Fraud Proof
Evidence submitted on the base chain showing that a layer 2 posted an incorrect result, causing that result to be rejected.
Modern implementations are interactive: the challenger and the proposer bisect the disputed computation step by step until they disagree about a single instruction, which the base chain can then execute itself cheaply and rule on. Both sides post bonds and the loser forfeits, which pays for the process and discourages frivolous challenges. The mechanism only works if the underlying data was published, since a challenger has to recompute the correct result, and if at least one honest party is watching, funded, and permitted to submit within the window. Where fault proofs are not yet live, or where only an approved list of challengers may act, users are relying on the operator's honesty rather than on the proof system.
In practice
In a bisection game the two parties narrow a dispute over an entire batch down to one instruction, so the base chain only has to execute a single step to settle it.
The common misunderstanding
Fraud proofs do not run by themselves; a person or service must notice the error and pay to submit the challenge before the window closes.