Adaptive attestation consensus.
Linear when fast. DAG when it matters.
Leader-based BFT protocols are fast when the leader is fast. But when a leader is slow, offline, or geographically distant, the entire network stalls waiting for a timeout before failover kicks in. Traditional solutions introduce complexity that creates more problems than it solves.
When a leader misses its slot, validators wait for a full timeout period before electing a replacement. During that window, zero blocks are produced. At scale, these gaps compound into measurable throughput loss.
Fork-choice rules and block rollback attempt to recover from leader failures, but they introduce state reversions, orphaned transactions, and edge cases that are difficult to reason about under adversarial conditions.
In standard BFT, only the leader produces blocks. Every other validator is silent until it is their turn or a vote is requested. Their compute capacity, network bandwidth, and liveness signal go unused.
Multi-tier failover cascades, jailing heuristics, and reputation-weighted re-election create hundreds of lines of consensus code. Each edge case is a potential liveness or safety bug that is difficult to audit.
Vexcel introduces a lightweight attestation layer between consensus rounds. When a leader is slow, non-leader validators produce attestation blocks — empty, zero-gas structures that prove the validator is alive, synchronized, and ready. These attestations are stored separately and referenced by the next leader, forming a DAG only when latency demands it.
Vexidus is an L1 blockchain running HyperSync BFT across 5 geographically distributed validators. Before Vexcel, leader failures triggered a cascade of rollbacks, fork-choice rules, and staggered failover. After Vexcel, those mechanisms were replaced entirely.
Fork-choice rules that selected the heaviest chain. Staggered failover with multi-tier timeout cascades. Block rollback that could revert finalized state. Jailing heuristics that punished validators for latency they could not control. Each edge case was a potential consensus bug.
A validator in Mumbai with 180ms round-trip to Europe would consistently miss tight timeouts. The failover mechanism would kick in, jail the validator, and redistribute its slots — only for the replacement to face similar latency issues. Production distribution became uneven and unpredictable.
330 lines of fork-choice, rollback, and staggered failover were removed. 420 lines of attestation logic were added. When a leader is slow, non-leaders produce attestation blocks stored in a separate data layer. The next leader references those attestations, forming a DAG. When the leader is fast, the chain is linear. No rollbacks. No forks. No wasted slots.
Vexcel is not specific to Vexidus. Any blockchain using leader-based BFT consensus can adopt the attestation pattern to eliminate rollback complexity and improve liveness under latency.
Any chain with rotating leaders and deterministic finality. Tendermint, HotStuff, and PBFT variants can integrate attestation blocks as a liveness layer without modifying their core safety proofs.
Shared sequencer networks with multiple operators face the same leader-latency problem. Attestation blocks let non-sequencing nodes prove liveness without competing for the sequencing role.
Networks spanning multiple continents face asymmetric latency by design. Vexcel turns that latency from a consensus liability into a liveness signal without penalizing distant validators.
Enterprise BFT networks with fixed validator sets can use attestations to detect node health, trigger maintenance alerts, and maintain audit trails of validator liveness without impacting chain throughput.
Sub-second block times amplify the cost of missed slots. Attestation blocks fill gaps with near-zero overhead, maintaining effective throughput even when individual leaders experience momentary delays.
New validators joining a network can produce attestation blocks before being assigned leader slots, proving their node is synchronized and operational. A verifiable onboarding signal instead of a trust assumption.