Verdict Desk reads a work posting, gig listing, or bounty brief and returns a verdict card — GO, CAUTION, or NO-GO — where every claim is a verbatim quote of the posting. The first verdict is free. The full evaluation, the claim checklist, and a recheck after the poster fixes the posting are paid products that unlock only when a PayPal order is captured.
Before anyone, a person or the AI agent working for them, spends days on a posting, they should know whether the money behind it is real. Today the proof is buried in prose: who funds the work, how the winner gets paid, and whether the poster can decline to pay at all. Verdict Desk reads the posting, quotes the exact lines that prove or fail each point, and says plainly what is still unproven. The full verdict is bought through PayPal. Our bet is simple: when demand can be checked before work starts, more of it gets posted and more of it gets done.
The problem. Freelancers, bounty hunters, and the AI agents now working for them choose work from prose. A posting can promise a prize it never funded, bury an identity demand in its terms, hide a clause that lets the poster decline to award at all, or contradict its own results page — and the worker finds out after the work is done. Job and task scams are a fraud shape every trust team knows; the quieter version, the unfunded or unawardable posting, wastes the same days without ever calling itself a scam.
What the desk does. Paste a posting or give its public URL. A deterministic check battery reads the text and runs six checks: core terms, fill status (has this already been awarded or closed?), payout route, identity demands and the stage they bite at, sponsor discretion, and fit. Every check returns pass, caution, fail, or unknown, and every claim on the card is printed with the quoted line it rests on. A check that cannot read its fact says unknown; the desk never guesses. Where a posting cites something checkable, the card labels the state verified; where the posting only claims funding in its own prose, the card labels it claimed, quoted. Those two states are never blurred.
The AI layer. An AI layer does the three reading jobs a deterministic battery cannot. It reads postings whose terms are written in loose natural language and proposes structured terms — payment route, funding claim, identity demand, award discretion — each pinned to the exact line it came from. It writes the plain language explanation on the verdict card, the paragraph a poster reads to understand what to fix. And it triages CAUTION cards, grouping the defects into the smallest set of edits that would move the verdict, which the paid recheck then verifies. The guard is structural and it is the product's honesty claim: the layer proposes, the battery verifies each proposal against a verbatim excerpt and against its own evidence and status, and the layer never sets the verdict. Proposals the battery cannot back appear on the full card as dropped, with the reason.
How PayPal carries the product. The worker pays for the full verdict; the poster pays for the recheck after fixing their posting. Two payers, two PayPal flows, both standard checkout. The desk creates the order through the PayPal Orders API, the buyer approves it (through the PayPal JavaScript SDK buttons, or the hosted approval page), the desk captures it, and the full evaluation unlocks on the captured order — not before. Capture is confirmed by webhook when a webhook id is configured, with the capture response and order status visible in the app and on the receipt view. The order lifecycle is complete: if a paid evaluation fails after capture — for example the posting page can no longer be fetched for the fresh full run — the capture is refunded automatically, and the buyer sees the refund on the receipt. Remove PayPal and there is no way to sell the full verdict, no order lifecycle to show, and no capture event that unlocks the deliverable; the payment is the mechanism of the product, not a decoration on it.
The three case studies in the demo are reconstructed from real cases with sponsor names and links changed. A bounty page whose own results post contradicts it (the fill check fails it). A builder offers page carrying a may-not-award clause and direct-to-wallet payment (CAUTION, quoted). And one clearly labelled illustrative sample with entry-stage identity demands and a sponsor-token payout (NO-GO).
Why this matters beyond one desk. Every check the desk sells is itself demand being verified before work starts, settled on a rail both people and agents can use. Posters who can prove their money is real get serious applicants; workers who can check terms in a minute stop burning days on postings that were never going to pay. That is the market this project advocates for: more agent economy activity, because more of it can be checked before it begins.
Built with
- PayPal developer platform — Orders API v2 (create, get, capture, refund), the PayPal JavaScript SDK for the approval flow, and webhook signature verification for capture confirmation. Sandbox environment.
- AI tools — the AI layer is implemented against an OpenAI-compatible chat model endpoint. When an API key is configured, the named model drafts term proposals (with verbatim quotes) and may draft the explanation; the deterministic battery verifies proposals, and a drafted explanation is accepted only sentence by sentence. As submitted and demoed, no model key is configured: proposals come from the local extractor shipped in the repository (
src/ai-layer.ts), explanations are composed from the verified check results, and the app labels which source produced each artifact. The verdict is always set by the deterministic battery. - Everything else — Node.js built-ins only (
node:http,node:test). No runtime dependencies.
