diff --git a/test/ml/plane-parity.test.ts b/test/ml/plane-parity.test.ts
new file mode 100644
index 0000000..f1e4c05
--- /dev/null
+++ b/test/ml/plane-parity.test.ts
@@ -0,0 +1,60 @@
+/**
+ * Neural-parity ts leg (WO14): the same golden sets the ruby and py
+ * legs run, same tolerance tiers. Env-gated — CI sets:
+ *   INTERSCRIPT_PLANE_ZIP=/tmp/model.zip
+ *   INTERSCRIPT_PLANE_GOLDEN=/tmp/golden/set.jsonl
+ *   INTERSCRIPT_PARITY_SMOKE=1   (corpus 2%; tight = per-row 2% + 0.5%)
+ */
+
+import { existsSync, readFileSync } from "node:fs"
+import { describe, expect, it } from "vitest"
+import { createPlaneModel } from "../../src/ml/index.js"
+
+const zipPath = process.env["INTERSCRIPT_PLANE_ZIP"]
+const goldenPath = process.env["INTERSCRIPT_PLANE_GOLDEN"]
+const smoke = process.env["INTERSCRIPT_PARITY_SMOKE"] === "1"
+
+const gated = zipPath && goldenPath && existsSync(zipPath) && existsSync(goldenPath)
+  ? describe
+  : describe.skip
+
+function lev(a: string, b: string): number {
+  const prev = Array.from({ length: b.length + 1 }, (_, j) => j)
+  for (let i = 1; i <= a.length; i++) {
+    const cur = [i]
+    for (let j = 1; j <= b.length; j++) {
+      cur[j] = Math.min(prev[j]! + 1, cur[j - 1]! + 1, prev[j - 1]! + (a[i - 1] !== b[j - 1] ? 1 : 0))
+    }
+    prev.splice(0, prev.length, ...cur)
+  }
+  return prev[b.length]!
+}
+
+gated("neural parity (ts plane leg)", () => {
+  it("matches the py-reference golden within the quantized tolerance", async () => {
+    const model = await createPlaneModel(new Uint8Array(readFileSync(zipPath!)))
+    const rows = readFileSync(goldenPath!, "utf8")
+      .split("\n")
+      .filter((l) => l.trim() !== "")
+      .map((l) => JSON.parse(l) as { input: string; output: string })
+
+    let corpusErrors = 0
+    let corpusChars = 0
+    let worstRowRate = 0
+    for (const row of rows) {
+      const out = await model.transform(row.input)
+      const dist = lev(out, row.output)
+      corpusErrors += dist
+      corpusChars += row.output.length
+      worstRowRate = Math.max(worstRowRate, dist / Math.max(1, row.output.length))
+    }
+    const corpusRate = corpusErrors / Math.max(1, corpusChars)
+    if (smoke) {
+      expect(corpusRate).toBeLessThan(0.02)
+    } else {
+      expect(worstRowRate).toBeLessThan(0.02)
+      expect(corpusRate).toBeLessThan(0.005)
+    }
+    await model.dispose()
+  })
+})