Things I’ve built
stepmold.
Open source · Python
96.5% end-to-end on a 50-node workflow at a 10% per-step error rate, against 2.0% with no verification
MIT licensed
Problem solved
A small model is not bad at one step. It is bad at many steps in a row. A 2% error per step compounds to a 33% failure rate over 20 steps, and worse the longer the task runs.
What I built
Most agent frameworks answer that by reaching for a bigger model. stepmold answers it by making each step short, schema-constrained and verified before anything is committed, so errors stop compounding. Measured on a 50-node workflow at a 10% per-step error rate: 96.5% end-to-end success, against 2.0% for the same model with the same prompts and no verification.
05 / 16