See the README.md in the repo for installation instructions and an overview of what's new in this version.

This release is associated with the model file in the release binary sdm_estimator_alpha0.9_m1000_e250_iter5__over_gpt4.1_o4mini_gemini2.5pro_granite3.3_8b_instruct__2025_07_26.zip, which contains an SDM estimator with an exemplar vector size of 1000 and alpha'=0.9, trained and calibrated over a subset of the OpenVerification1 dataset (see documentation/DATA.md).

This SDM estimator takes as input classification decisions and explanations from gpt-4.1-2025-04-14, o4-mini-2025-04-16 (reasoning_effort="high"), and gemini-2.5-pro, as well as the hidden states and output of an agreement classification over those explanations from granite-3.3-8b-instruct (which runs locally on the machine running the MCP Server). The Opus 4 model is the recommended tool-calling LLM.