diff --git a/CHANGELOG.md b/CHANGELOG.md
index 4f4dccda3a..2f09d9ff5e 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -10,6 +10,7 @@
 #### Bug Fixes
 
 - Fixed debug-only underflow in memory range-check trace generation when the first memory access is at `clk = 0` ([#2976](https://github.com/0xMiden/miden-vm/pull/2976)).
+- Aligned replay stack word access bounds with `StackInterface`, allowing the maximum valid start index for word reads and writes ([#3014](https://github.com/0xMiden/miden-vm/pull/3014)).
 - Replaced unsound `ptr::read` with safe unbox in panic recovery, removing UB from potential double-drop ([#2934](https://github.com/0xMiden/miden-vm/pull/2934)).
 - Reverted `InvokeKind::ProcRef` back to `InvokeKind::Exec` in `visit_mut_procref` and added an explanatory comment (#2893).
 - Fixed the release dry-run publish cycle between `miden-air` and `miden-ace-codegen`, and preserved leaf-only DAG imports with explicit snapshots ([#2931](https://github.com/0xMiden/miden-vm/pull/2931)).
diff --git a/processor/src/trace/parallel/processor.rs b/processor/src/trace/parallel/processor.rs
index 4007e7b88c..89c92291af 100644
--- a/processor/src/trace/parallel/processor.rs
+++ b/processor/src/trace/parallel/processor.rs
@@ -255,7 +255,7 @@ impl StackInterface for ReplayProcessor {
     }
 
     fn get_word(&self, start_idx: usize) -> Word {
-        debug_assert!(start_idx < MIN_STACK_DEPTH - 4);
+        debug_assert!(start_idx <= MIN_STACK_DEPTH - WORD_SIZE);
 
         let word_start_idx = MIN_STACK_DEPTH - start_idx - 4;
         let mut result: [Felt; WORD_SIZE] =
@@ -274,7 +274,7 @@ impl StackInterface for ReplayProcessor {
     }
 
     fn set_word(&mut self, start_idx: usize, word: &Word) {
-        debug_assert!(start_idx < MIN_STACK_DEPTH - 4);
+        debug_assert!(start_idx <= MIN_STACK_DEPTH - WORD_SIZE);
         let word_start_idx = MIN_STACK_DEPTH - start_idx - 4;
 
         // Reverse so word[0] ends up at the top of stack (highest internal index)
@@ -506,3 +506,45 @@ impl Stopper for ReplayStopper {
         }
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::trace::trace_state::{
+        AdviceReplay, ExecutionContextReplay, HasherResponseReplay, MastForestResolutionReplay,
+        MemoryReadsReplay, StackOverflowReplay, StackState, SystemState,
+    };
+
+    fn build_replay_processor() -> ReplayProcessor {
+        let system = SystemState {
+            clk: 0_u32.into(),
+            ctx: ContextId::root(),
+            fn_hash: Word::default(),
+            pc_transcript_state: Word::default(),
+        };
+        let stack = StackState::new([ZERO; MIN_STACK_DEPTH], MIN_STACK_DEPTH, ZERO);
+
+        ReplayProcessor::new(
+            system,
+            stack,
+            StackOverflowReplay::default(),
+            ExecutionContextReplay::default(),
+            AdviceReplay::default(),
+            MemoryReadsReplay::default(),
+            HasherResponseReplay::default(),
+            MastForestResolutionReplay::default(),
+            1_u32.into(),
+        )
+    }
+
+    #[test]
+    fn stack_set_word_allows_max_start_idx() {
+        let mut processor = build_replay_processor();
+        let start_idx = MIN_STACK_DEPTH - WORD_SIZE;
+        let word = [Felt::new(1), Felt::new(2), Felt::new(3), Felt::new(4)].into();
+
+        processor.set_word(start_idx, &word);
+
+        assert_eq!(processor.get_word(start_idx), word);
+    }
+}