Every measured figure on the site is an instrument reading, re-derived by one script, dated, and set beside a published benchmark. This page says what was counted, how, and against what. Nothing on it is typed by hand.
On 2026-09-03, one script read every change the platform has delivered from GitHub, across 2 production instances, from 2026-04-22 to 2026-09-01. That is 1,507 changes delivered and 58 opened and not delivered. One person approved every one of them.
Finished what it started is the delivered count over everything opened, delivered or not: 96.3%, stated as 96%. Delivered per week is the delivered count divided by the weeks between the first and the last: 79.5 a week, stated as 79. Right the first time is the share of delivered changes that needed no correction after they were opened, allowing a 10-minute window for the author to finish: 88.4%, stated as 88%. Without the window it is 81%. First line of code to live is the median from the first line written to the change going live, and going live is the same event as approval on these instances: 23 minutes. Opened to delivered is the median from the pull request opening to that same event: 16 minutes. Had to be undone counts changes reverted after going live: 2 of 1,507, 0.1%.
Every multiple is the ratio of two cited numbers and nothing else, and every multiple is floored, never rounded up. The site states no more than the instrument read.
LinearB, 2026 Software Engineering Benchmarks. 8.1 million pull requests from 4,800 teams, January to December 2025. Its elite tier merges more than 2 changes per developer per week and moves a change from first line of code to live in under 25 hours; the average team takes about 7 days. Across its whole sample, 84.4% of human-written changes were accepted and 32.7% of AI-generated ones. The elite figures are the comparators on every page, chosen because they are the boundary of the top tier, not the average.
Code review effort in 23,600 GitLab merge requests, 2025. arXiv 2507.23640. Up to 71% of merge requests needed changes after submission, so about 29% were right the first time.
Merge-request deviations, 2025. arXiv 2506.08860. Revert rates in the studied projects ran up to 7.6%.
The $1,000 cost per change.A labeled assumption, not a measurement: a US mid-level engineer, fully loaded, at about three delivered changes a week. The home page lets you replace it with your own figure, and every saving shown is arithmetic on the number you set. The $1.10 beside it is measured from the platform's own spend ledger, the same ledger its spend limits enforce against.
These numbers were produced by one person approving the work of one governed organization on 2 codebases. They are not a promise about your team, your backlog, or your codebase. The worked examples on the home page are arithmetic on the measured rate, and say so. Nothing on the site is a headcount claim. The sixty-day pilot exists to replace every figure here with your own.