Benchmark-gated
Every change is kept only if it raises a measured score, no vibes, no self-grading.
Self-improving AI · framework
SIA wraps any model or agent and autonomously raises its performance on a benchmark task, measured, not asserted. The scaffolding learns; you own the direction.
Every change is kept only if it raises a measured score, no vibes, no self-grading.
Point it at any model or agent. SIA improves the system around it, not the weights.
Runs the propose / test / keep loop on its own until the gains stop coming.