Settlement is not finality
Settlement is the discharge of an obligation. I owed you money; now I do not. In a
traditional system this is the moment a central bank’s books are updated.
Finality is the point after which that discharge cannot be undone. It is a legal concept
before it is a technical one, and the two do not always agree.
A payments business does not “wait for finality”. It chooses a confirmation depth at which the
residual reorg probability is acceptable relative to the value at risk, and that choice is a
business decision priced into the product, not a technical constant.
Ethereum’s move to proof-of-stake introduced economic finality via checkpointing, which is stronger: reverting a finalised checkpoint requires an attacker to burn a large fraction of staked ETH. It is still not a statute. It is a very expensive disincentive, which is a different kind of guarantee than TARGET2 offers, and worth being precise about when someone claims the two are equivalent.
Three different things called finality
The word covers at least three distinct guarantees, and a corridor operator has to know which one it is relying on.
Solana is the useful case for seeing the difference, because it exposes finality as progressive commitment levels rather than a single block count. A transaction is processed, then confirmed once a supermajority of validators has voted on its block, then finalized once enough confirmed blocks are rooted above it. Roughly 32 slots at 400ms reaches finalized, which is where the ~13 second figure in the table below comes from.
The
Arc waits for the finalized equivalent on every chain. That is the slower and more conservative option, chosen so that “settled” means one thing across five chains with five different consensus mechanisms.
confirmed level arrives in about a second and is what most consumer applications use. That choice is not a technical detail. It is a decision to credit a beneficiary against a state that is very unlikely, but not certain, to persist, and the gap between the two levels is a credit exposure whether or not anyone books it as one.The counter-intuitive part: block time is not speed
Here is the table that makes the point, and it is the reason Arc’s chain layer exists at all.
Polygon has blocks six times faster than Ethereum and settles nearly twice as slowly. Fast blocks that are individually less secure require more of them to reach the same confidence. Optimising for block time alone gets this exactly backwards.
Arc models each of these five chains with its own block time, finality depth, fee model, reorg depth distribution, and failure rates, and then selects a chain per transfer by scoring settlement time against fee, weighted by a speed preference. That is what “chain-agnostic” means concretely: the corridor picks per transfer rather than being wired to one chain.The full story is The fastest chain that settled last.
Failed is not dropped
A distinction that looks pedantic until you have to compensate for it.FailedThe transaction was mined into a block, executed, and reverted. It consumed gas. The fee is spent and gone.
DroppedThe transaction never made it into a block: evicted from the mempool, underpriced, or replaced. It cost nothing.
These require different compensation. Reversing a failed transaction must not reverse the gas
expense, because the gas was really spent: pretending otherwise misstates the expense and
produces a balanced ledger that lies about reality. Arc’s saga deliberately does not track or
reverse the network-fee journal for exactly this reason.
Arc’s chain simulator models both as distinct terminal states (failed and dropped), alongside pending → included → confirmed, precisely so the saga has to handle the difference rather than collapsing them into “error”.
Reorgs are not hypothetical
A reorganisation happens when the chain discards blocks it had previously built on and adopts a competing branch. Transactions in the discarded blocks return to the mempool: they are un-mined, and a transaction you had counted as included is suddenly pending again. For a payments business this is the nastiest failure mode available, because it is the one where you already told the customer it worked. Arc’s simulator produces reorgs two ways, and the second matters more:1
Probabilistically, during block production
Each chain has its own reorg rate and maximum depth, drawn from a seeded PRNG. A seeded run
reproduces the same reorg sequence every time, which is what makes the whole determinism claim
hold.
2
Deterministically, via forceReorg(depth)
The chaos suite needs a reorg at a precise moment. Tuning a probability until a test happens
to reorg is flaky and proves nothing, so an explicit trigger exists alongside the probabilistic
one.
reorg event carries revertedTransactions, so a consumer knows exactly what to undo without diffing state itself. Rollback clears each affected transaction’s block height, resets its status to pending, and pushes it to the front of the mempool to be re-mined promptly.
What this means for the customer promise
A corridor product has to make a promise about time, and there are only three honest shapes for it.
The third is what most consumer-facing “instant” products actually are, and it deserves to be named. Crediting a beneficiary before finality means the provider is lending against an unsettled position. That can be perfectly sound, priced, bounded, reserved against, but it is a lending decision, and a system that does not model it as one will eventually be surprised by it.
Arc takes the second shape. The
settle step broadcasts, advances the chain, then re-reads the
transaction and requires final before proceeding. Checking final rather than “no error was
thrown” is the difference between confirming settlement and merely confirming submission, and it
is a distinction that is easy to get wrong in code that looks correct.Next: the 2026 regulatory picture
GENIUS Act rulemaking, MiCA’s transitional cliff, Kenya’s VASP Act, and Nigeria’s ISA 2025.
Chain characteristics above are Arc’s simulation parameters, chosen to reflect published finality guidance for each network at the time of writing. See the chain layer for the source of truth and the bibliography for references.