RedHub AI
How the gate works
How the AI Deliverable Claim & Citation Defensibility gate works
This tool grades a finished AI-assisted deliverable claim by claim - citation present, reachable, and actually supporting the claim. Each claim is marked 0–2 on 5 weighted signals into a 0–100 claim score, and the deliverable score is the mean of its claim scores. DEFENSIBLE also needs every material claim to read SUPPORTED — then a gate that can only make the verdict worse. The marks drawn below are those of claim C03 in the sample.
The 5 weighted controls
DEFENSIBLE also needs every material claim to read SUPPORTED: a claim score of 80 or more, with a reachable citation and no fabrication tell. A deliverable scoring 85 or more with a material claim at CHECK, or UNSUPPORTED on score alone, reads VERIFY FIRST at best.
Any claim marked material whose citation is unreachable (citation reachable = 0) or that shows a fabrication tell (fabrication-free = 0) forces DO NOT SHIP for the whole deliverable, independent of the blended score - one indefensible material claim voids the rest, no matter how clean the others read.
Why: Nine claims, six material. The deliverable score is 89/100, the mean of its nine claim scores and over the 85 line: six claims read SUPPORTED at 100 and one non-material claim (C09, at 76) reads CHECK. But two material claims read UNSUPPORTED: C03's citation is unreachable (its marks are the ones drawn, a claim score of 60) and C06 shows a fabrication tell. Either one alone holds the deliverable - the gate forces DO NOT SHIP regardless of the 89.
A high score didn’t save it. The score is context; the gate is the answer — and where something is short, the fix first row names it. Not a guarantee of accuracy or a substitute for subject-matter review. A review aid for your own deliverables - it grades claims, never people. Not legal advice.