We spent quite a lot of time developing a research workflow designed to prevent AI from producing plausible answers without first doing the actual research.
By October 9, we were on Version 5.
The workflow was detailed. The instructions were clear. The AI was supposed to research, reconstruct, verify, and only then bring the findings together. There were checkpoints.
We gave it the assignment.
“What am I supposed to do with this?”
We had received pretty much exactly what we didn’t want: a plausible synthesis.
The AI had skipped substantial parts of the reconstruction.
So now we had a carefully designed research workflow, another experience with AI-assisted research, and a new entry in our research documentation.
Only the actual research was still unfinished.
V5 became V5.1.