diff --git a/crates/guest-program/src/l2/output.rs b/crates/guest-program/src/l2/output.rs
index 8b64cafc2a9..f81e7eec77c 100644
--- a/crates/guest-program/src/l2/output.rs
+++ b/crates/guest-program/src/l2/output.rs
@@ -67,3 +67,124 @@ impl ProgramOutput {
         encoded
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use ethrex_common::types::balance_diff::BalanceDiff;
+
+    /// Reproducer for an L1↔guest public-input encoding mismatch.
+    ///
+    /// The L1 verifier in `OnChainProposer.sol::_getPublicInputsFromCommitment`
+    /// reconstructs the public inputs and `sha256`-hashes them. For each entry in
+    /// `currentBatch.l2InMessageRollingHashes`, the contract appends:
+    ///
+    /// ```solidity
+    /// abi.encodePacked(
+    ///     publicInputs,
+    ///     bytes32(rh.chainId),   // 32 bytes — chainId is uint256 in the struct
+    ///     rh.rollingHash         // 32 bytes
+    /// );
+    /// ```
+    ///
+    /// — i.e. **64 bytes per entry**.
+    ///
+    /// The guest-side `ProgramOutput::encode` (used as the SP1/RISC0 public
+    /// commitment via `commit_slice(&output.encode())`) appends:
+    ///
+    /// ```ignore
+    /// for (chain_id, hash) in &self.l2_in_message_rolling_hashes {
+    ///     encoded.extend_from_slice(&chain_id.to_be_bytes()); // u64 → 8 bytes
+    ///     encoded.extend_from_slice(&hash.to_fixed_bytes());  // 32 bytes
+    /// }
+    /// ```
+    ///
+    /// — only **40 bytes per entry**, because `chain_id` is typed as `u64`
+    /// (`Vec<(u64, H256)>`) and `u64::to_be_bytes()` returns 8 bytes.
+    ///
+    /// Consequence: as soon as a batch contains any L2-in privileged
+    /// transactions (i.e. any L2-to-L2 messaging is exercised), the prover
+    /// commits to a public input that is shorter than the one the L1
+    /// reconstructs, the two `sha256` hashes diverge, and the proof fails
+    /// verification — bricking L2-to-L2 batches. The `l1_in_messages_rolling_hash`
+    /// path doesn't have this issue (it's a single bytes32). The
+    /// `BalanceDiff::chain_id` and the top-level `chain_id` fields are fine
+    /// because both are `U256` and round-trip through `to_big_endian()` as 32
+    /// bytes — so the bug is local to the `(u64, H256)` typing of
+    /// `l2_in_message_rolling_hashes`.
+    ///
+    /// This test pins the current (buggy) byte layout. Two reasonable shapes
+    /// for the fix:
+    /// 1. Change the field type to `Vec<(U256, H256)>` and use
+    ///    `chain_id.to_big_endian()` like the other fields.
+    /// 2. Keep the type but pad on encode: write 24 zero bytes followed by
+    ///    `chain_id.to_be_bytes()` so the wire format is 32 bytes.
+    /// Both make the per-entry contribution 64 bytes, matching L1.
+    #[test]
+    fn l2_in_message_rolling_hashes_chain_id_is_only_8_bytes_not_32() {
+        // Build an output where every other field is empty / zero so we can
+        // isolate the `l2_in_message_rolling_hashes` contribution.
+        let chain_id: u64 = 0x1234_5678_9abc_def0;
+        let rolling_hash = H256([0xCC; 32]);
+
+        let output = ProgramOutput {
+            initial_state_hash: H256::zero(),
+            final_state_hash: H256::zero(),
+            l1_out_messages_merkle_root: H256::zero(),
+            l1_in_messages_rolling_hash: H256::zero(),
+            l2_in_message_rolling_hashes: vec![(chain_id, rolling_hash)],
+            blob_versioned_hash: H256::zero(),
+            last_block_hash: H256::zero(),
+            chain_id: U256::zero(),
+            non_privileged_count: U256::zero(),
+            balance_diffs: Vec::<BalanceDiff>::new(),
+        };
+
+        let encoded = output.encode();
+
+        // Fixed prefix: 8 H256/U256 fields × 32 bytes = 256 bytes.
+        const PREFIX_LEN: usize = 32 * 8;
+
+        // Current (buggy) per-entry size: 8 (u64) + 32 (H256) = 40 bytes.
+        // L1's expected per-entry size: 32 (bytes32 chainId) + 32 (H256) = 64 bytes.
+        assert_eq!(
+            encoded.len(),
+            PREFIX_LEN + 40,
+            "Guest commits 40 bytes per (chainId, rollingHash) entry; L1 reads 64. \
+             If this assertion now fails because encoded.len() == PREFIX_LEN + 64, \
+             the chain_id encoding was widened to bytes32 — update the assertion."
+        );
+
+        // Verify the chain_id bytes sit at the expected offset and are exactly
+        // u64::to_be_bytes (8 bytes, no padding).
+        let chain_id_bytes = &encoded[PREFIX_LEN..PREFIX_LEN + 8];
+        assert_eq!(
+            chain_id_bytes,
+            &chain_id.to_be_bytes(),
+            "chain_id should be the bare 8-byte big-endian u64 — confirming the bug",
+        );
+
+        // After the chain_id (offset PREFIX_LEN+8), the rolling hash starts
+        // immediately. L1 expects it at PREFIX_LEN+32.
+        let hash_bytes = &encoded[PREFIX_LEN + 8..PREFIX_LEN + 8 + 32];
+        assert_eq!(
+            hash_bytes, &rolling_hash.0,
+            "rollingHash starts right after the 8-byte chain_id, with no left-pad",
+        );
+
+        // Build what L1's `abi.encodePacked(bytes32(chainId), rollingHash)` would
+        // produce so the diff is explicit in the test output.
+        let mut expected_l1_per_entry = [0u8; 64];
+        // bytes32(uint256(chainId)): left-pad u64 to 32 bytes.
+        expected_l1_per_entry[24..32].copy_from_slice(&chain_id.to_be_bytes());
+        expected_l1_per_entry[32..64].copy_from_slice(&rolling_hash.0);
+
+        let guest_per_entry = &encoded[PREFIX_LEN..];
+        assert_ne!(
+            guest_per_entry,
+            expected_l1_per_entry.as_slice(),
+            "Guest and L1 disagree on the per-entry layout — this is the bug. \
+             After the fix, this assertion will flip to assert_eq!.",
+        );
+    }
+}