Luca — ICAI CA-GPT

The model never writes a spreadsheet formula

The model emits an intermediate representation: inputs, rows and scenarios. A deterministic compiler turns that into the workbook, assigning every cell address, laying out sheets and wiring cross-sheet references.

Because the IR has no cell-address surface at all, a whole class of cross-sheet reference bugs becomes unrepresentable rather than merely discouraged by a prompt.

Clarify first, and remember what is still open

Answers carry an "Unconfirmed — answer these to refine" block and an "Assumptions" block. The orchestrator parses those questions back out, so when the user answers only some of them the system knows which are still open.

Five providers behind one health-scored registry

No part of the application talks to a model vendor directly. Every provider carries a health score; successes raise it and failures lower it, with a rate limit costing more than a transient error and an auth or quota failure zeroing it outright.

Cooldowns are typed per failure class. Below the health floor a provider is skipped entirely, and Azure OpenAI is the guaranteed floor, so the failure mode is a slower answer rather than no answer.

5 LLM providers11 chat modes10 compliance modules