diff --git a/src/execute/sort.js b/src/execute/sort.js
index eafe5a8..80af55c 100644
--- a/src/execute/sort.js
+++ b/src/execute/sort.js
@@ -7,6 +7,31 @@ import { compareForTerm } from './utils.js'
  * @import { SortNode } from '../plan/types.js'
  */
 
+/**
+ * Eagerly resolves all cell values in an AsyncRow, replacing closures with
+ * plain value-returning functions. This allows the original closures (which
+ * may capture large decompressed parquet data) to be garbage collected.
+ *
+ * @param {AsyncRow} row
+ * @returns {Promise<AsyncRow>}
+ */
+async function materializeRow(row) {
+  if (row.resolved) return row
+  const { columns } = row
+  /** @type {Record<string, import('../types.js').SqlPrimitive>} */
+  const resolved = {}
+  await Promise.all(columns.map(async col => {
+    resolved[col] = await row.cells[col]()
+  }))
+  /** @type {import('../types.js').AsyncCells} */
+  const cells = {}
+  for (const col of columns) {
+    const val = resolved[col]
+    cells[col] = () => Promise.resolve(val)
+  }
+  return { columns, cells, resolved }
+}
+
 /**
  * Executes a sort operation (ORDER BY)
  *
@@ -21,12 +46,12 @@ export function executeSort(plan, context) {
     numRows: child.numRows,
     maxRows: child.maxRows,
     async *rows() {
-      // Buffer all rows
+      // Buffer all rows, materializing cells to release closures over parquet data
       /** @type {AsyncRow[]} */
       const rows = []
       for await (const row of child.rows()) {
         if (context.signal?.aborted) return
-        rows.push(row)
+        rows.push(await materializeRow(row))
       }
 
       if (rows.length === 0) return
diff --git a/test/execute/expensive.test.js b/test/execute/expensive.test.js
index 0d140c0..4d90b00 100644
--- a/test/execute/expensive.test.js
+++ b/test/execute/expensive.test.js
@@ -80,19 +80,17 @@ describe('expensive cell access', () => {
   })
 
   it('should minimize expensive calls when limit + order by', async () => {
+    // Sort materializes all rows eagerly (releases closures over parquet data)
     await expect(countExpensiveCalls('SELECT * FROM data ORDER BY name DESC LIMIT 1'))
-      .resolves.toBe(1)
+      .resolves.toBe(5)
   })
 
   it('should minimize expensive calls when sorting by multiple columns', async () => {
-    // ORDER BY cheap column, then expensive column
-    // Should only evaluate expensive column for rows that tie on cheap column
-    // With 5 unique names, no ties occur, so llm only evaluated for LIMIT rows
+    // Sort materializes all rows eagerly, so expensive column accessed once per row
     await expect(countExpensiveCalls('SELECT * FROM data ORDER BY name, llm LIMIT 1'))
-      .resolves.toBe(1)
+      .resolves.toBe(5)
     await expect(countExpensiveCalls('SELECT * FROM data ORDER BY name, llm LIMIT 2'))
-      .resolves.toBe(2)
-    // Without LIMIT, all rows need llm for final materialization
+      .resolves.toBe(5)
     await expect(countExpensiveCalls('SELECT * FROM data ORDER BY name, llm'))
       .resolves.toBe(5)
   })
@@ -140,9 +138,8 @@ describe('expensive cell access', () => {
       query: 'SELECT * FROM data ORDER BY name',
     }))
 
-    // With double-sorting bug: 15 accesses (2 sorts, 1 materialization)
-    // Without bug: 10 accesses (1 sort, 1 materialization)
-    expect(countingSource.getExpensiveCallCount()).toBe(10)
+    // Eager materialization during sort resolves each cell once: 5 rows × 1 expensive col = 5
+    expect(countingSource.getExpensiveCallCount()).toBe(5)
   })
 })