Stationify

Research

AI that improves itself, in steps we can measure.

Stationify Research is building an AI system that makes complete, original software end to end, and then improves itself. Each round it studies what it made, finds where it falls short and writes its next version. A new version only counts as better when it beats the last one in the same fixed test, on the same inputs.

So far, as of 20 October 2026

complete builds
1,084
versions of the system
16
quality score out of 10, versions 1 to 10
4.1 → 5.6
quality score out of 10, versions 11 to 15, on a newer test
5.8 → 6.0
experiments, 44 of them finished
48
automated test sessions
205,910
measurement runs
83,619
AI evaluations
13,819

Quality score, version by version

Each point is one version of the system. Versions 1 to 10, built between 10 and 12 October 2026, were scored the same way on the same inputs. From version 11, each version is scored on a newer, fixed test with its own scale, so the second line doesn't continue the first. The dotted stretch passes version 13, which has no score on the newer test yet.