diff --git a/backends/v3/Cargo.toml b/backends/v3/Cargo.toml
index 996290ed3ba..588a2716fe1 100644
--- a/backends/v3/Cargo.toml
+++ b/backends/v3/Cargo.toml
@@ -71,6 +71,7 @@ prost-build = "0.12.1"
 [dev-dependencies]
 criterion = "0.3"
 itertools = "0.13"
+rustc-hash = "2"
 
 [features]
 default = ["ngrok"]
diff --git a/backends/v3/src/radix.rs b/backends/v3/src/radix.rs
index 532ec6ddcc8..e697dc01bd4 100644
--- a/backends/v3/src/radix.rs
+++ b/backends/v3/src/radix.rs
@@ -208,6 +208,7 @@ impl Allocator for RadixAllocator {
     }
 }
 
+#[derive(Debug)]
 struct RadixAllocation {
     prefix_node: NodeId,
     cached_prefix_len: usize,
@@ -631,6 +632,12 @@ fn shared_prefix(left: &[u32], right: &[u32], block_size: usize) -> usize {
 mod tests {
     use std::sync::Arc;
 
+    use rand::{
+        distributions::Uniform, prelude::Distribution, rngs::SmallRng, seq::SliceRandom,
+        SeedableRng,
+    };
+    use rustc_hash::FxHashSet;
+
     use super::*;
 
     #[test]
@@ -873,4 +880,114 @@ mod tests {
         // Clear out the whole trie.
         assert_eq!(trie.evict(10), vec![1, 2, 3, 0, 1]);
     }
+
+    #[test]
+    fn invariants_hold_on_many_insertions() {
+        const VOCAB_SIZE: u32 = 2;
+        const DATA_LEN: usize = 1_000_000;
+
+        //const MAX_PREFILL_LEN: usize = 2usize.pow(12);
+        const MAX_PREFILL_LEN: usize = 2usize.pow(14);
+        const MAX_DECODE_LEN: usize = 2usize.pow(4);
+
+        let vocab_range = Uniform::new(0, VOCAB_SIZE);
+        let data_range = Uniform::new(0, DATA_LEN);
+        let prefill_len_range = Uniform::new(0, MAX_PREFILL_LEN);
+        let decode_len_range = Uniform::new(0, MAX_DECODE_LEN);
+
+        let mut rng = SmallRng::seed_from_u64(64);
+        let data = (0..DATA_LEN)
+            .map(|_| vocab_range.sample(&mut rng))
+            .collect::<Vec<_>>();
+        let mut allocator = RadixAllocator::new(1, 4_000_000, None);
+
+        let mut allocations = Vec::new();
+
+        for i in 0..500 {
+            // Allocate until all blocks are used.
+            'allocation: loop {
+                // Use offset 0 half of the times for prefix sharing.
+                let prefill_offset = *[0, 0, 1, data_range.sample(&mut rng)]
+                    .choose(&mut rng)
+                    .unwrap();
+                let prefill_len = prefill_len_range.sample(&mut rng);
+                let decode_len = decode_len_range.sample(&mut rng);
+
+                let prefill =
+                    data[prefill_offset..data.len().min(prefill_offset + prefill_len)].to_vec();
+
+                let allocation = match allocator
+                    .allocate((prefill.len() + decode_len) as u32, Some(Arc::new(prefill)))
+                {
+                    Some(allocation) => allocation,
+                    None => break 'allocation,
+                };
+                let prefix_blocks = allocation.blocks[..allocation.prefix_len as usize]
+                    .iter()
+                    .copied()
+                    .collect::<FxHashSet<_>>();
+                let non_prefix_blocks = allocation.blocks[allocation.prefix_len as usize..]
+                    .iter()
+                    .copied()
+                    .collect::<FxHashSet<_>>();
+                let blockset = allocation.blocks.iter().copied().collect::<FxHashSet<_>>();
+
+                allocations.push((allocation, blockset, prefix_blocks, non_prefix_blocks));
+            }
+
+            // Check invariants. Skip first iteration, since there is no prefix sharing yet.
+            if i > 1 {
+                check_allocation_invariants(&allocations);
+            }
+
+            // Remove 20% of the allocations, randomly.
+            allocations.shuffle(&mut rng);
+            let remove_index = (allocations.len() as f64 * 0.8) as usize;
+            for (allocation, _, _, _) in allocations.drain(remove_index..) {
+                allocator.free(allocation.blocks.clone(), allocation.allocation_id);
+            }
+        }
+    }
+
+    fn check_allocation_invariants(
+        allocations: &[(
+            BlockAllocation,
+            FxHashSet<u32>,
+            FxHashSet<u32>,
+            FxHashSet<u32>,
+        )],
+    ) {
+        for i in 0..allocations.len() {
+            let (allocation, blockset, prefix_blocks, non_prefix_blocks) = &allocations[i];
+
+            // 0 is used for health checks, must not be used.
+            assert!(!blockset.contains(&0), "Block 0 must not be allocated");
+
+            // No duplicate blocks in an allocation.
+            assert_eq!(
+                allocation.blocks.len(),
+                blockset.len(),
+                "Duplicate blocks in allocation"
+            );
+
+            //eprintln!(
+            //    "Prefix blocks, non-prefix blocks: {} {}",
+            //    prefix_blocks.len(),
+            //    non_prefix_blocks.len()
+            //);
+
+            for (_, _, other_prefix_blocks, other_non_prefix_blocks) in &allocations[i + 1..] {
+                if !other_non_prefix_blocks.is_disjoint(non_prefix_blocks) {
+                    eprintln!(
+                        "overlapping: {:?}",
+                        non_prefix_blocks.intersection(other_non_prefix_blocks),
+                    );
+                }
+                assert!(
+                    other_non_prefix_blocks.is_disjoint(non_prefix_blocks),
+                    "Allocations share non-prefix blocks"
+                )
+            }
+        }
+    }
 }