About
A small lab, run locally.
Bastion is a small lab building AI tools that run on your own machine. Every change is tested against the last one before it ships.
What we build
Tools you run, and the research behind them.
How we test
Every change against the last one.
Nothing ships on a good feeling. A new version has to beat the one before it, on the same tasks, by more than the noise between runs.
- 01
Write the bar down first
Before anything runs, the log says what the change should improve and by how much.
- 02
Run old and new the same way
Both versions run on the same tasks with fixed seeds, scored by a judge the change cannot edit.
- 03
Keep it or revert it
A change stays only if it beats the last version by more than run-to-run noise. Most are reverted.
- 04
Check once on unseen data
What is kept is confirmed on a held-back set, and every decision is written to a ledger.