Devpost submission draft (English)

Title

ProofPay — escrow for AI work: frozen first, paid only when the output checks out

Tagline

Your money is authorized, not spent. Nine objective checks decide: pass → capture, fail → void.

Problem

Buying AI-generated content today is pay-first-and-pray: if the output has a wrong price, a fake source link, or stale data, the buyer has already paid. Refunds and disputes are slow, manual, and rare.

Solution

ProofPay wraps a PayPal AUTHORIZE/CAPTURE flow around AI generation:

1. Buyer places a task (deliverable spec + data snapshot). PayPal only freezes the funds.
2. An AI generates the deliverable strictly from the frozen snapshot (here: a deal brief built from a PriceScout price feed).
3. Nine objective, re-runnable checks verify the output — count, length, SKU↔source binding, price, currency, freshness window (both bounds), deadline sanity, amount, dedup + version.
4. 9/9 pass → capture. Any failure → void; the buyer pays $0.00 and gets the failed-item list.

Why it matters

It turns "trust the AI" into "verify the AI" — a reusable settlement layer for agentic commerce: content shops, research briefs, data summaries, any AI deliverable that can be checked objectively before money moves.

Tech

• PayPal sandbox Orders v2 (AUTHORIZE intent) + Payments v2 (capture/void), verified end-to-end in sandbox on 2026-10-09 (capture COMPLETED; void VOIDED).
• Python stdlib demo app (order / verification / receipt pages), deterministic generator, offline checker CLI. Snapshot format reuses the PriceScout feed.

Honest scope

Gaps: live-order amount binding; cross-task uniqueness and generator-version authenticity rely on self-reported values without a trusted execution log. The demo video is a local simulation that mirrors separate sandbox API tests; no money moves on screen. We show exactly what is verified and what is not.

AI tools used

Muse / GPT-class coding agents for implementation assistance (disclosed per rules).

Built With

• ai-agents
• fintech
• paypal
• python