HIATUS-bench v0.3 — A Minimal Benchmark for Discriminant Experimental Design (Dataset + Paper)

This repository provides the full experimental bundle for HIATUS-bench v0.3, a discriminant framework designed to evaluate informational transitions across controlled micro-worlds. Unlike standard benchmarks focused on predictive accuracy, HIATUS-bench evaluates the ability of models to actively select experiments that reduce uncertainty between competing hypotheses. The benchmark is built on a variational and geometric perspective of transitions, where competing models are evaluated based on their ability to discriminate between structured dynamics. The bundle includes:- core implementation- baseline policies (random, fixed, greedy)- micro-worlds (relaxation, oscillator)- configuration files- reproducible runs (seed42)- results and evaluation scripts The goal is to provide an auditable and minimal experimental setup for testing discriminant power across paradigms. This work is part of a broader research program on informational transitions, variational geometry, and adaptive evaluation frameworks. Disclaimer:Personal work (concept, structure, arbitrations) with AI assistance for writing, formatting, and a minimal Python POC. All artifacts and tests are provided for reproducibility.