diff --git a/Complexity.lean b/Complexity.lean
index f73741f988..6ec0742cdb 100644
--- a/Complexity.lean
+++ b/Complexity.lean
@@ -160,6 +160,7 @@ public import Complexity.Algorithms.WordRAM.ExactTriangle.CountPrime.Spread
 public import Complexity.Algorithms.WordRAM.ExactTriangle.CountPrime.TimeBound
 public import Complexity.Algorithms.WordRAM.ExactTriangle.CountPrime.Width
 public import Complexity.Algorithms.WordRAM.ExactTriangle.DyadicCost
+public import Complexity.Algorithms.WordRAM.ExactTriangle.DyadicCost.Asymptotics
 public import Complexity.Algorithms.WordRAM.ExactTriangle.DyadicParameters
 public import Complexity.Algorithms.WordRAM.ExactTriangle.DyadicParameters.Exec
 public import Complexity.Algorithms.WordRAM.ExactTriangle.DyadicParameters.Input
diff --git a/Complexity/Algorithms/WordRAM/ExactTriangle/DyadicCost/Asymptotics.lean b/Complexity/Algorithms/WordRAM/ExactTriangle/DyadicCost/Asymptotics.lean
new file mode 100644
index 0000000000..e1a0a05b7f
--- /dev/null
+++ b/Complexity/Algorithms/WordRAM/ExactTriangle/DyadicCost/Asymptotics.lean
@@ -0,0 +1,156 @@
+/-
+Copyright (c) 2026 Erik Demaine. All rights reserved.
+Released under Apache 2.0 license as described in the file LICENSE.
+Authors: Erik Demaine
+-/
+
+module
+-- upstream: candidate
+
+public import Complexity.Algorithms.WordRAM.ExactTriangle.DyadicCost
+import Mathlib.Analysis.SpecialFunctions.Pow.Asymptotics
+import Mathlib.Analysis.SpecialFunctions.Log.Base
+import Mathlib.Tactic.Linarith
+import Mathlib.Tactic.Positivity
+import Mathlib.Tactic.Ring
+
+/-!
+# Uniform strict-subcubic component budgets
+
+Mathlib's logarithm-versus-real-power theorem absorbs the logarithmic coefficient of the
+computed dyadic native budgets. The resulting threshold depends only on the original weight
+exponent; it is shared by all matrices, selected primes and reference memories.
+
+This is an asymptotic theorem about the existing proved component charges. It does not assemble
+an execution, account for the callable adapter or dispatch small inputs, and does not assert
+the source's tuned 3SUM exponent.
+
+## Main results
+
+* `exists_components_subcubic_threshold`: beyond one threshold, the actual component allowance
+  is at most `4*n^(3-1/1296)`, uniformly over original weights bounded by `n^κ`.
+-/
+
+public section
+
+namespace Cslib.Algorithms.WordRAM.ExactTriangle.DyadicCost
+
+open Filter Cslib.Matrix Cslib.Algorithms.ExactTriangle
+open ParameterChoice
+
+private noncomputable def coefficient (κ : ℝ) (n : ℕ) : ℝ :=
+  238169 + 12 * (Nat.log 2 n : ℝ) + 138412032 * ((19 * level n + 1 : ℕ) : ℝ) +
+    (140 * falsePositiveConst + 140) * κ * Real.log n
+
+private noncomputable def logCoefficient (κ : ℝ) : ℝ :=
+  238169 + 138412032 + (12 + 19 * 138412032) / Real.log 2 +
+    (140 * falsePositiveConst + 140) * κ
+
+private theorem logCoefficient_pos {κ : ℝ} (hκ : 1 ≤ κ) : 0 < logCoefficient κ := by
+  have hC : 0 ≤ falsePositiveConst := (by norm_num : (0 : ℝ) ≤ 1).trans one_le_falsePositiveConst
+  have hk : 0 ≤ κ := by linarith
+  have hlog : 0 < Real.log 2 := Real.log_pos (by norm_num)
+  unfold logCoefficient
+  positivity
+
+private theorem coefficient_le_log {n : ℕ} {κ : ℝ} (hn : 16 ≤ n) :
+    coefficient κ n ≤ logCoefficient κ * Real.log n := by
+  have hnR : (16 : ℝ) ≤ n := by exact_mod_cast hn
+  have hlog1 : 1 ≤ Real.log n :=
+    (Real.le_log_iff_exp_le (by linarith)).mpr
+      (Real.exp_one_lt_three.le.trans (by linarith))
+  have hl : (level n : ℝ) ≤ Nat.log 2 n := by exact_mod_cast level_le_log n
+  have hb : (Nat.log 2 n : ℝ) ≤ Real.log n / Real.log 2 := by
+    simpa only [Real.logb, Nat.cast_ofNat] using Real.natLog_le_logb n 2
+  unfold coefficient logCoefficient
+  push_cast
+  have hlevel := mul_le_mul_of_nonneg_left hl (by norm_num : (0 : ℝ) ≤ 19 * 138412032)
+  have hconstant := mul_le_mul_of_nonneg_left hlog1
+    (by norm_num : (0 : ℝ) ≤ 238169 + 138412032)
+  calc
+    _ ≤ (238169 + 138412032) + (12 + 19 * 138412032) * (Nat.log 2 n : ℝ) +
+        (140 * falsePositiveConst + 140) * κ * Real.log n := by nlinarith only [hlevel]
+    _ ≤ (238169 + 138412032) + (12 + 19 * 138412032) * (Real.log n / Real.log 2) +
+        (140 * falsePositiveConst + 140) * κ * Real.log n :=
+      by gcongr
+    _ ≤ (238169 + 138412032) * Real.log n +
+        (12 + 19 * 138412032) * (Real.log n / Real.log 2) +
+        (140 * falsePositiveConst + 140) * κ * Real.log n :=
+      by
+        gcongr
+        simpa only [mul_one] using hconstant
+    _ = _ := by ring
+
+private theorem coefficient_eventually_le_rpow {κ ε : ℝ} (hκ : 1 ≤ κ) (hε : 0 < ε) :
+    ∀ᶠ n : ℕ in atTop, coefficient κ n ≤ (n : ℝ) ^ ε := by
+  have hc := logCoefficient_pos hκ
+  have hlog : (fun n : ℕ => Real.log (n : ℝ)) =o[atTop] (fun n => (n : ℝ) ^ ε) :=
+    (isLittleO_log_rpow_atTop hε).comp_tendsto tendsto_natCast_atTop_atTop
+  filter_upwards [eventually_ge_atTop (16 : ℕ), hlog.bound (inv_pos.mpr hc)] with n hn hsmall
+  have hlog0 : 0 ≤ Real.log (n : ℝ) := Real.log_nonneg (by exact_mod_cast (by omega : 1 ≤ n))
+  have hpow0 : 0 ≤ (n : ℝ) ^ ε := Real.rpow_nonneg (Nat.cast_nonneg n) ε
+  rw [Real.norm_eq_abs, abs_of_nonneg hlog0, Real.norm_eq_abs, abs_of_nonneg hpow0] at hsmall
+  calc
+    _ ≤ logCoefficient κ * Real.log n := coefficient_le_log hn
+    _ ≤ logCoefficient κ * ((logCoefficient κ)⁻¹ * (n : ℝ) ^ ε) :=
+      mul_le_mul_of_nonneg_left hsmall hc.le
+    _ = _ := by rw [← mul_assoc, mul_inv_cancel₀ hc.ne', one_mul]
+
+-- Permit the fixed 1296-bit fast-threshold numeral during filter transport.
+set_option exponentiation.threshold 1296 in
+private theorem components_eventually_le {κ ε : ℝ} (hκ : 1 ≤ κ) (hε : 0 < ε) :
+    ∀ᶠ n : ℕ in atTop, ∀ (hn : 0 < n) (values : List ℤ) (mem : ℕ → ℕ) (p : ℕ),
+      WeightsBoundedBy
+        (fun u v : Fin n => values.getD (pairIndex (u, v)) 0)
+        (fun u v => values.getD (n * n + pairIndex (u, v)) 0)
+        (fun u v => values.getD (2 * (n * n) + pairIndex (u, v)) 0) ((n : ℝ) ^ κ) →
+      IsSelectedPrime
+        (fun u v : Fin n => values.getD (pairIndex (u, v)) 0)
+        (fun u v => values.getD (n * n + pairIndex (u, v)) 0)
+        (fun u v => values.getD (2 * (n * n) + pairIndex (u, v)) 0) (window n) p →
+      ((12 * Nat.log 2 n + 36 + (PrimeSelection.timeBound n (window n) + 16) +
+          FixedPrime.timeBound (m := level n) hn (groups n) p values mem : ℕ) : ℝ) ≤
+        4 * (n : ℝ) ^ (3 - 1 / 648 + ε : ℝ) := by
+  filter_upwards [coefficient_eventually_le_rpow hκ hε,
+    eventually_ge_atTop (2 ^ 1296 : ℕ)] with n hc hfast
+  intro hn values mem p hv hp
+  have hf := (blocks_pos_iff n).mpr hfast
+  have hb := components_timeBound_le hf hκ values mem hv hp
+  change (_ : ℝ) ≤ coefficient κ n * ((n : ℝ) ^ 3 / groups n) at hb
+  calc
+    _ ≤ coefficient κ n * ((n : ℝ) ^ 3 / groups n) := hb
+    _ ≤ (n : ℝ) ^ ε * ((n : ℝ) ^ 3 / groups n) :=
+      mul_le_mul_of_nonneg_right hc (by positivity)
+    _ ≤ (n : ℝ) ^ ε * (4 * (n : ℝ) ^ (3 - 1 / 648 : ℝ)) :=
+      mul_le_mul_of_nonneg_left (cubic_div_groups_le hn) (by positivity)
+    _ = _ := by
+      rw [mul_left_comm, ← Real.rpow_add (by exact_mod_cast hn)]
+      congr 2
+      ring
+
+/-- A single threshold for each original weight exponent bounds all actual component charges
+strictly below cubic. It is independent of the original matrices, chosen prime, reference
+memory and physical width. The existing budget's positive-input proof is explicit; positivity
+also follows from the threshold. This does not assert an execution of a composed solver. -/
+theorem exists_components_subcubic_threshold {κ : ℝ} (hκ : 1 ≤ κ) :
+    ∃ N : ℕ, 2 ^ 1296 ≤ N ∧
+      ∀ n ≥ N, ∀ (hn : 0 < n) (values : List ℤ) (mem : ℕ → ℕ) (p : ℕ),
+        WeightsBoundedBy
+          (fun u v : Fin n => values.getD (pairIndex (u, v)) 0)
+          (fun u v => values.getD (n * n + pairIndex (u, v)) 0)
+          (fun u v => values.getD (2 * (n * n) + pairIndex (u, v)) 0) ((n : ℝ) ^ κ) →
+        IsSelectedPrime
+          (fun u v : Fin n => values.getD (pairIndex (u, v)) 0)
+          (fun u v => values.getD (n * n + pairIndex (u, v)) 0)
+          (fun u v => values.getD (2 * (n * n) + pairIndex (u, v)) 0) (window n) p →
+        ((12 * Nat.log 2 n + 36 + (PrimeSelection.timeBound n (window n) + 16) +
+            FixedPrime.timeBound (m := level n) hn (groups n) p values mem : ℕ) : ℝ) ≤
+          4 * (n : ℝ) ^ (3 - 1 / 1296 : ℝ) := by
+  have h := components_eventually_le hκ (by norm_num : (0 : ℝ) < 1 / 1296)
+  rw [show (3 - 1 / 648 + 1 / 1296 : ℝ) = 3 - 1 / 1296 by norm_num] at h
+  obtain ⟨N, hN⟩ := eventually_atTop.mp h
+  refine ⟨max N (2 ^ 1296), le_max_right _ _, ?_⟩
+  intro n hn
+  exact hN n ((le_max_left _ _).trans hn)
+
+end Cslib.Algorithms.WordRAM.ExactTriangle.DyadicCost
diff --git a/ComplexityTests/WordRAMExactTriangleSubcubicBudget.lean b/ComplexityTests/WordRAMExactTriangleSubcubicBudget.lean
new file mode 100644
index 0000000000..a543ceffa7
--- /dev/null
+++ b/ComplexityTests/WordRAMExactTriangleSubcubicBudget.lean
@@ -0,0 +1,77 @@
+/-
+Copyright (c) 2026 Erik Demaine. All rights reserved.
+Released under Apache 2.0 license as described in the file LICENSE.
+Authors: Erik Demaine
+-/
+
+-- upstream: local-only (tests)
+
+import Complexity.Algorithms.WordRAM.ExactTriangle.DyadicCost.Asymptotics
+import Mathlib.Data.List.GetD
+
+/-!
+# Closed uniform subcubic component-budget clients
+
+For original constant-weight matrices, the clients discharge every weight, geometry and
+selected-prime premise beyond a genuine threshold. Both zero and nonzero signed input words
+are covered. These are budget clients, not executions of an assembled solver.
+-/
+
+namespace ComplexityTests.WordRAMExactTriangleSubcubicBudget
+
+open Cslib.Matrix Cslib.Algorithms.ExactTriangle
+open Cslib.Algorithms.WordRAM.ExactTriangle
+open ParameterChoice
+
+private def data (n : ℕ) (z : ℤ) : List ℤ := List.replicate (3 * (n * n)) z
+
+private theorem pair_bound {n : ℕ} (a b : Fin n) : pairIndex (a, b) < n * n := by
+  have h := Nat.mul_le_mul_right n (Nat.succ_le_of_lt a.isLt)
+  dsimp only [pairIndex]
+  nlinarith only [h, b.isLt]
+
+private def ConstantBudget (z : ℤ) : Prop :=
+  ∃ N : ℕ, 2 ^ 1296 ≤ N ∧ ∀ n ≥ N, ∃ hn : 0 < n, ∃ p,
+    IsSelectedPrime (fun _ _ : Fin n => z) (fun _ _ => z) (fun _ _ => z) (window n) p ∧
+      ∀ mem : ℕ → ℕ,
+        ((12 * Nat.log 2 n + 36 + (PrimeSelection.timeBound n (window n) + 16) +
+            FixedPrime.timeBound (m := level n) hn (groups n) p (data n z) mem : ℕ) : ℝ) ≤
+          4 * (n : ℝ) ^ (3 - 1 / 1296 : ℝ)
+
+-- Permit normalization of the existing concrete fast-threshold numeral in this client.
+set_option exponentiation.threshold 1296 in
+private theorem constant_budget (z : ℤ) (hz : |z| ≤ 1) : ConstantBudget z := by
+  obtain ⟨N, hthreshold, hcost⟩ :=
+    DyadicCost.exists_components_subcubic_threshold (by norm_num : (1 : ℝ) ≤ 1)
+  refine ⟨N, hthreshold, ?_⟩
+  intro n hnN
+  have hn : 0 < n := (Nat.two_pow_pos 1296).trans_le (hthreshold.trans hnN)
+  have hf : 0 < blocks n := (blocks_pos_iff n).mpr (hthreshold.trans hnN)
+  obtain ⟨p, hp⟩ := exists_isSelectedPrime (fun _ _ : Fin n => z)
+    (fun _ _ => z) (fun _ _ => z) (window n) (window_ge_sixteen hf)
+  have hab (a b : Fin n) : (data n z).getD (pairIndex (a, b)) 0 = z :=
+    List.getD_replicate z (by have h := pair_bound a b; omega)
+  have hbc (a b : Fin n) : (data n z).getD (n * n + pairIndex (a, b)) 0 = z :=
+    List.getD_replicate z (by have h := pair_bound a b; omega)
+  have hac (a b : Fin n) : (data n z).getD (2 * (n * n) + pairIndex (a, b)) 0 = z :=
+    List.getD_replicate z (by have h := pair_bound a b; omega)
+  have hzR : ((|z| : ℤ) : ℝ) ≤ 1 := by exact_mod_cast hz
+  have hnR : (1 : ℝ) ≤ n := by exact_mod_cast hn
+  have hv : WeightsBoundedBy (fun _ _ : Fin n => z) (fun _ _ => z) (fun _ _ => z)
+      ((n : ℝ) ^ (1 : ℝ)) := by
+    rw [Real.rpow_one]
+    exact ⟨fun _ _ => hzR.trans hnR, fun _ _ => hzR.trans hnR, fun _ _ => hzR.trans hnR⟩
+  refine ⟨hn, p, hp, ?_⟩
+  intro mem
+  apply hcost n hnN hn (data n z) mem p
+  · simpa only [hab, hbc, hac] using hv
+  · simpa only [hab, hbc, hac] using hp
+
+-- A genuine zero-triangle family, with all cost hypotheses discharged.
+example : ConstantBudget 0 := constant_budget 0 (by norm_num)
+
+-- Nonzero original signed words use the same quantified component theorem.
+example : ConstantBudget 1 := constant_budget 1 (by norm_num)
+example : ConstantBudget (-1) := constant_budget (-1) (by norm_num)
+
+end ComplexityTests.WordRAMExactTriangleSubcubicBudget
diff --git a/UPSTREAM.md b/UPSTREAM.md
index f6773b93e4..c4c13ef8f0 100644
--- a/UPSTREAM.md
+++ b/UPSTREAM.md
@@ -331,8 +331,12 @@ Do not add a ledger row for an ordinary candidate or repeat its source marker he
   `ExactTriangle/DyadicCost.lean` (#2145) simplifies actual preparation/product/query charges
   to `n³/g`, absorbs the actual counter-based selector cost using `g^648≤n`, and applies the
   existing selected-prime failed-scan estimate once globally. Its combined component allowance
-  has an explicit logarithmic factor and a `3-1/648` real-power scale; logarithmic factors
-  are not absorbed into that exponent. It does not assert a composed execution theorem.
+  has an explicit logarithmic factor and a `3-1/648` real-power scale.
+  `DyadicCost/Asymptotics.lean` absorbs that factor via Mathlib's existing logarithm-growth
+  theorem: beyond a threshold depending only on the weight exponent, all actual component
+  allowances are at most `4*n^(3-1/1296)`, uniformly over matrices, selected primes and memories.
+  This existential asymptotic threshold is not an algorithm input. Neither result asserts
+  a composed execution theorem or the source's tuned exponent.
   Callable adaptation, small-case solver dispatch (#2146), outer host and #1394 remain open.
   `Algorithms/ExactTriangle/{Splitting,Witnesses}.lean` adapts the complete finite splitting and
   witness recovery component ([#1447](https://github.com/cslib-community/complexitylib/issues/1447)),