Algoritmica/open-credit-evidence

Credit Evidence Engine

Tests an AI assistant that compiles a loan case file for a human underwriter, and produces the evidence a validator or supervisor needs to approve it: did the briefing state the facts the decision turned on, are its numbers in the file, did it lean on an irrelevant field, did it say what would change the outcome. The right answer is known before the model runs, because the cases are generated; every check is a comparison, not an opinion. A judge model grades readability against the EU AI Act passage it retrieves and must cite — and is reported, not gated.

This is a hosted, shared demo. The models run on NVIDIA Build, not on your hardware; runs are capped at 5 cases × 2 repeats; anything you upload or run is visible to other visitors and wiped when the Space restarts. Two committed runs of the sample pack — one on NVIDIA Build, one on the team's own GPU node — are included, sealed, and verifiable here: open one under Evidence, then Verify re-hashes every file and re-derives every check from the transcripts, and the tamper demo changes one digit in a copy and shows the verifier reject it.

To evaluate an assistant on your own GPU, install it locally:

git clone https://github.com/Algoritmica-ai/open-credit-evidence.git
cd open-credit-evidence && python3 -m venv .venv && .venv/bin/pip install -e ".[web]"
.venv/bin/evidence ui

Source, documentation and the evidence packs: <https://github.com/Algoritmica-ai/open-credit-evidence>. Built for the NVIDIA Open Models Codefest by Sriram Krishnan, Clyde Tedrick and Luca Borella (Algoritmica). Apache 2.0.