CV-IR is a tri-state validator that checks whether
a code tree actually satisfies a list of acceptance criteria —
not by trusting the AI's self-report, but by running programmatic
probes and a separate LLM judge whose citations must be
literal substrings of the code. Verdicts come back as
pass, fail, or honest indeterminate.
The headline case (go-chi/chi):
I wrote an intent that said "this router doesn't bundle authentication."
The validator returned FAIL
with cited evidence — middleware/basic_auth.go exists.
The intent's claim was wrong about the code I had been pointing
the validator at. CV-IR catches mismatches between what the
intent claims and what the code does — the kind of overclaim no
self-reporting AI can flag.
Describe what a piece of software is supposed to do or be. An LLM blind to the target codebase proposes 5 acceptance criteria. The validator then runs probes + a separately-prompted LLM judge against one of four demo repositories. Verdicts come back with substring-verified citations. Total round trip: ~60 seconds.