Under proof of work, a block is never final. It becomes progressively harder to reverse as blocks accumulate on top of it, because reversing means redoing that work faster than the network is producing it. Confidence is probabilistic and it grows with depth, which is why exchanges wait for a number of confirmations rather than for an event.
Proof of stake can offer something stronger. Because validators are known — they have posted collateral to identifiable accounts — the protocol can require them to vote on blocks and can punish contradictory votes. Ethereum's design groups slots into epochs and asks validators to attest as each epoch passes. When a supermajority of staked ether attests to a checkpoint, it is justified; when a second checkpoint is justified on top of it, the first is finalised. Finalisation is not a probability. Reverting a finalised checkpoint requires that at least a third of all staked ether has voted for contradictory checkpoints, and the protocol can identify and destroy exactly that stake.
This is what distinguishes economic finality from probabilistic settlement. The guarantee is not that reversal is computationally hard; it is that reversal is attributable and expensive in a way that cannot be recovered. An attacker who reverts a finalised block loses a third of the network's entire stake in the process. Under proof of work an attacker with enough hardware can reorganise the chain and keep the hardware.
Three assumptions carry that argument, and they should be stated. Slashing must be enforceable, which means the honest majority must be able to include the evidence in a block — an adversary controlling block production for long enough could delay it. Validators must be sufficiently independent: if a large share of stake runs the same client, the same cloud provider or the same staking operator, a correlated failure can look like a coordinated attack, and the protocol's correlation penalties are deliberately harsh precisely because it cannot tell the difference. And there is a bootstrapping question — a node syncing from genesis with no recent view of the chain cannot distinguish the real history from a fabricated one written by validators who have since withdrawn, which is why proof-of-stake clients rely on weak subjectivity checkpoints rather than on the chain alone.
The inactivity leak handles the case where finality stalls because too many validators are offline. Rather than halting, the protocol drains the stake of non-participating validators until the remaining active set is a supermajority again and finality resumes. It is a mechanism for surviving a partition at the cost of the absent, and it means a network split does not simply stop — it prunes.
For anyone reading settlement times rather than protocol specifications, the practical figures follow from the structure: attestation is continuous, finalisation happens on epoch boundaries, and the worst case is a small number of epochs rather than an unbounded wait. It is why proof-of-stake chains can offer a firm settlement guarantee to an institution in minutes, and why the design question that matters is not the yield but who holds the stake.