RaptorCast: getting a block to two hundred validators before the slot ends

A 300 millisecond slot leaves very little time to move a block the size of Monad’s gas limit to every validator that must vote. Gossip that fans a block out one full copy at a time will not do it. RaptorCast is the propagation protocol Monad uses instead: erasure-coded chunks, so a validator can reconstruct the block from a subset of pieces, and a sending pattern that does not wait for a single neighbour to finish a full copy.

The vote that does not wait

MIP-10, discussed through 2026 and not the default behaviour of the network as this desk writes, would let validators cast a vote before the entire block has been reassembled locally. The idea is to take propagation off the critical path of the vote. It is the sort of change that sounds small on a slide and is large on a wide-area network, where the last 10 percent of a block is often the slowest 10 percent. Until it is in the client you are peering with, the live rule is the older one: the block arrives, then the vote.

This is also why “200 validators in 30 countries” is an engineering statement and not only a governance one. Every extra region adds delay. Linear consensus complexity does not help if the block itself cannot cross the ocean inside the slot. RaptorCast is the part of the stack that has to make the geography fit.

Finality you cannot propagate is a number on a benchmark host.

What a trader notices

Almost nothing, until a slot is missed. A missed proposal is a 300 millisecond hole, not a ten-minute reorg. CertiK’s August snapshot put the successful proposal rate at 99.903 percent. Holes still happen. Systems that assume every slot produces a block, and that price their orders on that assumption, will occasionally find the book one slot older than they think. Build for the hole. Do not build for the average.

Any three pieces, if four were sent

Take a block and cut it into pieces. Code the pieces so a receiver who missed one can still rebuild the file. That is erasure coding, and it is the move RaptorCast makes instead of hoping a full copy arrives intact. Codes in the Raptor family are built for a sender who does not want a conversation about which piece failed. The sender emits coded chunks. The receiver stops when it has enough. Nobody spends a round trip asking for the one that got lost, because on a 300 millisecond clock a round trip is a large part of the slot.

The ocean is inside the budget

Light takes on the order of 40 milliseconds to cross the Atlantic, before a router adds anything. A slot of 300 milliseconds can pay for a few of those crossings and very few acknowledgements. Gossip that hands the entire block to one neighbour, waits until that neighbour has it, and only then continues, spends the slot on politeness. Spraying coded chunks to many validators at once spends it on coverage. A set of operators in more than 30 countries, the scale 2026 accounts of the validator set keep citing, is a harder physics problem than a dozen machines on one rack. It is also the scale at which “decentralised” stops being a caption and becomes a delay you can measure.