Method. A 200-line file; both sides apply one to three random edits — modify, insert, delete — either anywhere in the file or confined to the same small window, which is what two people working in one function looks like. Both engines see byte-identical input. Measured 2026-09-11 against git merge-file (git 2.45.1); default and --zdiff3 agreed on every pair.
Why the gap. The two engines are built on different substrates and end up asking different questions of the same three files. That mechanism is worth seeing in full, so it has its own section below.
The number that keeps this honest: a merge engine can drive its conflict rate to zero by silently dropping an edit, so every clean merge above is audited, per trial, by the checker below. Where both engines merged cleanly the output was byte-identical — all 1,957 of those pairs; there was no pair Git merged and TOVIO refused; and on every clean merge, no inserted line went missing and no removed line came back.
That second result is what lets the same figure be stated over one population rather than two: because no pair exists that Git merged and TOVIO refused, TOVIO's conflicts are a subset of Git's. So of the 843 pairs Git stopped on, TOVIO settles 647 — 77% — with every line of both sides accounted for, and stops on the rest. The advantage narrows as the collision tightens: in the last row, both developers inside the same eight lines, it is 3.3× and TOVIO still conflicts on 23.3% of pairs. Two people rewriting the same line is a question for a human, and no version control system can merge it away.