diff --git a/.github/workflows/bench-main.yml b/.github/workflows/bench-main.yml
index 17a90125..f5fb8d87 100644
--- a/.github/workflows/bench-main.yml
+++ b/.github/workflows/bench-main.yml
@@ -245,8 +245,8 @@ jobs:
   #   - decompile: the inverse of compile. Roundtrip correctness is
   #     checked elsewhere (`ix validate` / roundtrip tests) — this only
   #     measures speed and memory.
-  #   - aiur-recursive: proves the IxVM typecheck of two fixed constants
-  #     (Nat.add_comm, Array.extract_append), runs the in-circuit
+  #   - aiur-recursive: proves the IxVM typecheck of three fixed constants
+  #     (Nat.add_comm, Vector.append, String.split), runs the in-circuit
   #     verifier over each proof, then proves THAT — the cost of
   #     recursion at kernel scale. A fixed subset instead of the
   #     Vectors.csv fan-out: its `env` field names the (already-cached)
diff --git a/.github/workflows/bench-pr.yml b/.github/workflows/bench-pr.yml
index db88afa8..a4e19ad4 100644
--- a/.github/workflows/bench-pr.yml
+++ b/.github/workflows/bench-pr.yml
@@ -32,9 +32,9 @@
 #   - zisk / sp1 / ooc: `execute`.
 #   - compile: `ix compile <env>.lean` → `<env>.ixe`.
 #   - decompile: `ix decompile` over the compile run's fresh `.ixe`.
-#   - aiur-recursive: a fixed IxVM statement (Nat.add_comm; the heavy-tier
-#     Array.extract_append is dropped for now — significantly more
-#     expensive even at 50 queries) proved and recursively verified via
+#   - aiur-recursive: fixed IxVM statements (Nat.add_comm plus the
+#     heavy-tier Vector.append and String.split; Array.extract_append
+#     is dropped for now — it OOMs on CI) proved and recursively verified via
 #     bench-typecheck --recursive; always schedules exactly one run
 #     regardless of BENCH_ENVS (the constants resolve in whichever env's
 #     .ixe the entry carries).
diff --git a/Ix/Cli/BenchCmd.lean b/Ix/Cli/BenchCmd.lean
index 9df9c336..cdfb361a 100644
--- a/Ix/Cli/BenchCmd.lean
+++ b/Ix/Cli/BenchCmd.lean
@@ -223,7 +223,7 @@ def backendSpecs : List BackendSpec := [
   -- large closures exceed the RAM ceiling and land as `status: oom` rows
   -- that bmf drops), which is why it's marked `unscheduled`: a testbed for
   -- local `--mode recursive` runs only — never uploaded to bencher, never
-  -- plotted. CI tracks the same measurement over the fixed two-constant
+  -- plotted. CI tracks the same measurement over the fixed three-constant
   -- subset via the aiur-recursive backend below.
   { name := "aiur", defaultMode := "prove", inputs := .perConstant,
     testbeds := [("prove", "aiur-check-prove-x64-32x"),
@@ -251,7 +251,7 @@ def backendSpecs : List BackendSpec := [
   -- `recursiveConstants`): IxVM recursion on real statements — prove each
   -- constant's typecheck, execute the in-circuit multi-stark verifier
   -- over the fresh proof, then prove THAT execution. The same measurement
-  -- as the aiur recursive mode above, but over a fixed two-constant
+  -- as the aiur recursive mode above, but over a fixed three-constant
   -- subset so CI schedules exactly one entry instead of the whole
   -- Vectors.csv fan-out.
   { name := "aiur-recursive", defaultMode := "prove", inputs := .fixedConfigs,
@@ -343,18 +343,18 @@ def findBackend (name : String) : Option BackendSpec :=
 /-- The `aiur-recursive` backend's rows: fixed IxVM statements, proved and
     recursively verified by `bench-typecheck --recursive` under the
     recursion-tuned parameters (50 queries at log-blowup 2, ~100-bit
-    soundness — this benchmarks *secure* recursion). Currently only
-    `Nat.add_comm` (cheap tier — the small end of the IxVM cost range):
-    `Array.extract_append` (heavy tier — kernel scale) is dropped for now,
-    its recursion being significantly more expensive even at 50 queries;
-    re-add it when the verifier prover slims down. A stage that exceeds
-    the CI RAM ceiling lands as an OOM row — the honest signal that
-    secure recursion at that scale does not yet fit. (Measured 2026-08 at
-    100 queries: `Nat.add_comm`'s outer prove peaked ~195 GiB on the
-    ~123 GiB runners; the 50-query halving is what gives it a chance to
-    fit.) -/
+    soundness — this benchmarks *secure* recursion). `Nat.add_comm`
+    (cheap tier) is the small end of the IxVM cost range;
+    `Vector.append` and `String.split` (heavy tier — kernel scale) are
+    the expensive end. `Array.extract_append` is dropped for now — it
+    OOMs on CI even at 50 queries; re-add it when the verifier prover
+    slims down. A stage that exceeds the CI RAM ceiling lands as an OOM
+    row — the honest signal that secure recursion at that scale does not
+    yet fit. (Measured 2026-08 at 100 queries: `Nat.add_comm`'s outer
+    prove peaked ~195 GiB on the ~123 GiB runners; the 50-query halving
+    is what gives it a chance to fit.) -/
 def recursiveConstants : List String :=
-  ["Nat.add_comm"]
+  ["Nat.add_comm", "Vector.append", "String.split"]
 
 def BackendSpec.testbedFor (b : BackendSpec) (mode : String) : Option String :=
   (b.testbeds.find? (·.1 == mode)).map (·.2)