TLDR: This is a mashup of the honest-majority availability post and @vbuterin’s erasure coding note, with the aim to achieve scalable fork-free sharding.
Construction
Suppose there are n validators and that we have \textrm{SHARD_COUNT} shards. For every collation body B to be pushed to the VMC, the proposer prepares an erasure coding with n pieces B_1, ..., B_n, one for each validator. The erasure coding is such that any \textrm{SHARD_COUNT} pieces are sufficient to reconstruct B. The proposer also needs to prove that the Merkle root r of the pieces B_i corresponds to a faithful erasure coding of B (the hard part!).
We now require a BLS threshold signature of r from (n + \textrm{SHARD_COUNT})/2 validators before B can pushed to the VMC. This can be guaranteed if the honest majority assumption is strengthened from n/2 to (n + \textrm{SHARD_COUNT})/2.
If collation sizes are capped at 1MB and \textrm{SHARD_COUNT} = 100 then each honest validator only has to download 10kB (the size of a piece) per collation per shard. That is, we can have honest-majority availability votes across all shards for just the bandwidth cost of a single shard, thereby achieving scalability.