In February a first working prototype of the AI analyst chat was completed. The prototype was exercised against live market data drawn from a small number of sources, with the principal question being whether a chat-based interface could produce analytical responses of the calibre that a serious investor would find worth reading.

The early results were uneven but encouraging. The model was capable of structured, useful analysis when constrained correctly; without constraint it tended towards the generic and the credulous. The work of the following months was largely a matter of designing the prompt architecture and data-grounding layer that would coerce useful behaviour reliably.