Radar / Ideas / Claim-check CI: a merge gate that…

Claim-check CI: a merge gate that verifies every factual claim in AI-generated docs

daily ideamoderateJEV confidence 0.54JEV monetization 0.482026-10-15

OutcomeNo AI-generated whitepaper, RFP response, or changelog ships with an unsupported factual claim.

The problem

Teams draft docs with AI ten times faster than before, but verification has not kept up: this is the validation bottleneck. One wrong number in an RFP response or whitepaper costs deals and credibility. Manual review does not scale, and existing checkers police grammar and brand voice, not whether each claim is actually supported by a source.

The idea

A CI action that takes a markdown doc plus a folder of source materials, runs a Deep Review-style agentic pipeline (segment claims, verify each against sources in parallel, synthesize a verdict), and emits a typed Jev judgment per claim: supported, unsupported, or needs-source. The check blocks the merge or posts an annotated PR review when unsupported claims remain. Priced per verified document, not per seat, so docs teams can start on one repo.

Why now

Google Deep Review went public on Oct 14 with exactly this four-stage agentic verification pattern, validated on 16,000+ conference manuscripts. Typed decision models (Jev) make the pass/fail verdict machine-readable, so it can live inside CI instead of inside a human's inbox.

What it combines

google-deep-review (agentic verification pipeline: segment, budget, verify in parallel, synthesize) + autotrust-jev-35b (typed judgments turn verdicts into CI-readable gates) + stepfun-step-5-preview-open-bf16-weights-released (cheap 1M-context reader to run verification agents at a fraction of frontier API cost). The mix turns a research demo into shippable CI infrastructure.

MVP

Weekend MVP: a GitHub Action that splits a markdown file into claims, checks each against a provided sources folder, and posts supported/unsupported verdicts as a PR comment. Deliberately skip: the parallel agentic pipeline, the Jev typed gate, and billing. Prove verdict quality on ten real docs first.

How it makes money

Budget owner: Head of Technical Writing / docs platform lead, already paying for docs tooling and review workflows. Pricing anchor: GC AI publishes $500/seat/month for AI document review (https://www.vaquill.ai/blog/legal-ai-pricing-benchmark); Spellbook charges $99-$299/user/month for AI-assisted review (https://beforeyou-sign.com/blog/best-ai-contract-review-tools). Willingness-to-pay evidence: Harvey charges roughly $1,200/lawyer/month with 98% gross revenue retention across ~100,000 lawyers, proving buyers pay four figures for AI document review (https://github.com/bra1ndump/skillpack/blob/HEAD/agent-runs/2026-03-24_01-37_run_business/deep-dive_business/findings.md). Replacement budget: displaces manual review hours and review-tool seats.

The moat

Harvey and Writer cannot copy this in one quarter: it serves a different buyer (docs/engineering teams, not legal or marketing), on a different business model (per-document CI usage, not per-seat enterprise), through a channel their enterprise sales motion cannot reach (the GitHub Actions long tail). Copying it would mean building a second go-to-market from scratch.

Distribution

B2B2C via the GitHub Actions marketplace and docs platforms (Mintlify, ReadMe). Wedge: docs-as-code teams already running CI on every docs PR; land one repo, expand to the org.

Why it wins

Writer.com polices brand voice; Harvey and Legora review contracts for legal teams. Nobody sells factual claim verification as a CI gate for engineering and docs teams on per-document pricing.

Risks

Verification quality on nuanced or implicit claims is the hard part; a gate that cries wolf gets disabled. De-risk by measuring precision and recall against human-labeled claims before wiring the merge block.

Build it with

Google Deep ReviewThe four-stage agentic verification pipeline pattern to replicate in CI
autotrust JEV-35B: 35B-A3B open decision model with calibrated probabilitiesTyped judgments (supported/unsupported/needs-source) as the machine-readable merge gate
StepFun Step-5-Preview — open BF16 weights releasedCheap long-context reader for the verification agents

Repo to start from

claimcheck-ci: GitHub Action with a claim splitter, a source-grounded verifier, and a PR comment reporter.

Evidence

Get the week's best AI launches, plus 3 ideas worth building

One email every Saturday. Ranked by traction, not hype. Free.