What finality actually promises

Probabilistic versus economic finality

Under longest-chain consensus, a block becomes progressively harder to displace as work accumulates on top of it. Reversal never becomes impossible; it becomes expensive, and the depth at which you stop worrying is a risk decision you are making, not a property the protocol hands you.

Bonded systems make the cost explicit. Reverting a finalised block requires a large fraction of bonded stake to contradict itself, and the protocol destroys that stake when it does. The promise is not "cannot happen" but "cannot happen without a specified, attributable, very large loss".

What the guarantee is conditional on

Every finality claim carries conditions, and the conditions are where the interesting failures live:

• An honest-majority or bonded-supermajority assumption that must actually hold.
• Network synchrony assumptions that may be violated during an outage or partition.
• Client correctness — a consensus bug can finalise something the rules do not permit.
• Social layer behaviour, which decides what happens after any of the above.

Why this matters operationally

The question "how many confirmations before we credit a customer" is a risk-appetite question dressed as a technical one. The right depth depends on the size of the transfer, the cost of reversal on that network, and what the business loses if it is wrong.

Written that way it becomes reviewable: a policy with a threshold, a rationale and a review date. Written as "we wait for finality", it is a phrase standing in for a decision nobody has made.

The teaching point

Students arrive expecting consensus to deliver certainty, and the useful shift is to see it as delivering a priced guarantee instead. Once finality is understood as a cost, comparing two networks becomes a comparison of costs and conditions rather than a contest of adjectives.

The consensus units of the fundamentals course end here deliberately: not with which mechanism is better, but with what each one is asking you to assume.