DNA replication is highly accurate because several layers of molecular discrimination reduce copying errors before they become permanent sequence changes.

The first layer is nucleotide selectivity. A replicative DNA polymerase favors an incoming nucleotide whose base forms the correct complementary pair with the template. Correct pairing also helps place the reactive groups in the geometry required for phosphodiester-bond formation.

Even so, an incorrect nucleotide is occasionally incorporated. Many replicative DNA polymerases then use a second activity called proofreading.

Proofreading removes a newly mispaired nucleotide

Suppose the template contains G but the polymerase accidentally adds T to the new strand. The resulting G-T pair has an abnormal geometry. The polymerase can pause and transfer the newly synthesized $3'$ end to a proofreading site with $3'\rightarrow5'$ exonuclease activity.

An exonuclease removes nucleotides from the end of a nucleic-acid strand. The proofreading site excises the incorrect terminal nucleotide, restoring a correctly paired $3'$ end. The polymerase can then resume synthesis in the normal $5'\rightarrow3'$ direction.

Schematically:

wrong nucleotide incorporated
        ↓
mispaired 3′ end recognized
        ↓
3′→5′ exonuclease removes terminal nucleotide
        ↓
5′→3′ synthesis resumes

Proofreading therefore does not mean that DNA synthesis reverses direction. Polymerization remains $5'\rightarrow3'$; a separate exonuclease reaction temporarily removes a nucleotide from the new strand's $3'$ end.

Fidelity is multiplicative

Replication accuracy reflects several successive filters:

1. correct nucleotide selection before bond formation;
2. proofreading immediately after many misincorporations;
3. post-replicative repair systems that can correct some mismatches that escaped proofreading.

Because each layer removes a large fraction of the errors remaining from the previous layer, the final error frequency can be much lower than the selectivity of nucleotide incorporation alone.

Proofreading is therefore a local quality-control mechanism operating at the replication fork. It corrects many copying mistakes while the newly synthesized strand still identifies where the error occurred.