kapso Files

The new project page

This release ships no code. src/, MANIFEST.in and the dependency set are byte-identical to 0.3.8 — upgrading changes nothing about how Kapso runs. The entire diff is README.md and docs/images/.

It gets a version because README.md is the PyPI long_description, and the project page is the only place this work reaches anyone arriving through pip.

An animated hero

The static architecture figure is replaced by the loop actually running: a tree of experiments expanding under ideate → implement → evaluate → feedback, the performance curve climbing beside it, and trajectories plus public knowledge flowing into the learning engine as the knowledge bank fills.

One frame for every benchmark

RelBench, MLE-Bench, ALE-Bench and the Claude Code baseline were drawn in three different styles across four READMEs. They are now one figure family — shared frame, shared type, shared colour encoding — and every bar carries the system's name and its organisation's mark.

Two encoding fixes came out of the redraw:

• a baseline's delta rendered in our accent colour, so Claude Code's −22% on recommendations — a loss — read as a win
• the AUROC groups now declare their axis floor in the metric line ("scale from 78") rather than truncating silently

IOAI, set up properly

The IOAI 2026 item was one dense sentence carrying four claims and none of the context that makes the result hard. It now introduces the olympiad, then the AI track, then the rule — once the clock starts, no human may solve, correct, or improve anything — before either result lands.

The human claim is also stronger and more checkable: outscored all 471 contestants, 536.07 total, which is what the benchmark README states.

Full Changelog: https://github.com/Leeroo-AI/kapso/compare/v0.3.8...v0.3.9