evaluation-methodology

An evaluation methodology for context-augmentation experiments: a format-matched control that separates whether added context helps because of its information or its format. Pre-registered instrument, mutation-audited scorer, worked example in SVG reference resolution.

Diagnostic of physics literacy in frontier LLMs. Evaluates induction, formulation, and prediction inside unfamiliar physics frameworks via dual-judge inter-rater reliability with human-audit resolution.

MCDM using TOPSIS: A technique for multi-criteria decision making, utilizing the TOPSIS method to evaluate alternatives based on multiple criteria.

Code, data, and paper for an evaluation-protocol-adjusted residual audit of public LLM benchmark scores — predicting score variation from public metadata and using the residuals as a review queue.

Runtime-verifiable audit harness that falsifies coding-agent intervention claims — preregistered, oracle-checked, reports null.

The instrument makes the winner: reconciling two contradictory 2026 clinical-AI evaluations (Real-POCQi vs Nature Medicine). A 2x2 instrument x rater decomposition shows the evaluation instrument (pairwise vs rubric), not the model, drives the disagreement.