In this article
Small steps are the method
A typical change touches 6 files and 178 lines. A quarter of changes touch 38 lines or fewer.
Of the changes to product code, 82% carry a test or a check in the same commit. 416 commit messages name their proof.
Long plans taught us this. One early plan listed 764 conditions to meet, and 188 never got a test. Plans go stale; small tested changes do not.
A test fails first, then the fix
On 18 October a progress line said 26 of 59 searches had run. In fact 55 had. The code counted a search only when it cost money.
The fix began with a test that held two free, empty searches. It failed at 0, then passed at 2.
On 15 October the first live research agent worked for 7 minutes, and its save was refused. Its scores pointed at evidence a later step had already replaced. So the save matched nothing. A database test now fails without that fix.
Pretend tests pass. Real data bites.
On 24 September every CRM sync part passed its tests against a pretend CRM. The first sync into a real local copy broke in seven places. Each is named in the commit.
One of those seven would have read every lead again every hour, quietly, forever.
On 28 September a 15-word cap on message lines was tested on 15 real leads. It refused 7 of them, so the cap became 18.
Rules are locked, and locks are tested
The rules for judging leads are numbered and locked. Blind checks drive each change, and a fixed, blind-labeled set guards the starting point. The lead search post has the check table and the 820-row lock.
Mach's writing rules list 196 kinds of text it may write, each with limits code checks. 154 check scripts guard the screens.
Scripts that change stored records do a test run by default. Records are backed up before they are cleared or merged.
What broke, and the rule it left
On 18 October the app was down for 15 minutes. An old build folder slipped an unfinished database change into a release. The next release clashed with it.
Since that day, the release folder is wiped before every release. The outage is written in the journal.
By 20 October there were 4,640 tests in the Java product and 516 in the agent service.
The checklist we would hand a friend
Write the failing test before the fix, and name the test in the commit.
Block a release on checks a person did not run by hand. Treat counters as code: an untested progress number lies sooner than anything else.
Make scripts do a test run by default, back up before merges and keep the journal honest, outages included.
Common questions
What does 82% with tests mean for the other 18%?
Most are docs, images and settings. A product-code change with no test in the same commit is rare. Every existing check still runs against it.
Where the numbers come from
Every count, date and story comes from Mach's commit history and journal, exported 20 October 2026. That history holds 1,225 commits since 23 July.
