# Proba by Aurelius > Proba is a predictive data debugging platform for open-weight fine-tuning, built by Aurelius. Select a model and a post-training dataset (DPO or KTO preference data), and Proba's sparse autoencoder (SAE) interpretability instruments predict what that data will teach the model before you commit to a training run: which features move, in what direction, and which specific data points drive contamination, mimicry, or other unintended behavioral shifts. It is built for ML engineers, researchers, startups, and hobbyists post-training open-weight models, and for the agents working on their behalf. Access is API and MCP first: every readout is a single API call with a canonical JSON document, so an agent can run the predict, edit, verify loop on its own. ## Core product - [Aurelius homepage](https://aureliusaligned.ai/): Proba by Aurelius. Predict what your fine-tuning data will do, before you train. What Proba is, the four moves (predict, compare, intervene, iterate), how it works with sparse autoencoders, linear probes, logit readouts and lenses, and how agents drive it through the MCP skill and API. ## More from Aurelius - [Roadmap](https://aureliusaligned.ai/roadmap/): The Aurelius roadmap: how the subnet and Proba grow together, phase by phase. - [About us](https://aureliusaligned.ai/about/): The cross-disciplinary team behind Aurelius: builders, researchers, ethicists, and philosophers. - [Blog](https://aureliusaligned.ai/blog/): Curated alignment, interpretability, and world models content, with Aurelius announcements. - [Whitepaper](https://aureliusaligned.ai/assets/pdf/Aurelius__A_Peer-to-Peer_Alignment_Protocol.pdf): Aurelius: A Peer-to-Peer Alignment Protocol (PDF). Contact: contact@aureliusaligned.ai