Premise: LLM outputs are untrusted. BELGI only accepts outputs that can be verified against a locked spec + cross-file invariants, with a default NO-GO posture on unknowns or missing evidence.

Mental model: lock -> verify -> seal. Gates decide GO / NO-GO and produce replayable, tamper-evident artifacts.

I’d love feedback on:

where this fails in real CI pipelines,

missing invariant classes / bypass vectors,

whether default NO-GO is usable at scale (and how you’d stage it).