Related work: governed substrate evaluation

[headsoftsoftware]

I’m studying whether durable evidence, provenance, route selection, and proof obligations can improve language-model agents beyond ordinary prompting or retrieval. I released a small process-isolated study with explicit controls: 18/48 raw, 14/48 flat evidence, and 48/48 governed routing on a frozen task family.
The paper is here: https://headsoft.org/research/equal-evidence/
DOI: https://doi.org/10.5281/zenodo.21367631
I would appreciate pointers on how AgentBench could best test whether this kind of governance improves long-horizon agent behavior without conflating it with extra retries or retrieval.