Every chain tells you a transaction is confirmed, but the strength of that statement varies enormously. Finality is the point at which reversal stops being realistically possible.
Proof of work finality is probabilistic
In a chain secured by mining, a transaction is included in a block and then buried under later blocks. Nothing marks it as irreversible.
Reversing it would require rebuilding the chain from before that block with more accumulated work than the network produced in the meantime.
Each additional block makes that harder, so confidence approaches certainty without ever formally arriving. Confirmations are a measure of cost to reverse, not a status flag.
Stake-based designs can finalise explicitly
Systems built on staked capital can have validators vote on checkpoints, and once a supermajority has voted, the protocol treats the block as final.
Reversing it would require those validators to contradict their own votes, which the protocol detects and punishes by destroying part of their stake.
Finality here is economic rather than statistical: it has a defined moment and a defined price.
Confirmation counts are business decisions
An exchange choosing how long to wait is balancing the cost of a reversal against the cost of making customers wait.
Small deposits clear faster than large ones on the same chain, because the amount at risk is what determines how much security is worth buying.
The same reasoning explains why a merchant accepting a modest payment may treat a single confirmation as sufficient while a custodian moving a treasury waits considerably longer.
Nothing about the chain changes between those two decisions. Only the size of the loss that a reversal would cause has changed.
Layered systems inherit the weakest guarantee
A rollup can order transactions instantly, but its settlement depends on data or proofs reaching a base chain, so its real finality is the base chain's.
Bridges face the same problem in a sharper form. Releasing funds on one chain before the source is final is exactly how reorganisation attacks against bridges have worked.
A user sees a near-instant confirmation on the fast side and assumes the whole path is settled, when the guarantee is still pending underneath.
Why the published number is not the whole answer
Two chains quoting the same confirmation count can offer very different assurances, because the cost of producing competing blocks differs by orders of magnitude between networks.
Security also depends on how concentrated block production is. A network where a small group produces most blocks can be reorganised without any of the effort the raw numbers suggest.