Verifiable decisions

Verify a model's decision yourself

Below is a real credit-risk model's decision, cryptographically signed. You do not have to trust it. Your browser re-runs the decision from the model's exact integer form, reproduces it to the bit, and checks the signature, all on this page, no server. Then forge the receipt or corrupt the signature and watch it get caught. This is what a model decision should be: reproducible and provable, not "trust us".

checking your browser…

The signed decision

applicant prediction model signature

Decide for anyoneruns in your browser

Move the sliders. The same integer model runs on your machine and reproduces a decision for any applicant, bit for bit.

Repayment status uses the dataset's own scale: -2 paid early, 0 on time, a positive number means that many months late.

reproduced in your browser

Verification log

ready.

A real credit model (LightGBM) trained on the public UCI "default of credit card clients" dataset, lowered to a portable integer forest that reproduces its decision bit for bit on any machine (CPU, GPU, this page), plus an Ed25519-signed receipt. It is a deliberately ordinary model: the thing we are proving is reproducibility and authenticity, not accuracy. The lab's determinism and Babel demos prove the same idea for physics and compilers. Built by Coherence Energy Labs.