Imagine you're driving a car with two gearboxes: a fast manual for clear highways and a slow-but-unstoppable automatic for fog. Most drivers pick one and stick with it. The interesting engineering question is: can you switch between them seamlessly, while the car is moving, without stalling? Steelhead says yes — and does it without even needing a separate signal to agree on which gearbox to use. The committed claim: you can interleave a partially synchronous commit rule (fast when the network behaves) and an asynchronous commit rule (live no matter what) on a single shared DAG, switching between them adaptively with zero extra messages. The protocol designates every k-th round for asynchronous decision via a common coin that hides the leader until after votes are cast, and all other rounds use the faster partially synchronous path. Validators periodically replay the committed DAG under each candidate period k, pick the one with the fewest expected message delays, and fall back to k=1 (pure asynchronous) when output stalls. This lands squarely in the DAG-based BFT consensus lineage that includes Mysticeti (partially synchronous, low latency) and Mahi-Mahi (asynchronous, censorship resistant). The core tension in DAG consensus is the tradeoff between speed and liveness guarantees: partially synchronous protocols are fast when GST holds but halt under true asynchrony, while asynchronous protocols survive anything but pay a latency tax in good conditions. Steelhead's architecture is generic over pairs of DAG commit rules sharing a committee — instantiated here with Mysticeti/Mahi-Mahi at n ≥ 3f+1 and two BlueBottle variants at n ≥ 5f+1. The integrity story is notably strong for a systems paper. The authors provide mechanized safety and liveness proofs in Lean 4 — not just pen-and-paper arguments. The protocol is evaluated in simulation, where it matches the partially synchronous protocol's throughput in healthy networks, tracks close to the asynchronous protocol when the partially synchronous one stalls, and adapts quickly in both directions. The gap is the absence of a real-world deployment or even a prototype implementation under adversarial network conditions. The design's key elegance is its zero-overhead property: Steelhead sends no messages beyond the DAG's existing blocks, not even to coordinate the period k. The common coin is opened only on rounds that need a hidden leader. This means you can bolt it onto existing DAG infrastructure without rearchitecting the messaging layer — a significant practical advantage for blockchain teams already running Mysticeti or similar. The next milestone to watch is whether Steelhead can demonstrate its adaptive switching under real adversarial network partitions, not just simulated delays. The simulation results are encouraging but the hard question is whether the period-replay mechanism converges fast enough under Byzantine behavior combined with genuine network instability. The obvious experiment not run here: a full implementation benchmark against Mysticeti and Mahi-Mahi individually under real network topologies with actual Byzantine nodes.
