Inspiration

Freelance and AI-assisted work has a trust gap. Buyers don't want to release payment for an incomplete result, and workers don't want to finish a milestone without knowing the funds exist. Today that decision usually comes down to a chat message or a manual dispute.

Prooflane tries a narrower, auditable alternative: authorize the payment first, evaluate concrete delivery evidence against explicit checks, and let deterministic policy decide whether capture is safe.

What it does

Prooflane creates a milestone with a price and acceptance checks. It then creates a PayPal Sandbox order with AUTHORIZE intent, so approval reserves the money without capturing it. Once the order is approved and authorized, the worker submits delivery evidence.

An AI verifier evaluates every acceptance check and returns pass/fail, a reason and a confidence score. The model never controls PayPal. Application code captures the payment only when every check passed, confidence meets the threshold (90%), the authorization belongs to this milestone, and it hasn't been released before. Ambiguous evidence stops for review.

Every assessment lands in a milestone ledger built with AG Grid, and the same flow is exposed to AI agents through an MCP server: an agent can create a milestone, reserve payment through PayPal, submit proof, and get paid only when the proof holds up. That makes Prooflane a trust layer for agentic commerce, where neither side has to take the other's word.

How we built it

Node.js HTTP server with no runtime dependencies
PayPal OAuth 2.0, Orders v2 with AUTHORIZE, Payments v2 authorization capture
Pluggable OpenAI-compatible structured verifier (demo: Qwen 2.5 7B Instruct running locally via Ollama), with output normalization that fails closed
AG Grid Community milestone ledger
MCP server (stdio JSON-RPC, no dependencies) exposing the payment flow to agents
Deterministic release state machine with duplicate-capture protection
Node test suite covering AI response handling, PayPal API sequencing, unsafe release rejection and single capture

Challenges we ran into

The hard part was keeping probabilistic interpretation away from an irreversible payment action. Prooflane treats model output as an evidence assessment only; a separate deterministic policy owns the capture decision and rejects missing authorization, failed checks, low confidence and repeated release attempts.

PayPal's approval, authorization and capture are separate lifecycle steps, so the UI also had to make the difference between reserved funds and captured funds obvious.

Accomplishments that we're proud of

Real PayPal Sandbox runs end to end: order → approval → authorization → AI verification → capture, both from the web console and from an AI agent over MCP
The AI rejected a vague "trust me" claim (20% confidence, every check failed with a reason) and approved concrete evidence (100%, every check passed)
Uncertain evidence fails closed; PayPal credentials stay server-side
Order, authorization and capture IDs of the verified runs are published in the repo

What we learned

The useful boundary isn't "AI decides whether to pay." It's "AI turns messy evidence into a structured assessment, while code enforces the money invariants." PayPal's authorize-then-capture flow makes that separation concrete.

What's next

Durable storage, signed evidence artifacts, a human review queue, webhook reconciliation, authorization-expiry handling, and PayPal partner capabilities for a compliant platform-fee model. Before live use it would need legal review; this prototype is not an escrow service.