Ideas · Builder Notes

Kognita 0.3: ticking the definition of done only when the code agrees

After midnight on 5 October I finished the Kognita alpha 0.3 checklist: classifier replay, the injection relabel test, the conformance kit, a gateway latency gate and the Tier 0 rollup.

Before

The 0.3 definition of done still had five open lines, and the gateway had no latency number. Some lines described work that already existed, some needed a test, and I didn't want to tick any of them from memory.

After

Each line got its own pull request and its own test. The classifier replay definition was written down and checked: replay uses the recorded label and never calls the classifier again. The injection relabel test and the conformance kit were ticked once their tests held.

I added a benchmark that builds the gateway the way kognita serve does and fails if overhead is 50 ms or more. Over 30 sequential calls, classifier time excluded, the maximum was about 37 ms and the mean about 9.5 ms. The machine and the sample went into the changelog with the numbers.

The rollup line 'All Tier 0 defects closed' was ticked last, after every box under it was re-checked against the code.

Why it matters

Writing down 'the note was wrong' was more useful than quietly tidying the roadmap. One drift stayed open on purpose: a typed classification could still replace the classifier's label, including a lower one. I fixed that in 0.5.

The latency gate on a maximum turned out to be fragile. A single slow CI run could fail it, so in 0.4 it moved to a 95th percentile and a mean over 100 calls. I'd start with a percentile next time.

Sources

Kognita is an open-source alpha. This note comes from Kognita PR #23, Kognita PR #24, Kognita PR #25, Kognita PR #26, Kognita PR #27 in the Kognita repository.

All ideasEmail Ahmed