We’ve released v1 of Allium: an LLM-native specification language that sits between prose (where contradictions hide), and code (where intent and accident are indistinguishable).

LLMs treat everything in a codebase as intentional, including workarounds and expedient hacks. Prose specs help, but without formal structure they rely on the author’s diligence to surface contradictions before they become bugs. Allium makes that structure explicit so the format does more of the work.

Every change to the language is debated by a nine-AI review panel. Each panellist has distinct priorities and the debate follows a structured protocol: present, respond, rebut, synthesise, verdict. Check out the TEAM file for the personas that collaborated on v1.

Looking forward to working with the panel on v2!