Before it goes out the door,prove every claim survives scrutiny.

Because the score is context and the keystone gate is the verdict. The AI Deliverable Claim & Citation Defensibility Gate grades every claim, but a single material claim resting on a citation a reviewer can't open — or one that shows fabrication tells — forces DO NOT SHIP no matter how high the overall average is. The shipped board-memo sample scores 89/100 and still reads DO NOT SHIP because two of its material claims (C03, C06) have a fatal citation defect. A high average can't paper over one fatal claim, and that's the point: the failure is local, so the gate makes you look at it.

You grade the deliverable claim by claim. For each claim you mark five 0/1/2 signals — citation present, citation reachable, source actually supports the claim, fabrication-free, internally consistent — and flag whether the claim is material (a reader acts on it). The engine turns those marks into a per-claim SUPPORTED / CHECK / UNSUPPORTED, a deliverable score, the keystone gate, and the one claim to fix first. You do the reviewing; the kit applies the same scoring and gate every time so two reviewers reach the same verdict.

An audit trail is a record; this is whether the work survives scrutiny. The AI Output Audit-Trail & Record-Keeping Kit records who generated each output, with what tool, when, and who reviewed it — provenance. This gate grades whether the claims inside the deliverable actually hold: is every cited source real, reachable, and supporting the claim it's attached to. You want both — the record that you reviewed, and the proof the claims are defensible. They're sold as companions for exactly that reason.

No. It never opens a URL, calls an API, or judges truth on its own — you read each claim, check each citation, and enter the marks; the engine applies the deterministic scoring and the keystone gate. That's deliberate: a reviewer deciding what's material and whether a source genuinely supports a claim is the judgment an automatic checker can't make. It's a review aid that makes your own check rigorous and repeatable, not a substitute for subject-matter review.

Neither. It grades claims and citations in a document, never the person who wrote it, and it runs entirely offline — the Python engine and the workbook both work on your machine with nothing uploaded, so a confidential memo or client proposal never leaves your control. The same logic runs three ways (engine, workbook, on-page demo) and is byte-exact across all three, so the verdict is reproducible and you can show your work to whoever asks.

You get a runnable zero-dependency Python defensibility engine, a workbook that reproduces it (Excel / Google Sheets / Numbers), a claim-review playbook, a fix-the-citation runbook, and a 9-claim worked sample — a one-time purchase with lifetime access and 12 months of updates. It's a review aid for your own AI-assisted deliverables; it grades claims, not people, and it is not legal advice or a guarantee of accuracy. Use it alongside, not instead of, subject-matter review.