As of 2026-09-14. Engine: MarkEngineV3, Mardo's own, by Floof Logic. Full, auditable source: ENGINEERING.md on GitHub · What produces these: /engineering/

Benchmarks

MarkEngineV3's numbers, each with the gate, the machine, and the date behind it. What produces them is in Engineering; the full, auditable source of both parts is ENGINEERING.md on GitHub.

How these are measured

Latency and memory come from Floof Logic's performance gate, which measures MarkEngineV3 with the app launched against a hard per-frame maximum of 16.67 ms where one applies; no percentile excuses a missed maximum. Large-file figures come from the large-file qualification gate on deterministic generated corpora with a matched native control. Mardo's source is private, so these gates are not runnable by a reader; the value of this page is that every number names the gate that made it, the machine, and the date, and its history here is a public diff.

Speed

First paint is three to five milliseconds from one kilobyte to sixteen megabytes, the same work at every size. Typing on a megabyte of prose commits a glyph in a median of 2.77 ms, a fortieth of the budget. The cold-launch worst sample varies with host load and is quoted as a median.

Memory

Settled memory of the whole process. The controls are a matched native text view measured the same way.

An eight-megabyte manual settles around 43 MiB; a four-gigabyte log settles 2.1 MiB above a native text view; a fifty-gigabyte file settles at 44.6 MiB, 1.8 MiB above a control that is sixteen hundred times smaller. Memory tracks what you are viewing, not the size of the file.

Large files

"Publish" is the time to put the bytes under the viewport on screen; the exact render follows immediately after.

Drop the scrollbar at byte four billion and the text is on screen in under ten milliseconds, because opening and jumping never do work that grows with the file.

Correctness and proof

Each row is one of Floof Logic's release gates; the count is the last green run's, dated.

Floof Logic keeps 219 dated evidence records, each with the exact result behind a claim.

Head to head on the numbers both sides publish

Downright is the only competitor that publishes performance figures, so it is the only fair numeric comparison. Its figures are from its own benchmark on a 5,000-line, 120 KB file, read from PERFORMANCE.md on 2026-09-11. Mardo's are the gate rows above.

On every row both projects measure, Mardo's number is smaller, its input is larger, and its gate runs against a hard maximum rather than a manual target. Downright earns the comparison by publishing at all; the rest of the field publishes no numbers to compare.