diff --git a/LeanFrontier/NumberTheory/MarkovTree.lean b/LeanFrontier/NumberTheory/MarkovTree.lean
new file mode 100644
index 0000000..c89eb1a
--- /dev/null
+++ b/LeanFrontier/NumberTheory/MarkovTree.lean
@@ -0,0 +1,137 @@
+import LeanFrontier.NumberTheory.MarkovEquation
+
+/-!
+# Descent for ordered positive Markov triples
+
+The Vieta involution from `LeanFrontier.NumberTheory.MarkovEquation`
+is the local move underlying the Markov tree.
+
+For a positive solution ordered `x ≤ y ≤ z`, the exceptional root
+`(1, 1, 1)` is the only case in which the largest two coordinates can
+coincide without a descending Vieta move. Away from that root, jumping
+the largest coordinate produces another positive coordinate at or below
+the middle one, and hence strictly decreases the largest coordinate.
+-/
+
+namespace LeanFrontier.MarkovTree
+
+private theorem root_of_middle_eq_largest
+    {x y z : ℤ}
+    (hx : 0 < x)
+    (hxy : x ≤ y)
+    (hyz_eq : y = z)
+    (h : MarkovEquation.IsSolution x y z) :
+    x = 1 ∧ y = 1 ∧ z = 1 := by
+  subst z
+  unfold MarkovEquation.IsSolution at h
+
+  have hy : 0 < y := lt_of_lt_of_le hx hxy
+
+  have hsq : x ^ 2 ≤ y ^ 2 := by
+    have hprod : 0 ≤ (y - x) * (y + x) := by
+      exact mul_nonneg (sub_nonneg.mpr hxy) (by linarith)
+    nlinarith
+
+  have hx1 : 1 ≤ x := by
+    have h1 : 0 + 1 ≤ x := Int.add_one_le_iff.mpr hx
+    simpa using h1
+
+  have hx_lt_two : x < 2 := by
+    by_contra hx_not
+    have hx2 : 2 ≤ x := le_of_not_gt hx_not
+    nlinarith
+
+  have hx_le_one : x ≤ 1 := by
+    have hlt : x < 1 + 1 := by simpa using hx_lt_two
+    exact Int.lt_add_one_iff.mp hlt
+
+  have hx_eq : x = 1 := le_antisymm hx_le_one hx1
+  subst x
+
+  have hy1 : 1 ≤ y := by
+    have h1 : 0 + 1 ≤ y := Int.add_one_le_iff.mpr hy
+    simpa using h1
+
+  have hy_le : y ≤ 1 := by
+    by_contra hy_not
+    have hy_gt_one : 1 < y := lt_of_not_ge hy_not
+    have hy2' : 1 + 1 ≤ y := Int.add_one_le_iff.mpr hy_gt_one
+    have hy2 : 2 ≤ y := by simpa using hy2'
+    nlinarith
+
+  have hy_eq : y = 1 := le_antisymm hy_le hy1
+  subst y
+
+  exact ⟨rfl, rfl, rfl⟩
+
+/--
+For a positive ordered Markov triple other than `(1, 1, 1)`, the Vieta
+jump in the largest coordinate is again positive, lies at or below the
+middle coordinate, and therefore strictly decreases the largest coordinate.
+-/
+theorem jump_descends_ordered_positive
+    {x y z : ℤ}
+    (hx : 0 < x)
+    (hxy : x ≤ y)
+    (hyz : y ≤ z)
+    (h : MarkovEquation.IsSolution x y z)
+    (hne : ¬ (x = 1 ∧ y = 1 ∧ z = 1)) :
+    0 < MarkovEquation.jump x y z ∧
+      MarkovEquation.jump x y z ≤ y ∧
+      MarkovEquation.jump x y z < z := by
+  have hy : 0 < y := lt_of_lt_of_le hx hxy
+  have hz : 0 < z := lt_of_lt_of_le hy hyz
+
+  have hyz_ne : y ≠ z := by
+    intro hyz_eq
+    exact hne (root_of_middle_eq_largest hx hxy hyz_eq h)
+
+  have hyz_lt : y < z := lt_of_le_of_ne hyz hyz_ne
+
+  have hjump_pos : 0 < MarkovEquation.jump x y z :=
+    MarkovEquation.jump_pos hx hz h
+
+  have hsq : x ^ 2 ≤ y ^ 2 := by
+    have hprod : 0 ≤ (y - x) * (y + x) := by
+      exact mul_nonneg (sub_nonneg.mpr hxy) (by linarith)
+    nlinarith
+
+  have hx1 : 1 ≤ x := by
+    have h1 : 0 + 1 ≤ x := Int.add_one_le_iff.mpr hx
+    simpa using h1
+
+  have hx_minus_one : 0 ≤ x - 1 := sub_nonneg.mpr hx1
+  have htail : 0 ≤ (x - 1) * y ^ 2 :=
+    mul_nonneg hx_minus_one (sq_nonneg y)
+
+  have hpoly : x ^ 2 + 2 * y ^ 2 - 3 * x * y ^ 2 ≤ 0 := by
+    nlinarith
+
+  have hfactor :
+      (y - z) * (y - MarkovEquation.jump x y z) =
+        x ^ 2 + 2 * y ^ 2 - 3 * x * y ^ 2 := by
+    calc
+      (y - z) * (y - MarkovEquation.jump x y z) =
+          y ^ 2 - 3 * x * y ^ 2 + z * MarkovEquation.jump x y z := by
+            rw [MarkovEquation.jump]
+            ring
+      _ = x ^ 2 + 2 * y ^ 2 - 3 * x * y ^ 2 := by
+            rw [MarkovEquation.mul_jump_eq h]
+            ring
+
+  have hleft_neg : y - z < 0 := by linarith
+
+  have hjump_le : MarkovEquation.jump x y z ≤ y := by
+    have hright_nonneg : 0 ≤ y - MarkovEquation.jump x y z := by
+      by_contra hn
+      have hright_neg : y - MarkovEquation.jump x y z < 0 :=
+        lt_of_not_ge hn
+      have hpositive :
+          0 < (y - z) * (y - MarkovEquation.jump x y z) :=
+        mul_pos_of_neg_of_neg hleft_neg hright_neg
+      nlinarith
+    exact sub_nonneg.mp hright_nonneg
+
+  exact ⟨hjump_pos, hjump_le, lt_of_le_of_lt hjump_le hyz_lt⟩
+
+end LeanFrontier.MarkovTree
diff --git a/Submissions/markov-tree-descent.json b/Submissions/markov-tree-descent.json
new file mode 100644
index 0000000..ff024ee
--- /dev/null
+++ b/Submissions/markov-tree-descent.json
@@ -0,0 +1,17 @@
+{
+  "protocol_version": "0.1",
+  "submission_id": "markov-tree-descent",
+  "producer": {
+    "type": "autonomous_agent",
+    "model": "gpt-5.6-sol",
+    "agent": "chatgpt"
+  },
+  "origin_mode": "target_driven",
+  "statement_origin": "machine",
+  "proof_origin": "machine",
+  "entrypoints": [
+    "LeanFrontier.MarkovTree.jump_descends_ordered_positive"
+  ],
+  "base_mathlib_revision": "v4.33.1",
+  "source_context": "Extension of the accepted Markov-equation Vieta-jumping submission. The agent independently surveyed the LeanFrontier corpus and identified several extension directions it considered worthwhile, including Markov-tree descent, a Ford-circle/Mathlib geometry bridge, a Stern-Brocot positive-rational development, and Thue-Morse cube-freeness. It recommended Markov-tree descent as the best combination of mathematical interest, reuse of accepted LeanFrontier results, and realistic formalization scope. The human operator then selected that recommended Markov-tree direction from the agent's shortlist; the human did not author or materially edit the formal theorem statement or proof. The formal statement and Lean proof in this submission were authored by the agent. The accepted MarkovEquation module deliberately leaves the tree's descent ordering for separate work. This submission proves the local descent step for a positive ordered Markov triple x <= y <= z away from the root (1,1,1): the existing Vieta jump in the largest coordinate stays positive, is at most the middle coordinate, and therefore strictly decreases the largest coordinate. The proof imports and depends on the accepted MarkovEquation API, in particular jump_pos and mul_jump_eq, rather than restating the Vieta involution. No claim of new mathematics is made; this is a formalization of the classical Markov-tree descent mechanism."
+}