Petri AI Testing 'Closes' possible solution without looking (github.com)

A developer has added BonepokeOS — a compact, 300-line cognitive architecture that measures "narrative metabolism" (a metric for narrative coherence) — to a fork of the Petri AI-testing framework and pushed a minimal integration to GitHub. The patch includes bonepoke.py (the core system) and target.py (a 10-line Petri adapter) and can be invoked quickly with the model-role flag: --model-role target=petri.bonepoke_adapter:BonepokeTarget. The author frames this as an exploratory, hypothesis-free experiment: they want to see how a system specifically designed to evaluate narrative coherence behaves when subjected to professional red-teaming. This is significant because it demonstrates how Petri can accept lightweight, purpose-built cognitive evaluators and run them as alternative model roles, enabling fast, reproducible probing of specific model behaviors. For the AI/ML community, BonepokeOS offers a concrete instrument for measuring coherence-related failure modes, which can help surface subtle vulnerabilities (e.g., hallucination, incoherent story drift) that standard benchmarks miss. Technically, the low-touch adapter and small codebase mean labs can wire it into existing Petri test harnesses in minutes, facilitating practical adversarial testing and iterative improvement of narrative-sensitive systems.