The difficulty rule bounds a catch-up, not Cardano's twenty seconds

A second solution on the same parent spends an output that is already gone, so only one of them lands. That does not hold WHAT to one block per Cardano block. A transaction may spend an output created earlier in the same Cardano block, and Cardano's mempools accept a transaction whose input is still unconfirmed.

So a Stratum server tied to a Cardano block-producing node can see the last block's transaction in the mempool, mine on that unconfirmed tip and propagate the spend. Any Cardano block producer already holding the parent accepts it and can include both in one block. Cardano's twenty seconds is the base pace for a server that follows only confirmed blocks, not a floor under the chain.

whitepaper §5, §9

Two chains, linked by different things

Fig. 10 — Two chains, interleaved. Cardano's row is linked by stake and carries one slot per block; WHAT's headers are linked to each other by work, and each lands inside one of those blocks as a transaction. The breaks in the top row each stand for about thirty Cardano blocks, drawn unequal because the intervals are random.

whitepaper §6, §9

A surge is absorbed in about log₂ k blocks, however fast they come

Every early block tightens the target by close to a doubling, so an arrival of factor k is absorbed in about log₂(k) blocks, around ten for a thousandfold arrival. The same arithmetic counted in extra blocks is why there is no land-grab window.

How fast those blocks land depends on what the miners' servers follow. On servers that follow only confirmed blocks, at most one lands per Cardano block: the catch-up runs at Cardano's twenty-second base and sheds the rest, k/30 discoveries per Cardano block with one landing. That is the case Fig. 5 draws. Servers that follow the mempool are not held to it. The first blocks of a large catch-up can come seconds apart and share a Cardano block, and each one halves the rate the next is found at, so the run slows toward ten minutes by itself.

After it, intervals are back at ten minutes and orphaned submissionsA solution that arrived second for the same block and is simply invalid, the equivalent of a stale share. It names an output that has already been spent, so Cardano rejects it before any script runs: it pays no fee, and it costs its finder only the hashes behind it. near four per cent of discoveries, which is where simulation puts them for servers that follow confirmed blocks; following the mempool narrows each race to the time a transaction takes to propagate. There is no steady state in which surplus work substitutes for speed. A surge is reported as a higher difficulty behind ten-minute blocks, which is the surge reported as work.

The estimate of the work behind the chain is unaffected throughout. It reads the targets and attested times of the blocks that did land, and a block that shares a Cardano block with its parent still records its own attested time. See how the hashrate estimate is made.

Fig. 5 — Inside the catch-up window, for servers that follow only confirmed blocks, against the size of the arrival that started it. One curve flattens onto Cardano's ~20-second base while the other climbs away: surplus work turning into orphaned submissions instead of speed, until the target catches up. The knee near thirty times the target rate is where the base starts to bind. Servers that follow the mempool can run below it. Neither curve is a steady state — after catch-up the interval is ten minutes at every value on this axis. Anchored on the paper's two simulated points and labelled simulation on its face; it is not chain history, it does not mean the hashrate estimate degrades, and it is not a forecast that any amount of hashpower turns up.

whitepaper §5

A lost race costs you the hashes and nothing else

An orphaned submission names an output that has already been spent, so Cardano rejects it before any script runs: it is dropped from the mempool and pays no fee. Collateral is forfeited only when a script runs and fails, and none runs here. The loser is out the electricity behind the hashes and nothing more.

All of which rests on the validator accepting or rejecting each submission the same way every time, which is worth seeing directly.