diff --git a/crates/blockchain/src/block_builder.rs b/crates/blockchain/src/block_builder.rs
index 919a7719..9af5ff83 100644
--- a/crates/blockchain/src/block_builder.rs
+++ b/crates/blockchain/src/block_builder.rs
@@ -117,7 +117,8 @@ pub(crate) fn build_block(
     let compacted = if config.enable_proposer_aggregation {
         compact_attestations(selected, head_state, slot)?
     } else {
-        keep_best_proof_per_data(selected, slot)
+        let running_votes = build_running_votes(head_state);
+        keep_best_proof_per_data(selected, &running_votes, slot)
     };
     metrics::observe_block_proposal_phase("compact", compact_start.elapsed());
 
@@ -661,30 +662,28 @@ fn compact_attestations(
 ///
 /// The block format permits at most one entry per `AttestationData`: `on_block`
 /// rejects duplicates (`StoreError::DuplicateAttestationData`). When proposer
-/// aggregation is disabled we therefore cannot keep every selected proof, nor
-/// can we merge them. For each group sharing an `AttestationData` we keep the
-/// single proof covering the most validators (ties broken by first occurrence)
-/// and drop the rest. No leanVM aggregation runs; coverage is whatever the best
-/// individual proof already had, which is the cost of skipping aggregation.
+/// aggregation is disabled we can neither keep every selected proof nor merge
+/// them, so for each group sharing an `AttestationData` exactly one proof
+/// survives and the rest are dropped. No leanVM aggregation runs.
+///
+/// The surviving proof is the one adding the most NEW voters, with ties
+/// broken by larger absolute participant count, then first occurrence.
 fn keep_best_proof_per_data(
     entries: Vec<(AggregatedAttestation, SingleMessageAggregate)>,
+    running_votes: &HashMap<H256, HashSet<u64>>,
     block_slot: u64,
 ) -> Vec<(AggregatedAttestation, SingleMessageAggregate)> {
-    // Preserve first-occurrence order of distinct AttestationData; for each,
-    // track the index of the best (most participants) entry seen so far.
+    // Group entry indices by AttestationData, preserving first-occurrence order.
     let mut order: Vec<AttestationData> = Vec::new();
-    let mut best_index: HashMap<AttestationData, usize> = HashMap::new();
+    let mut groups: HashMap<AttestationData, Vec<usize>> = HashMap::new();
     for (i, (att, _)) in entries.iter().enumerate() {
-        match best_index.entry(att.data.clone()) {
+        match groups.entry(att.data.clone()) {
             std::collections::hash_map::Entry::Vacant(e) => {
                 order.push(e.key().clone());
-                e.insert(i);
+                e.insert(vec![i]);
             }
             std::collections::hash_map::Entry::Occupied(mut e) => {
-                let current_best = entries[*e.get()].0.aggregation_bits.count_ones();
-                if att.aggregation_bits.count_ones() > current_best {
-                    e.insert(i);
-                }
+                e.get_mut().push(i);
             }
         }
     }
@@ -702,15 +701,44 @@ fn keep_best_proof_per_data(
         "Skipping attestation compaction"
     );
 
+    // Pick the surviving index per group: most new coverage (over in-state
+    // voters plus those already claimed by earlier groups this block), then
+    // densest proof, then earliest occurrence. Computed over `entries` before
+    // it is moved into `items` for extraction.
+    let mut claimed: HashMap<H256, HashSet<u64>> = HashMap::new();
+    let mut best_per_data: Vec<usize> = Vec::with_capacity(order.len());
+    for data in &order {
+        let target_root = data.target.root;
+        let prior = running_votes.get(&target_root);
+        let block_claimed = claimed.get(&target_root);
+        let best = *groups[data]
+            .iter()
+            .max_by_key(|&&idx| {
+                let (att, proof) = &entries[idx];
+                let marginal = proof
+                    .participant_indices()
+                    .filter(|vid| {
+                        prior.is_none_or(|voted| !voted.contains(vid))
+                            && block_claimed.is_none_or(|voted| !voted.contains(vid))
+                    })
+                    .count();
+                (marginal, att.aggregation_bits.count_ones(), Reverse(idx))
+            })
+            .expect("group is non-empty");
+        // Record the winner's voters so a later group sharing this target root
+        // does not treat them as new coverage.
+        claimed
+            .entry(target_root)
+            .or_default()
+            .extend(entries[best].1.participant_indices());
+        best_per_data.push(best);
+    }
+
     let mut items: Vec<Option<(AggregatedAttestation, SingleMessageAggregate)>> =
         entries.into_iter().map(Some).collect();
-    order
-        .iter()
-        .map(|data| {
-            items[best_index[data]]
-                .take()
-                .expect("best index taken once")
-        })
+    best_per_data
+        .into_iter()
+        .map(|idx| items[idx].take().expect("best index taken once"))
         .collect()
 }
 
@@ -1320,6 +1348,173 @@ mod tests {
         );
     }
 
+    /// When several proofs share one `AttestationData` but the proposer cannot
+    /// aggregate, only one may be kept. Selecting by absolute participant count
+    /// starves a small subnet whose votes are the only ones still missing from
+    /// the target's in-state coverage: a larger subnet already counted in-state
+    /// keeps winning yet adds nothing, so the target never reaches 2/3.
+    ///
+    /// This mirrors the union coverage `compact_attestations` achieves when
+    /// aggregation is enabled: the kept proof must be the one adding the most
+    /// NEW voters over the target's in-state voter set, not the largest one.
+    #[test]
+    fn keep_best_proof_per_data_prefers_marginal_coverage_over_absolute_size() {
+        let target_root = H256([9u8; 32]);
+        let att_data = AttestationData {
+            slot: 5,
+            head: Checkpoint::default(),
+            target: Checkpoint {
+                slot: 5,
+                root: target_root,
+            },
+            source: Checkpoint::default(),
+        };
+
+        // Two candidate proofs for the same data: a 3-validator subnet already
+        // fully counted in-state, and a 2-validator subnet {3, 7} that is the
+        // only remaining new coverage.
+        let large = (
+            AggregatedAttestation {
+                aggregation_bits: make_bits(&[0, 1, 2]),
+                data: att_data.clone(),
+            },
+            SingleMessageAggregate::empty(make_bits(&[0, 1, 2])),
+        );
+        let small = (
+            AggregatedAttestation {
+                aggregation_bits: make_bits(&[3, 7]),
+                data: att_data.clone(),
+            },
+            SingleMessageAggregate::empty(make_bits(&[3, 7])),
+        );
+        let entries = vec![large, small];
+
+        // In-state, validators {0, 1, 2} already voted for this target.
+        let mut running_votes: HashMap<H256, HashSet<u64>> = HashMap::new();
+        running_votes.insert(target_root, HashSet::from([0, 1, 2]));
+
+        let out = keep_best_proof_per_data(entries, &running_votes, att_data.slot);
+
+        assert_eq!(
+            out.len(),
+            1,
+            "a block carries one entry per AttestationData"
+        );
+        let kept = &out[0].0;
+        assert_eq!(
+            kept.aggregation_bits.count_ones(),
+            2,
+            "marginal coverage, not absolute participant count, drives the choice"
+        );
+        assert!(
+            kept.aggregation_bits.get(3).unwrap_or(false)
+                && kept.aggregation_bits.get(7).unwrap_or(false),
+            "the small subnet {{3, 7}} adds the only new coverage and must be kept over the \
+             larger {{0, 1, 2}} subnet already counted in-state"
+        );
+    }
+
+    /// With no in-state votes for the target (a fresh target), marginal
+    /// coverage equals absolute count, so `keep_best_proof_per_data` keeps the
+    /// larger proof, preserving the pre-existing behavior.
+    #[test]
+    fn keep_best_proof_per_data_keeps_largest_when_target_is_fresh() {
+        let att_data = AttestationData {
+            slot: 5,
+            head: Checkpoint::default(),
+            target: Checkpoint {
+                slot: 5,
+                root: H256([9u8; 32]),
+            },
+            source: Checkpoint::default(),
+        };
+        let small = (
+            AggregatedAttestation {
+                aggregation_bits: make_bits(&[0]),
+                data: att_data.clone(),
+            },
+            SingleMessageAggregate::empty(make_bits(&[0])),
+        );
+        let large = (
+            AggregatedAttestation {
+                aggregation_bits: make_bits(&[1, 2]),
+                data: att_data.clone(),
+            },
+            SingleMessageAggregate::empty(make_bits(&[1, 2])),
+        );
+        let entries = vec![small, large];
+
+        let out = keep_best_proof_per_data(entries, &HashMap::new(), att_data.slot);
+
+        assert_eq!(out.len(), 1);
+        assert_eq!(
+            out[0].0.aggregation_bits.count_ones(),
+            2,
+            "with an empty in-state voter set the larger proof wins on absolute count"
+        );
+    }
+
+    /// Two distinct `AttestationData` entries can share one `target.root`
+    /// (differing only in `slot`, `head`, or `source`). The STF unions all
+    /// their voters onto that target, so the collapse must score each group's
+    /// candidates against voters already claimed by earlier groups in this
+    /// block, not just the frozen in-state snapshot. Otherwise both groups keep
+    /// the same subnet and the block under-covers the target.
+    #[test]
+    fn keep_best_proof_per_data_accounts_for_voters_claimed_by_earlier_groups() {
+        let target_root = H256([9u8; 32]);
+        let target = Checkpoint {
+            slot: 5,
+            root: target_root,
+        };
+        // Same target root, different attestation slots -> distinct data roots.
+        let data_a = AttestationData {
+            slot: 5,
+            head: Checkpoint::default(),
+            target,
+            source: Checkpoint::default(),
+        };
+        let data_b = AttestationData {
+            slot: 6,
+            head: Checkpoint::default(),
+            target,
+            source: Checkpoint::default(),
+        };
+
+        // A has a single proof {3, 7}. B offers {3, 7} (already claimed by A)
+        // and {8, 9} (the only new coverage). B's {3, 7} is listed first, so
+        // absent claim-tracking the first-occurrence tiebreak would keep it.
+        let entry = |bits: &[usize], data: &AttestationData| {
+            (
+                AggregatedAttestation {
+                    aggregation_bits: make_bits(bits),
+                    data: data.clone(),
+                },
+                SingleMessageAggregate::empty(make_bits(bits)),
+            )
+        };
+        let entries = vec![
+            entry(&[3, 7], &data_a),
+            entry(&[3, 7], &data_b),
+            entry(&[8, 9], &data_b),
+        ];
+
+        let out = keep_best_proof_per_data(entries, &HashMap::new(), data_a.slot);
+
+        assert_eq!(out.len(), 2, "one entry per distinct AttestationData");
+        assert!(
+            out[0].0.aggregation_bits.get(3).unwrap_or(false)
+                && out[0].0.aggregation_bits.get(7).unwrap_or(false),
+            "group A keeps its only proof {{3, 7}}"
+        );
+        assert!(
+            out[1].0.aggregation_bits.get(8).unwrap_or(false)
+                && out[1].0.aggregation_bits.get(9).unwrap_or(false),
+            "group B must keep {{8, 9}}: {{3, 7}} was already claimed by group A for \
+             the same target root, so it adds no new coverage"
+        );
+    }
+
     /// Regression test for leanSpec PR #716: build_block must absorb
     /// gap-closing attestations whose source is justified on the head
     /// chain but older than `latest_justified` (e.g., a sibling fork