This is a race between two offline machines.

The source file was a 44-page Ohio Bureau of Criminal Investigation prosecutor summary concerning an officer-involved critical incident in Findlay, Ohio. The exercise used a counterfactual prompt: assume the subject survived, was charged with felonious assault on a peace officer, and the prosecutor was preparing the State’s case-in-chief.

The offline AI system was asked to identify the strongest direct-examination witness and explain why, based only on the BCI report. Its answer selected the officer with the cleanest observation of the alleged act: the officer who saw the subject turn and point the handgun directly at him. The typewriter's redacted output is here.

A separate citation check found no hallucinated support. The quoted lines and page references were verified against the report: the muzzle observation, the limitation in the first officer's account, the second officer's view of the subject's body mechanics, the direct-pointing language, the recovery of the firearm, and the timing witness all checked out.

The run took 3 minutes and 25 seconds. The model used roughly 22 percent of its available context.

This work does not require the cloud. It does not fail when the internet fails. It does not meter every question. It sits inside the office, making review faster while preserving human judgment.