+      FROM generate_series(1, ${PRED_ROWS:-1000000} / 16) ORDER BY value)
+TO 'sql_benchmarks/predicate_eval/scratch/drift_split/f14.parquet'
+STORED AS PARQUET;
+
+COPY (SELECT value AS seq, value % 2    AS a_sel, value % 1000 AS b_sel
+      FROM generate_series(1, ${PRED_ROWS:-1000000} / 16) ORDER BY value)
+TO 'sql_benchmarks/predicate_eval/scratch/drift_split/f15.parquet'
+STORED AS PARQUET;
+
+set datafusion.optimizer.repartition_file_scans = false;
+
+CREATE EXTERNAL TABLE t
+STORED AS PARQUET
+LOCATION 'sql_benchmarks/predicate_eval/scratch/drift_split/';
diff --git a/benchmarks/sql_benchmarks/predicate_eval/load/ints.sql b/benchmarks/sql_benchmarks/predicate_eval/load/ints.sql
index 1d3f030915b14..f36b283c2a68a 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/load/ints.sql
+++ b/benchmarks/sql_benchmarks/predicate_eval/load/ints.sql
@@ -1,8 +1,8 @@
--- Sixteen independent integer columns, each uniform on [0,100). The predicate
--- `cN < k` therefore has selectivity ~k%. All columns are equally cheap to
--- evaluate, so only selectivity (not cost) distinguishes orderings here. The
--- multipliers are all coprime to 100, which keeps the residues uniform and the
--- columns mutually decorrelated. PRED_ROWS sizes the table.
+-- Seventeen integer columns: c0..c15 uniform on [0,100), so `cN < k` has
+-- selectivity ~k%, all equally cheap and mutually decorrelated (the multipliers
+-- are coprime to 100); plus `c_sel`, NULL on 10% of rows and uniform on [0,100)
+-- elsewhere, so `c_sel < 5` is true on 4% of rows and NULL on 10% -- a selective
+-- conjunct that is not NULL-free. PRED_ROWS sizes the table.
 CREATE TABLE t AS
 SELECT
   (value * 1)  % 100 AS c0,
@@ -20,5 +20,6 @@ SELECT
   (value * 31) % 100 AS c12,
   (value * 33) % 100 AS c13,
   (value * 37) % 100 AS c14,
-  (value * 39) % 100 AS c15
+  (value * 39) % 100 AS c15,
+  CASE WHEN value % 10 = 0 THEN NULL ELSE (value * 11) % 100 END AS c_sel
 FROM generate_series(1, ${PRED_ROWS:-1000000});
diff --git a/benchmarks/sql_benchmarks/predicate_eval/load/ints_wide.sql b/benchmarks/sql_benchmarks/predicate_eval/load/ints_wide.sql
new file mode 100644
index 0000000000000..cddce6f8e5a5a
--- /dev/null
+++ b/benchmarks/sql_benchmarks/predicate_eval/load/ints_wide.sql
@@ -0,0 +1,71 @@
+-- `ints.sql` widened to 64 integer columns, each uniform on [0,100), so `cN < k`
+-- again has selectivity ~k%. c0..c15 reuse the same multipliers, so a predicate
+-- over them has exactly the same selectivities as on `ints` and only the width of
+-- the batches flowing through the filter changes. PRED_ROWS sizes the table.
+CREATE TABLE t AS
+SELECT
+  (value * 1)   % 100 AS c0,
+  (value * 3)   % 100 AS c1,
+  (value * 7)   % 100 AS c2,
+  (value * 9)   % 100 AS c3,
+  (value * 11)  % 100 AS c4,
+  (value * 13)  % 100 AS c5,
+  (value * 17)  % 100 AS c6,
+  (value * 19)  % 100 AS c7,
+  (value * 21)  % 100 AS c8,
+  (value * 23)  % 100 AS c9,
+  (value * 27)  % 100 AS c10,
+  (value * 29)  % 100 AS c11,
+  (value * 31)  % 100 AS c12,
+  (value * 33)  % 100 AS c13,
+  (value * 37)  % 100 AS c14,
+  (value * 39)  % 100 AS c15,
+  (value * 41)  % 100 AS c16,
+  (value * 43)  % 100 AS c17,
+  (value * 47)  % 100 AS c18,
+  (value * 49)  % 100 AS c19,
+  (value * 51)  % 100 AS c20,
+  (value * 53)  % 100 AS c21,
+  (value * 57)  % 100 AS c22,
+  (value * 59)  % 100 AS c23,
+  (value * 61)  % 100 AS c24,
+  (value * 63)  % 100 AS c25,
+  (value * 67)  % 100 AS c26,
+  (value * 69)  % 100 AS c27,
+  (value * 71)  % 100 AS c28,
+  (value * 73)  % 100 AS c29,
+  (value * 77)  % 100 AS c30,
+  (value * 79)  % 100 AS c31,
+  (value * 81)  % 100 AS c32,
+  (value * 83)  % 100 AS c33,
+  (value * 87)  % 100 AS c34,
+  (value * 89)  % 100 AS c35,
+  (value * 91)  % 100 AS c36,
+  (value * 93)  % 100 AS c37,
+  (value * 97)  % 100 AS c38,
+  (value * 99)  % 100 AS c39,
+  (value * 101) % 100 AS c40,
+  (value * 103) % 100 AS c41,
+  (value * 107) % 100 AS c42,
+  (value * 109) % 100 AS c43,
+  (value * 111) % 100 AS c44,
+  (value * 113) % 100 AS c45,
+  (value * 117) % 100 AS c46,
+  (value * 119) % 100 AS c47,
+  (value * 121) % 100 AS c48,
+  (value * 123) % 100 AS c49,
+  (value * 127) % 100 AS c50,
+  (value * 129) % 100 AS c51,
+  (value * 131) % 100 AS c52,
+  (value * 133) % 100 AS c53,
+  (value * 137) % 100 AS c54,
+  (value * 139) % 100 AS c55,
+  (value * 141) % 100 AS c56,
+  (value * 143) % 100 AS c57,
+  (value * 147) % 100 AS c58,
+  (value * 149) % 100 AS c59,
+  (value * 151) % 100 AS c60,
+  (value * 153) % 100 AS c61,
+  (value * 157) % 100 AS c62,
+  (value * 159) % 100 AS c63
+FROM generate_series(1, ${PRED_ROWS:-1000000});
diff --git a/benchmarks/sql_benchmarks/predicate_eval/load/markers.sql b/benchmarks/sql_benchmarks/predicate_eval/load/markers.sql
index 6b8fff9385ea0..d1c818fe0f293 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/load/markers.sql
+++ b/benchmarks/sql_benchmarks/predicate_eval/load/markers.sql
@@ -1,15 +1,8 @@
--- Wide-string dataset: five markers embedded in `PRED_FILL`-wide filler so that a
--- non-matching `regexp_like` must scan the whole value (every string predicate
--- is "expensive"). Selectivities are coprime so the predicates are independent:
---
---   'aaa'  present in ~90%  of rows  (value % 10  <> 0)
---   'bbb'  present in ~86%  of rows  (value % 7   <> 0)
---   'ccc'  present in ~80%  of rows  (value % 5   <> 0)
---   'ddd'  present in ~75%  of rows  (value % 4   <> 0)
---   'rare' present in ~0.1% of rows  (value % 1009 = 5)   <- the selective one
---
--- PRED_FILL sets the filler width per marker (the string-column width knob: ~6*PRED_FILL
--- chars per row), and PRED_ROWS sizes the table.
+-- Wide-string dataset: one column `s` holding five markers in PRED_FILL-wide
+-- filler, so a non-matching `regexp_like` must scan the whole value. 'aaa' ~90%,
+-- 'bbb' ~86%, 'ccc' ~80%, 'ddd' ~75%, 'rare' ~0.1%; the moduli are coprime, so the
+-- markers are independent. PRED_FILL is the string-width knob (~6*PRED_FILL chars
+-- per row) and PRED_ROWS sizes the table.
 CREATE TABLE t AS
 SELECT
   repeat('q', ${PRED_FILL:-30})
diff --git a/benchmarks/sql_benchmarks/predicate_eval/load/mixed.sql b/benchmarks/sql_benchmarks/predicate_eval/load/mixed.sql
index a51c1040daca6..bc931e8a40af0 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/load/mixed.sql
+++ b/benchmarks/sql_benchmarks/predicate_eval/load/mixed.sql
@@ -1,15 +1,7 @@
--- Mixed-cost dataset: cheap integer columns (`cN < k` ~ k% selectivity)
--- alongside one wide string column carrying three markers matched by expensive
--- `regexp_like`:
---
---   'rare' present in ~0.1% of rows  (value % 1009 = 5)
---   'ten'  present in ~10%  of rows  (value % 10   = 0)
---   'aaa'  present in ~90%  of rows  (value % 10  <> 0)
---
--- This lets a single table mix cheap integer compares with expensive regexp
--- scans at independently chosen selectivities (e.g. a cheap, unselective compare
--- next to an expensive, selective regexp). PRED_FILL is the string-width knob;
--- PRED_ROWS sizes the table.
+-- Mixed-cost dataset: cheap integer columns c0..c3 uniform on [0,100) (`cN < k`
+-- ~k%) alongside one wide string column `s` carrying three markers matched by an
+-- expensive `regexp_like`: 'rare' ~0.1%, 'ten' ~10%, 'aaa' ~90%. PRED_FILL is the
+-- string-width knob and PRED_ROWS sizes the table.
 CREATE TABLE t AS
 SELECT
   (value * 1)  % 100 AS c0,
diff --git a/benchmarks/sql_benchmarks/predicate_eval/predicate_eval.benchmark.template b/benchmarks/sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
index 0030a7e946eca..4a23594d3b53b 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
+++ b/benchmarks/sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
@@ -1,34 +1,21 @@
-# Shared template for every predicate_eval benchmark. Each qNN.benchmark sets
-# its `subgroup` directive and then includes this template with parameters:
-#   SUBGROUP  subgroup name, also the query sub-directory          (e.g. costsel)
-#   QPAD      zero-padded query id, also the query file stem       (e.g. 01)
-#   DATASET   load script stem under load/                         (e.g. markers)
-#   NAME      criterion display name                               (e.g. costsel_q01_regexp_selective_last)
-# Optional (consumed by the load scripts via ${...:-default}):
-#   PRED_ROWS      synthetic row count                                  (default 1_000_000)
-#   PRED_FILL      filler chars per marker = string-column width knob    (default 30)
-#
-# The run SQL lives in queries/${SUBGROUP}/q${QPAD}.sql so the WHERE clause is
-# readable on its own. The table is always named `t`, so the assert and cleanup
-# are uniform across datasets.
-#
-# The suite is implementation-agnostic and sets no engine config of its own: it
-# measures DataFusion's built-in left-deep `AND` short-circuit by default. To
-# evaluate a predicate-ordering system under test, set its native config via the
-# environment (the bench harness builds its SessionContext with
-# SessionConfig::from_env), e.g.
-#   DATAFUSION_EXECUTION_ADAPTIVE_FILTER_REORDERING=true
+# Shared template for every predicate_eval benchmark. Each qNN.benchmark sets its
+# `subgroup` directive and includes this template with parameters:
+#   QUERY     query file stem under queries/  (e.g. costsel/q01)
+#   DATASET   load script stem under load/    (e.g. markers)
+#   NAME      criterion display name          (e.g. costsel_q01_regexp_selective_last)
+# Optional, read by the load scripts as ${...:-default}:
+#   PRED_ROWS  synthetic row count                          (default 1000000)
+#   PRED_FILL  filler chars per marker, the string-width knob (default 30)
+# The knob sweeps reuse another subgroup's query file through QUERY, and the table
+# is always named `t`, so cleanup is uniform across datasets.
 
 load sql_benchmarks/predicate_eval/load/${DATASET}.sql
 
 name ${NAME}
 group predicate_eval
 
-assert I
-SELECT count(*) > 0 FROM t;
-----
-true
+run sql_benchmarks/predicate_eval/queries/${QUERY}.sql
 
-run sql_benchmarks/predicate_eval/queries/${SUBGROUP}/q${QPAD}.sql
+result sql_benchmarks/predicate_eval/results/${NAME}.csv
 
 cleanup sql_benchmarks/predicate_eval/init/cleanup.sql
diff --git a/benchmarks/sql_benchmarks/predicate_eval/predicate_eval.suite b/benchmarks/sql_benchmarks/predicate_eval/predicate_eval.suite
index af1a326cd8c51..4ee33994d37f4 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/predicate_eval.suite
+++ b/benchmarks/sql_benchmarks/predicate_eval/predicate_eval.suite
@@ -1,4 +1,4 @@
-description = "Conjunctive filter evaluation micro-benchmarks covering predicate cost, selectivity, cardinality, width, scale, correlation, and drift"
+description = "Conjunctive filter evaluation micro-benchmarks covering predicate cost, selectivity, cardinality, width, scale, correlation, drift, and nullable predicates"
 
 query_pattern = "q{QUERY_ID_PADDED}.benchmark"
 
diff --git a/benchmarks/sql_benchmarks/predicate_eval/queries/cardinality/q30.sql b/benchmarks/sql_benchmarks/predicate_eval/queries/cardinality/q30.sql
index 3be840e917383..dbd410f9575a0 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/queries/cardinality/q30.sql
+++ b/benchmarks/sql_benchmarks/predicate_eval/queries/cardinality/q30.sql
@@ -1,6 +1,4 @@
--- Hidden: cheap integer compares; `c1 < 5` matches ~5%, the `c0 < 90` family
--- ~90%. k = 2 here. q30..q33 sweep k = 2/4/8/16 with one ~5% predicate written
--- last among ~90% ones.
+-- k = 2: one ~90% compare then one ~5% compare (q30..q33 sweep k = 2/4/8/16).
 SELECT count(*) FROM t
 WHERE c0 < 90
   AND c1 < 5;
diff --git a/benchmarks/sql_benchmarks/predicate_eval/queries/cardinality/q31.sql b/benchmarks/sql_benchmarks/predicate_eval/queries/cardinality/q31.sql
index 4ba84f8124be9..d21dd6273eefa 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/queries/cardinality/q31.sql
+++ b/benchmarks/sql_benchmarks/predicate_eval/queries/cardinality/q31.sql
@@ -1,4 +1,4 @@
--- k = 4: three ~90% compares followed by one ~5% compare. See q30.
+-- k = 4: three ~90% compares then one ~5% compare.
 SELECT count(*) FROM t
 WHERE c0 < 90
   AND c1 < 90
diff --git a/benchmarks/sql_benchmarks/predicate_eval/queries/cardinality/q32.sql b/benchmarks/sql_benchmarks/predicate_eval/queries/cardinality/q32.sql
index d9e920cc62574..83037bdfc5798 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/queries/cardinality/q32.sql
+++ b/benchmarks/sql_benchmarks/predicate_eval/queries/cardinality/q32.sql
@@ -1,4 +1,4 @@
--- k = 8: seven ~90% compares followed by one ~5% compare. See q30.
+-- k = 8: seven ~90% compares then one ~5% compare.
 SELECT count(*) FROM t
 WHERE c0 < 90
   AND c1 < 90
diff --git a/benchmarks/sql_benchmarks/predicate_eval/queries/cardinality/q33.sql b/benchmarks/sql_benchmarks/predicate_eval/queries/cardinality/q33.sql
index 2408427ab7632..890ff4205d68c 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/queries/cardinality/q33.sql
+++ b/benchmarks/sql_benchmarks/predicate_eval/queries/cardinality/q33.sql
@@ -1,4 +1,4 @@
--- k = 16: fifteen ~90% compares followed by one ~5% compare. See q30.
+-- k = 16: fifteen ~90% compares then one ~5% compare.
 SELECT count(*) FROM t
 WHERE c0 < 90
   AND c1 < 90
diff --git a/benchmarks/sql_benchmarks/predicate_eval/queries/cardinality/q34.sql b/benchmarks/sql_benchmarks/predicate_eval/queries/cardinality/q34.sql
new file mode 100644
index 0000000000000..2c5b586e54af8
--- /dev/null
+++ b/benchmarks/sql_benchmarks/predicate_eval/queries/cardinality/q34.sql
@@ -0,0 +1,10 @@
+-- k = 8 as in q32, but over the 64-column `ints_wide` table.
+SELECT count(*) FROM t
+WHERE c0 < 90
+  AND c1 < 90
+  AND c2 < 90
+  AND c3 < 90
+  AND c4 < 90
+  AND c5 < 90
+  AND c6 < 90
+  AND c7 < 5;
diff --git a/benchmarks/sql_benchmarks/predicate_eval/queries/correlation/q70.sql b/benchmarks/sql_benchmarks/predicate_eval/queries/correlation/q70.sql
index 86e33534c705d..f9d7017e618ce 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/queries/correlation/q70.sql
+++ b/benchmarks/sql_benchmarks/predicate_eval/queries/correlation/q70.sql
@@ -1,6 +1,4 @@
--- Hidden: `x` and `ind` are independent, each ~20%, so the conjunction matches
--- ~4% and the second predicate is just as selective among the first's survivors
--- as on its own. Baseline for the correlation sweep. cf. q71, q72.
+-- Independent: `x` and `ind` are uncorrelated, each ~20%, so the pair matches ~4%.
 SELECT count(*) FROM t
 WHERE x < 20
   AND ind < 20;
diff --git a/benchmarks/sql_benchmarks/predicate_eval/queries/correlation/q71.sql b/benchmarks/sql_benchmarks/predicate_eval/queries/correlation/q71.sql
index eda61cc289e92..4b5925e80a809 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/queries/correlation/q71.sql
+++ b/benchmarks/sql_benchmarks/predicate_eval/queries/correlation/q71.sql
@@ -1,6 +1,4 @@
--- Hidden: `x_pos` is a copy of `x`, so `x < 20 AND x_pos < 20` still matches
--- ~20% (not the ~4% independence would imply) -- the second predicate removes
--- none of the first's survivors. cf. q70.
+-- Positively correlated: `x_pos` is a copy of `x`, so the pair still matches ~20%.
 SELECT count(*) FROM t
 WHERE x < 20
   AND x_pos < 20;
diff --git a/benchmarks/sql_benchmarks/predicate_eval/queries/correlation/q72.sql b/benchmarks/sql_benchmarks/predicate_eval/queries/correlation/q72.sql
index ff987524da6ed..63ec8d57de3f4 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/queries/correlation/q72.sql
+++ b/benchmarks/sql_benchmarks/predicate_eval/queries/correlation/q72.sql
@@ -1,6 +1,4 @@
--- Hidden: `x_anti` is `99 - x`, so `x < 50 AND x_anti < 50` is empty -- the
--- second predicate removes all of the first's survivors, though each matches
--- ~50% alone. cf. q70.
+-- Anti-correlated: `x_anti` is `99 - x`, so the pair is empty though each matches ~50%.
 SELECT count(*) FROM t
 WHERE x < 50
   AND x_anti < 50;
diff --git a/benchmarks/sql_benchmarks/predicate_eval/queries/correlation/q73.sql b/benchmarks/sql_benchmarks/predicate_eval/queries/correlation/q73.sql
index 5e1e822e92eca..6dc76a8cab23f 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/queries/correlation/q73.sql
+++ b/benchmarks/sql_benchmarks/predicate_eval/queries/correlation/q73.sql
@@ -1,11 +1,5 @@
--- Hidden: `c0 = 1` is a perfect proxy for the s1/s2/s3 regexes -- after the
--- cheap proxy, each of those keeps every survivor while the equally selective
--- (~30%) s4 regex still discards ~70%. The optimal order is [c0, s4, s1/s2/s3]
--- (one informative regex on 30% of rows, the three redundant ones on 9%), but
--- the four regexes are marginally identical -- same width, same marker offset,
--- same cost, same selectivity -- so ranking them takes their *joint*
--- distribution with the proxy. Written with the redundant regexes first,
--- grouped with their proxy, as an author naturally would.
+-- Redundant proxy: `c0 = 1` implies the s1/s2/s3 regexes, while the marginally
+-- identical s4 regex is independent of it. Written proxy-first, regexes grouped.
 SELECT count(*) FROM t
 WHERE c0 = 1
   AND regexp_like(s1, 'a.a')
diff --git a/benchmarks/sql_benchmarks/predicate_eval/queries/cost/q10.sql b/benchmarks/sql_benchmarks/predicate_eval/queries/cost/q10.sql
index b089ebc7a192a..cbe0e937efde7 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/queries/cost/q10.sql
+++ b/benchmarks/sql_benchmarks/predicate_eval/queries/cost/q10.sql
@@ -1,6 +1,4 @@
--- Hidden: both predicates match ~10%, but `regexp_like(s, 'ten')` scans the
--- string (expensive) while `c0 < 10` is a cheap compare. Equal selectivity,
--- unequal cost; expensive one written first. cf. q11 (opposite order).
+-- Equal selectivity (~10%), unequal cost: the expensive regexp written first. cf. q11.
 SELECT count(*) FROM t
 WHERE regexp_like(s, 'ten')
   AND c0 < 10;
diff --git a/benchmarks/sql_benchmarks/predicate_eval/queries/cost/q11.sql b/benchmarks/sql_benchmarks/predicate_eval/queries/cost/q11.sql
index 82d748c93b3b2..54167ff19494f 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/queries/cost/q11.sql
+++ b/benchmarks/sql_benchmarks/predicate_eval/queries/cost/q11.sql
@@ -1,5 +1,4 @@
--- Same two predicates as q10 (both ~10%; regexp expensive, compare cheap),
--- opposite written order. cf. q10.
+-- q10 with the cheap compare written first.
 SELECT count(*) FROM t
 WHERE c0 < 10
   AND regexp_like(s, 'ten');
diff --git a/benchmarks/sql_benchmarks/predicate_eval/queries/costsel/q01.sql b/benchmarks/sql_benchmarks/predicate_eval/queries/costsel/q01.sql
index bc029ed5d8297..1618c54814cfa 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/queries/costsel/q01.sql
+++ b/benchmarks/sql_benchmarks/predicate_eval/queries/costsel/q01.sql
@@ -1,7 +1,5 @@
--- Hidden in the data: the five markers have very different selectivities --
--- 'aaa' ~90%, 'bbb' ~86%, 'ccc' ~80%, 'ddd' ~75%, 'rare' ~0.1% -- while every
--- regexp_like costs about the same. 'rare' (most selective) is written last.
--- cf. q02 (most selective written first).
+-- Five equally expensive regexps of very different selectivity ('rare' ~0.1%, the
+-- rest 75-90%), the selective one written last. cf. q02.
 SELECT count(*) FROM t
 WHERE regexp_like(s, 'aaa')
   AND regexp_like(s, 'bbb')
diff --git a/benchmarks/sql_benchmarks/predicate_eval/queries/costsel/q02.sql b/benchmarks/sql_benchmarks/predicate_eval/queries/costsel/q02.sql
index 7f7fc61831ff0..dc57aa8d88da0 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/queries/costsel/q02.sql
+++ b/benchmarks/sql_benchmarks/predicate_eval/queries/costsel/q02.sql
@@ -1,5 +1,4 @@
--- Same predicates and hidden selectivities as q01 ('rare' ~0.1% is the
--- selective one, the rest 75-90%), but with 'rare' written first. cf. q01.
+-- q01 with the selective regexp written first.
 SELECT count(*) FROM t
 WHERE regexp_like(s, 'rare')
   AND regexp_like(s, 'aaa')
diff --git a/benchmarks/sql_benchmarks/predicate_eval/queries/costsel/q03.sql b/benchmarks/sql_benchmarks/predicate_eval/queries/costsel/q03.sql
index a583a498b211c..a9d110c65a058 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/queries/costsel/q03.sql
+++ b/benchmarks/sql_benchmarks/predicate_eval/queries/costsel/q03.sql
@@ -1,6 +1,4 @@
--- Hidden: `c0 < 90` matches ~90% (cheap integer compare); `regexp_like(s,
--- 'rare')` matches ~0.1% (scans the wide string). The cheaper predicate is the
--- less selective one.
+-- Cheap unselective compare (~90%) then expensive selective regexp (~0.1%). cf. q04.
 SELECT count(*) FROM t
 WHERE c0 < 90
   AND regexp_like(s, 'rare');
diff --git a/benchmarks/sql_benchmarks/predicate_eval/queries/costsel/q04.sql b/benchmarks/sql_benchmarks/predicate_eval/queries/costsel/q04.sql
new file mode 100644
index 0000000000000..2c5a253664264
--- /dev/null
+++ b/benchmarks/sql_benchmarks/predicate_eval/queries/costsel/q04.sql
@@ -0,0 +1,4 @@
+-- q03 the other way round, so the as-written order is already the best one.
+SELECT count(*) FROM t
+WHERE regexp_like(s, 'rare')
+  AND c0 < 90;
diff --git a/benchmarks/sql_benchmarks/predicate_eval/queries/drift/q80.sql b/benchmarks/sql_benchmarks/predicate_eval/queries/drift/q80.sql
index b8cb61e85a478..63e55ce473a7f 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/queries/drift/q80.sql
+++ b/benchmarks/sql_benchmarks/predicate_eval/queries/drift/q80.sql
@@ -1,7 +1,5 @@
--- The non-obvious property: selectivity changes across the scan. Rows arrive in
--- `seq` order; `a_sel = 0` matches ~0.1% in the first 10% of rows and ~50%
--- after, `b_sel = 0` is the mirror -- so which predicate is more selective flips
--- partway through. cf. q81 (opposite order).
+-- Selectivity drifts across the scan: `a_sel = 0` matches ~0.1% over the first 2%
+-- of rows and ~50% after, and `b_sel = 0` is the mirror.
 SELECT count(*) FROM t
 WHERE a_sel = 0
   AND b_sel = 0;
diff --git a/benchmarks/sql_benchmarks/predicate_eval/queries/drift/q81.sql b/benchmarks/sql_benchmarks/predicate_eval/queries/drift/q81.sql
deleted file mode 100644
index d65ef475cc0e0..0000000000000
--- a/benchmarks/sql_benchmarks/predicate_eval/queries/drift/q81.sql
+++ /dev/null
@@ -1,5 +0,0 @@
--- Same drifting predicates as q80 (a_sel/b_sel flip which is more selective
--- partway through the scan), opposite written order. cf. q80.
-SELECT count(*) FROM t
-WHERE b_sel = 0
-  AND a_sel = 0;
diff --git a/benchmarks/sql_benchmarks/predicate_eval/queries/drift/q82.sql b/benchmarks/sql_benchmarks/predicate_eval/queries/drift/q82.sql
new file mode 100644
index 0000000000000..57bb65d2162a3
--- /dev/null
+++ b/benchmarks/sql_benchmarks/predicate_eval/queries/drift/q82.sql
@@ -0,0 +1,5 @@
+-- Late flip: one stream reads f=00..f=15 in order, so the drift lands halfway
+-- through the scan and a warm-up-and-freeze decision is wrong for half the rows.
+SELECT count(*) FROM t
+WHERE a_sel = 0
+  AND b_sel = 0;
diff --git a/benchmarks/sql_benchmarks/predicate_eval/queries/drift/q83.sql b/benchmarks/sql_benchmarks/predicate_eval/queries/drift/q83.sql
new file mode 100644
index 0000000000000..7ecc0fb1c52f4
--- /dev/null
+++ b/benchmarks/sql_benchmarks/predicate_eval/queries/drift/q83.sql
@@ -0,0 +1,5 @@
+-- Per-partition skew: one whole file per stream, so each stream sees a single fixed
+-- profile and one pooled decision is backwards for half of them.
+SELECT count(*) FROM t
+WHERE a_sel = 0
+  AND b_sel = 0;
diff --git a/benchmarks/sql_benchmarks/predicate_eval/queries/neutral/q60.sql b/benchmarks/sql_benchmarks/predicate_eval/queries/neutral/q60.sql
index b217f56953272..76f5a154daedf 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/queries/neutral/q60.sql
+++ b/benchmarks/sql_benchmarks/predicate_eval/queries/neutral/q60.sql
@@ -1,5 +1,4 @@
--- Hidden: four integer compares of equal cost, each ~50% selective. Nothing is
--- selective and the costs are equal, so the predicates are interchangeable.
+-- Four equally cheap compares, each ~50%: nothing to reorder.
 SELECT count(*) FROM t
 WHERE c0 < 50
   AND c1 < 50
diff --git a/benchmarks/sql_benchmarks/predicate_eval/queries/neutral/q61.sql b/benchmarks/sql_benchmarks/predicate_eval/queries/neutral/q61.sql
index 7029a3d9f8f7d..43225f2e4809a 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/queries/neutral/q61.sql
+++ b/benchmarks/sql_benchmarks/predicate_eval/queries/neutral/q61.sql
@@ -1,6 +1,4 @@
--- Hidden: four regexp scans of about equal cost, all unselective ('aaa' ~90%,
--- 'bbb' ~86%, 'ccc' ~80%, 'ddd' ~75%). Like q60 the predicates are
--- interchangeable, but here each one is expensive.
+-- Four equally expensive regexps, all unselective (75-90%): nothing to reorder.
 SELECT count(*) FROM t
 WHERE regexp_like(s, 'aaa')
   AND regexp_like(s, 'bbb')
diff --git a/benchmarks/sql_benchmarks/predicate_eval/queries/nulls/q90.sql b/benchmarks/sql_benchmarks/predicate_eval/queries/nulls/q90.sql
new file mode 100644
index 0000000000000..ce4e40d9acfa9
--- /dev/null
+++ b/benchmarks/sql_benchmarks/predicate_eval/queries/nulls/q90.sql
@@ -0,0 +1,6 @@
+-- The selective conjunct (~4%) is nullable, so AND pre-selection is disabled and
+-- writing it first gates nothing. cf. q91.
+SELECT count(*) FROM t
+WHERE c_sel < 5
+  AND c0 < 90
+  AND c1 < 90;
diff --git a/benchmarks/sql_benchmarks/predicate_eval/queries/nulls/q91.sql b/benchmarks/sql_benchmarks/predicate_eval/queries/nulls/q91.sql
new file mode 100644
index 0000000000000..abca8bb1f90a2
--- /dev/null
+++ b/benchmarks/sql_benchmarks/predicate_eval/queries/nulls/q91.sql
@@ -0,0 +1,5 @@
+-- q90 with the nullable selective conjunct written last, which is the same work.
+SELECT count(*) FROM t
+WHERE c0 < 90
+  AND c1 < 90
+  AND c_sel < 5;
diff --git a/benchmarks/sql_benchmarks/predicate_eval/queries/scale/q50.sql b/benchmarks/sql_benchmarks/predicate_eval/queries/scale/q50.sql
deleted file mode 100644
index 03a0f1c0db285..0000000000000
--- a/benchmarks/sql_benchmarks/predicate_eval/queries/scale/q50.sql
+++ /dev/null
@@ -1,6 +0,0 @@
--- Same predicates as costsel/q03 (`c0 < 90` ~90% cheap, `regexp_like(s, 'rare')`
--- ~0.1% expensive). q50..q53 sweep table size; here PRED_ROWS=5_000, roughly a
--- single batch.
-SELECT count(*) FROM t
-WHERE c0 < 90
-  AND regexp_like(s, 'rare');
diff --git a/benchmarks/sql_benchmarks/predicate_eval/queries/scale/q51.sql b/benchmarks/sql_benchmarks/predicate_eval/queries/scale/q51.sql
deleted file mode 100644
index 28174a5df4f44..0000000000000
--- a/benchmarks/sql_benchmarks/predicate_eval/queries/scale/q51.sql
+++ /dev/null
@@ -1,4 +0,0 @@
--- q50 at PRED_ROWS=100_000 (~12 batches). See q50.
-SELECT count(*) FROM t
-WHERE c0 < 90
-  AND regexp_like(s, 'rare');
diff --git a/benchmarks/sql_benchmarks/predicate_eval/queries/scale/q52.sql b/benchmarks/sql_benchmarks/predicate_eval/queries/scale/q52.sql
deleted file mode 100644
index 74938c4634f78..0000000000000
--- a/benchmarks/sql_benchmarks/predicate_eval/queries/scale/q52.sql
+++ /dev/null
@@ -1,4 +0,0 @@
--- q50 at PRED_ROWS=5_000_000 (~610 batches). See q50.
-SELECT count(*) FROM t
-WHERE c0 < 90
-  AND regexp_like(s, 'rare');
diff --git a/benchmarks/sql_benchmarks/predicate_eval/queries/scale/q53.sql b/benchmarks/sql_benchmarks/predicate_eval/queries/scale/q53.sql
deleted file mode 100644
index 8edb4d4a057d2..0000000000000
--- a/benchmarks/sql_benchmarks/predicate_eval/queries/scale/q53.sql
+++ /dev/null
@@ -1,4 +0,0 @@
--- q50 at PRED_ROWS=50_000_000 (~6100 batches); builds a ~9 GB table. See q50.
-SELECT count(*) FROM t
-WHERE c0 < 90
-  AND regexp_like(s, 'rare');
diff --git a/benchmarks/sql_benchmarks/predicate_eval/queries/selectivity/q20.sql b/benchmarks/sql_benchmarks/predicate_eval/queries/selectivity/q20.sql
index 3638f757a720d..b834a4bd7b87a 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/queries/selectivity/q20.sql
+++ b/benchmarks/sql_benchmarks/predicate_eval/queries/selectivity/q20.sql
@@ -1,6 +1,5 @@
--- Hidden: two equally cheap integer compares of unequal selectivity -- `c4 < 95`
--- matches ~95%, `c0 < 5` matches ~5%. Less selective one written first.
--- cf. q21 (opposite order).
+-- Two equally cheap compares of unequal selectivity (~95%, ~5%), less selective
+-- first. cf. q21.
 SELECT count(*) FROM t
 WHERE c4 < 95
   AND c0 < 5;
diff --git a/benchmarks/sql_benchmarks/predicate_eval/queries/selectivity/q21.sql b/benchmarks/sql_benchmarks/predicate_eval/queries/selectivity/q21.sql
index 5181faf38784f..1f8531a51077b 100644
--- a/benchmarks/sql_benchmarks/predicate_eval/queries/selectivity/q21.sql
+++ b/benchmarks/sql_benchmarks/predicate_eval/queries/selectivity/q21.sql
@@ -1,5 +1,4 @@
--- Same two equally-cheap compares as q20 (`c4 < 95` ~95%, `c0 < 5` ~5%),
--- opposite written order. cf. q20.
+-- q20 with the selective compare written first.
 SELECT count(*) FROM t
 WHERE c0 < 5
   AND c4 < 95;
diff --git a/benchmarks/sql_benchmarks/predicate_eval/queries/width/q40.sql b/benchmarks/sql_benchmarks/predicate_eval/queries/width/q40.sql
deleted file mode 100644
index 1b3df3e937eb3..0000000000000
--- a/benchmarks/sql_benchmarks/predicate_eval/queries/width/q40.sql
+++ /dev/null
@@ -1,9 +0,0 @@
--- Same predicate set and hidden selectivities as costsel/q01 ('rare' ~0.1%, the
--- rest 75-90%); only the string-column width differs across q40/q41/q42. Narrow:
--- PRED_FILL=2, ~12 chars/row.
-SELECT count(*) FROM t
-WHERE regexp_like(s, 'aaa')
-  AND regexp_like(s, 'bbb')
-  AND regexp_like(s, 'ccc')
-  AND regexp_like(s, 'ddd')
-  AND regexp_like(s, 'rare');
diff --git a/benchmarks/sql_benchmarks/predicate_eval/queries/width/q41.sql b/benchmarks/sql_benchmarks/predicate_eval/queries/width/q41.sql
deleted file mode 100644
index a03b576d9c959..0000000000000
--- a/benchmarks/sql_benchmarks/predicate_eval/queries/width/q41.sql
+++ /dev/null
@@ -1,7 +0,0 @@
--- q40 with wide strings: PRED_FILL=30, ~186 chars/row. See q40.
-SELECT count(*) FROM t
-WHERE regexp_like(s, 'aaa')
-  AND regexp_like(s, 'bbb')
-  AND regexp_like(s, 'ccc')
-  AND regexp_like(s, 'ddd')
-  AND regexp_like(s, 'rare');
diff --git a/benchmarks/sql_benchmarks/predicate_eval/queries/width/q42.sql b/benchmarks/sql_benchmarks/predicate_eval/queries/width/q42.sql
deleted file mode 100644
index cb55d828e32ab..0000000000000
--- a/benchmarks/sql_benchmarks/predicate_eval/queries/width/q42.sql
+++ /dev/null
@@ -1,7 +0,0 @@
--- q40 with extra-wide strings: PRED_FILL=170, ~1KB/row. See q40.
-SELECT count(*) FROM t
-WHERE regexp_like(s, 'aaa')
-  AND regexp_like(s, 'bbb')
-  AND regexp_like(s, 'ccc')
-  AND regexp_like(s, 'ddd')
-  AND regexp_like(s, 'rare');
diff --git a/benchmarks/sql_benchmarks/predicate_eval/results/cardinality_q30_k2.csv b/benchmarks/sql_benchmarks/predicate_eval/results/cardinality_q30_k2.csv
new file mode 100644
index 0000000000000..23d7344bc8c8c
--- /dev/null
+++ b/benchmarks/sql_benchmarks/predicate_eval/results/cardinality_q30_k2.csv
@@ -0,0 +1,2 @@
+count(*)
+50000
diff --git a/benchmarks/sql_benchmarks/predicate_eval/results/cardinality_q31_k4.csv b/benchmarks/sql_benchmarks/predicate_eval/results/cardinality_q31_k4.csv
new file mode 100644
index 0000000000000..94f68af265629
--- /dev/null
+++ b/benchmarks/sql_benchmarks/predicate_eval/results/cardinality_q31_k4.csv
@@ -0,0 +1,2 @@
+count(*)
+40000
diff --git a/benchmarks/sql_benchmarks/predicate_eval/results/cardinality_q32_k8.csv b/benchmarks/sql_benchmarks/predicate_eval/results/cardinality_q32_k8.csv
new file mode 100644
index 0000000000000..23d7344bc8c8c
--- /dev/null
+++ b/benchmarks/sql_benchmarks/predicate_eval/results/cardinality_q32_k8.csv
@@ -0,0 +1,2 @@
+count(*)
+50000
diff --git a/benchmarks/sql_benchmarks/predicate_eval/results/cardinality_q33_k16.csv b/benchmarks/sql_benchmarks/predicate_eval/results/cardinality_q33_k16.csv
new file mode 100644
index 0000000000000..6291042419967
--- /dev/null
+++ b/benchmarks/sql_benchmarks/predicate_eval/results/cardinality_q33_k16.csv
@@ -0,0 +1,2 @@
+count(*)
+10000
diff --git a/benchmarks/sql_benchmarks/predicate_eval/results/cardinality_q34_k8_wide64.csv b/benchmarks/sql_benchmarks/predicate_eval/results/cardinality_q34_k8_wide64.csv
new file mode 100644
index 0000000000000..23d7344bc8c8c
--- /dev/null
+++ b/benchmarks/sql_benchmarks/predicate_eval/results/cardinality_q34_k8_wide64.csv
@@ -0,0 +1,2 @@
+count(*)
+50000
diff --git a/benchmarks/sql_benchmarks/predicate_eval/results/correlation_q70_independent.csv b/benchmarks/sql_benchmarks/predicate_eval/results/correlation_q70_independent.csv
new file mode 100644
index 0000000000000..23d7344bc8c8c
--- /dev/null
+++ b/benchmarks/sql_benchmarks/predicate_eval/results/correlation_q70_independent.csv
@@ -0,0 +1,2 @@
+count(*)
+50000
diff --git a/benchmarks/sql_benchmarks/predicate_eval/results/correlation_q71_positive.csv b/benchmarks/sql_benchmarks/predicate_eval/results/correlation_q71_positive.csv
new file mode 100644
index 0000000000000..e3795c1d30546
--- /dev/null
+++ b/benchmarks/sql_benchmarks/predicate_eval/results/correlation_q71_positive.csv
@@ -0,0 +1,2 @@
+count(*)
+200000
diff --git a/benchmarks/sql_benchmarks/predicate_eval/results/correlation_q72_anti.csv b/benchmarks/sql_benchmarks/predicate_eval/results/correlation_q72_anti.csv
new file mode 100644
index 0000000000000..4303b90458538
--- /dev/null
+++ b/benchmarks/sql_benchmarks/predicate_eval/results/correlation_q72_anti.csv
@@ -0,0 +1,2 @@
+count(*)
+0
diff --git a/benchmarks/sql_benchmarks/predicate_eval/results/correlation_q73_redundant_proxy.csv b/benchmarks/sql_benchmarks/predicate_eval/results/correlation_q73_redundant_proxy.csv
new file mode 100644
index 0000000000000..1cabb81336225
--- /dev/null
+++ b/benchmarks/sql_benchmarks/predicate_eval/results/correlation_q73_redundant_proxy.csv
@@ -0,0 +1,2 @@
+count(*)
+90000
diff --git a/benchmarks/sql_benchmarks/predicate_eval/results/cost_q10_expensive_first.csv b/benchmarks/sql_benchmarks/predicate_eval/results/cost_q10_expensive_first.csv
new file mode 100644
index 0000000000000..6291042419967
--- /dev/null
+++ b/benchmarks/sql_benchmarks/predicate_eval/results/cost_q10_expensive_first.csv
@@ -0,0 +1,2 @@
+count(*)
+10000
diff --git a/benchmarks/sql_benchmarks/predicate_eval/results/cost_q11_cheap_first.csv b/benchmarks/sql_benchmarks/predicate_eval/results/cost_q11_cheap_first.csv
new file mode 100644
index 0000000000000..6291042419967
--- /dev/null
+++ b/benchmarks/sql_benchmarks/predicate_eval/results/cost_q11_cheap_first.csv
@@ -0,0 +1,2 @@
+count(*)
+10000
diff --git a/benchmarks/sql_benchmarks/predicate_eval/results/costsel_q01_regexp_selective_last.csv b/benchmarks/sql_benchmarks/predicate_eval/results/costsel_q01_regexp_selective_last.csv
new file mode 100644
index 0000000000000..5eb0b761fa8b7
--- /dev/null
+++ b/benchmarks/sql_benchmarks/predicate_eval/results/costsel_q01_regexp_selective_last.csv
@@ -0,0 +1,2 @@
+count(*)
+508
diff --git a/benchmarks/sql_benchmarks/predicate_eval/results/costsel_q02_regexp_selective_first.csv b/benchmarks/sql_benchmarks/predicate_eval/results/costsel_q02_regexp_selective_first.csv
new file mode 100644
index 0000000000000..5eb0b761fa8b7
--- /dev/null
+++ b/benchmarks/sql_benchmarks/predicate_eval/results/costsel_q02_regexp_selective_first.csv
@@ -0,0 +1,2 @@
+count(*)
+508
diff --git a/benchmarks/sql_benchmarks/predicate_eval/results/costsel_q03_cheap_unselective_then_expensive_selective.csv b/benchmarks/sql_benchmarks/predicate_eval/results/costsel_q03_cheap_unselective_then_expensive_selective.csv
new file mode 100644
index 0000000000000..1c82d80b6f382
--- /dev/null
+++ b/benchmarks/sql_benchmarks/predicate_eval/results/costsel_q03_cheap_unselective_then_expensive_selective.csv
@@ -0,0 +1,2 @@
+count(*)
+893
diff --git a/benchmarks/sql_benchmarks/predicate_eval/results/costsel_q04_expensive_selective_then_cheap_unselective.csv b/benchmarks/sql_benchmarks/predicate_eval/results/costsel_q04_expensive_selective_then_cheap_unselective.csv
new file mode 100644
index 0000000000000..1c82d80b6f382
--- /dev/null
+++ b/benchmarks/sql_benchmarks/predicate_eval/results/costsel_q04_expensive_selective_then_cheap_unselective.csv
@@ -0,0 +1,2 @@
+count(*)
+893
diff --git a/benchmarks/sql_benchmarks/predicate_eval/results/drift_q80_a_then_b.csv b/benchmarks/sql_benchmarks/predicate_eval/results/drift_q80_a_then_b.csv
new file mode 100644
index 0000000000000..c543899cb162a
--- /dev/null
+++ b/benchmarks/sql_benchmarks/predicate_eval/results/drift_q80_a_then_b.csv
@@ -0,0 +1,2 @@
+count(*)
+1000
diff --git a/benchmarks/sql_benchmarks/predicate_eval/results/drift_q82_late_flip.csv b/benchmarks/sql_benchmarks/predicate_eval/results/drift_q82_late_flip.csv
new file mode 100644
index 0000000000000..571ea61b2741d
--- /dev/null
+++ b/benchmarks/sql_benchmarks/predicate_eval/results/drift_q82_late_flip.csv
@@ -0,0 +1,2 @@
+count(*)
+992
diff --git a/benchmarks/sql_benchmarks/predicate_eval/results/drift_q83_per_partition_skew.csv b/benchmarks/sql_benchmarks/predicate_eval/results/drift_q83_per_partition_skew.csv
new file mode 100644
index 0000000000000..571ea61b2741d
--- /dev/null
+++ b/benchmarks/sql_benchmarks/predicate_eval/results/drift_q83_per_partition_skew.csv
@@ -0,0 +1,2 @@
+count(*)
+992
diff --git a/benchmarks/sql_benchmarks/predicate_eval/results/neutral_q60_cheap_uniform.csv b/benchmarks/sql_benchmarks/predicate_eval/results/neutral_q60_cheap_uniform.csv
new file mode 100644
index 0000000000000..a1c39233b50b9
--- /dev/null
+++ b/benchmarks/sql_benchmarks/predicate_eval/results/neutral_q60_cheap_uniform.csv
@@ -0,0 +1,2 @@
+count(*)
+150000
diff --git a/benchmarks/sql_benchmarks/predicate_eval/results/neutral_q61_expensive_uniform.csv b/benchmarks/sql_benchmarks/predicate_eval/results/neutral_q61_expensive_uniform.csv
new file mode 100644
index 0000000000000..5d7624926c7c4
--- /dev/null
+++ b/benchmarks/sql_benchmarks/predicate_eval/results/neutral_q61_expensive_uniform.csv
@@ -0,0 +1,2 @@
+count(*)
+514286
diff --git a/benchmarks/sql_benchmarks/predicate_eval/results/nulls_q90_nullable_selective_first.csv b/benchmarks/sql_benchmarks/predicate_eval/results/nulls_q90_nullable_selective_first.csv
new file mode 100644
index 0000000000000..ce62ca3d04d63
--- /dev/null
+++ b/benchmarks/sql_benchmarks/predicate_eval/results/nulls_q90_nullable_selective_first.csv
@@ -0,0 +1,2 @@
+count(*)
+20000
diff --git a/benchmarks/sql_benchmarks/predicate_eval/results/nulls_q91_nullable_selective_last.csv b/benchmarks/sql_benchmarks/predicate_eval/results/nulls_q91_nullable_selective_last.csv
new file mode 100644
index 0000000000000..ce62ca3d04d63
--- /dev/null
+++ b/benchmarks/sql_benchmarks/predicate_eval/results/nulls_q91_nullable_selective_last.csv
@@ -0,0 +1,2 @@
+count(*)
+20000
diff --git a/benchmarks/sql_benchmarks/predicate_eval/results/scale_q50_5k.csv b/benchmarks/sql_benchmarks/predicate_eval/results/scale_q50_5k.csv
new file mode 100644
index 0000000000000..8cedd33498804
--- /dev/null
+++ b/benchmarks/sql_benchmarks/predicate_eval/results/scale_q50_5k.csv
@@ -0,0 +1,2 @@
+count(*)
+5
diff --git a/benchmarks/sql_benchmarks/predicate_eval/results/scale_q51_100k.csv b/benchmarks/sql_benchmarks/predicate_eval/results/scale_q51_100k.csv
new file mode 100644
index 0000000000000..5066201ce8c28
--- /dev/null
+++ b/benchmarks/sql_benchmarks/predicate_eval/results/scale_q51_100k.csv
@@ -0,0 +1,2 @@
+count(*)
+90
diff --git a/benchmarks/sql_benchmarks/predicate_eval/results/scale_q52_5m.csv b/benchmarks/sql_benchmarks/predicate_eval/results/scale_q52_5m.csv
new file mode 100644
index 0000000000000..41c4f54ec18eb
--- /dev/null
+++ b/benchmarks/sql_benchmarks/predicate_eval/results/scale_q52_5m.csv
@@ -0,0 +1,2 @@
+count(*)
+4461
diff --git a/benchmarks/sql_benchmarks/predicate_eval/results/scale_q53_50m.csv b/benchmarks/sql_benchmarks/predicate_eval/results/scale_q53_50m.csv
new file mode 100644
index 0000000000000..defd47ee70c37
--- /dev/null
+++ b/benchmarks/sql_benchmarks/predicate_eval/results/scale_q53_50m.csv
@@ -0,0 +1,2 @@
+count(*)
+44600
diff --git a/benchmarks/sql_benchmarks/predicate_eval/results/selectivity_q20_unselective_first.csv b/benchmarks/sql_benchmarks/predicate_eval/results/selectivity_q20_unselective_first.csv
new file mode 100644
index 0000000000000..23d7344bc8c8c
--- /dev/null
+++ b/benchmarks/sql_benchmarks/predicate_eval/results/selectivity_q20_unselective_first.csv
@@ -0,0 +1,2 @@
+count(*)
+50000
diff --git a/benchmarks/sql_benchmarks/predicate_eval/results/selectivity_q21_selective_first.csv b/benchmarks/sql_benchmarks/predicate_eval/results/selectivity_q21_selective_first.csv
new file mode 100644
index 0000000000000..23d7344bc8c8c
--- /dev/null
+++ b/benchmarks/sql_benchmarks/predicate_eval/results/selectivity_q21_selective_first.csv
@@ -0,0 +1,2 @@
+count(*)
+50000
diff --git a/benchmarks/sql_benchmarks/predicate_eval/results/width_q40_narrow.csv b/benchmarks/sql_benchmarks/predicate_eval/results/width_q40_narrow.csv
new file mode 100644
index 0000000000000..5eb0b761fa8b7
--- /dev/null
+++ b/benchmarks/sql_benchmarks/predicate_eval/results/width_q40_narrow.csv
@@ -0,0 +1,2 @@
+count(*)
+508
diff --git a/benchmarks/sql_benchmarks/predicate_eval/results/width_q41_wide.csv b/benchmarks/sql_benchmarks/predicate_eval/results/width_q41_wide.csv
new file mode 100644
index 0000000000000..5eb0b761fa8b7
--- /dev/null
+++ b/benchmarks/sql_benchmarks/predicate_eval/results/width_q41_wide.csv
@@ -0,0 +1,2 @@
+count(*)
+508
diff --git a/benchmarks/sql_benchmarks/predicate_eval/results/width_q42_xwide.csv b/benchmarks/sql_benchmarks/predicate_eval/results/width_q42_xwide.csv
new file mode 100644
index 0000000000000..5eb0b761fa8b7
--- /dev/null
+++ b/benchmarks/sql_benchmarks/predicate_eval/results/width_q42_xwide.csv
@@ -0,0 +1,2 @@
+count(*)
+508
diff --git a/benchmarks/sql_benchmarks/predicate_eval/scratch/.gitignore b/benchmarks/sql_benchmarks/predicate_eval/scratch/.gitignore
new file mode 100644
index 0000000000000..4bed5da93fb28
--- /dev/null
+++ b/benchmarks/sql_benchmarks/predicate_eval/scratch/.gitignore
@@ -0,0 +1 @@
+*.parquet