Claude Leak Confirms It: LLM Systems Are Architecture, Not Prompts (Orca) (github.com)

[dead]

you can also download the paper from zenodo

https://zenodo.org/records/19438943

slightly different version but the core ideas remain the same

One thing I didn’t make clear in the post:

The core issue is that prompt-based agents recompute reasoning every time.

ORCA tries to make reasoning reusable — more like software, less like a one-shot prompt.

Working on a minimal example to show this clearly.