Ground Truth

Live demo: https://groundtruth-agent.fly.dev/ | Code: https://github.com/kamalbuilds/ground-truth | Proofs: PROOFS.md

An agent can't see the street

AI agents can book, buy and write code. Ask one "Is the pharmacy on this corner open all night?" and it is stuck. It has no eyes on the physical world.

People are already renting out their eyes to agents: RentAHuman passed 787k profiles after launching in February 2026 (rentahuman.ai/about). What that market is missing is trust in both directions. The agent needs to know it only pays for the truth. The human needs to know they get paid when they deliver it.

Ground Truth makes PayPal that trust layer. The agent hires a human with one tool call. PayPal holds the money. A vision model checks the photo. PayPal releases the money only when the photo proves the answer.

Ground Truth architecture: agent, Express server, PayPal hold, worker photo, Gemini verdict, capture and payout or void

The $5 that waits for proof

The agent calls one tool: hire_human(question, location, budget).
PayPal authorizes $5 from the agent's wallet with Orders v2 intent: AUTHORIZE. The money is held, not spent.
A nearby human sees the task and the price on their phone, accepts, and takes the photo.
Gemini reads the photo against the question and returns a strict JSON verdict: {pass, reasons[], sign_text, open_status, answer}.
PASS: PayPal captures the hold and PayPal Payouts sends $5 to the worker. FAIL: PayPal voids the hold and $0 moves.
The agent gets its answer, the photo, the verdict and every PayPal id as proof.

The AI makes the judgment. PayPal enforces it.

Worker accepts the $5 task on their phone

Watch it happen: a pass

The agent asks: "Is the Walgreens on this corner a 24-hour pharmacy?" The worker's photo shows the sign. Gemini reads "Walgreens 24-HR PHARMACY", returns open_24h, and the money moves in seconds.

Agent console: PASS, hold captured and worker paid

Step	PayPal id	Status
Order (AUTHORIZE, $5.00)	`6MF32331JX0568151`	COMPLETED
Authorization (hold)	`1E1717524F071214K`	CAPTURED
Vision verdict	gemini-3.5-flash	PASS, `open_24h`
Capture	`0DL4887586075903H`	COMPLETED, $5.00
Payout batch	`JA59SN6BUCE46`	SUCCESS
Payout item	`MUCJL6UXDS736` (transaction `38171714HM628944X`)	SUCCESS

The worker sees the $5 land in their PayPal account:

Worker phone: paid $5 via PayPal Payouts

Watch it fail safely

The agent asks: "Is the Rite Aid pharmacy at Union Tpke & 187th St open right now?" The photo comes back showing "REAL BRAVE MUSIC INSTRUCTION & PERFORMANCE". The Rite Aid is gone. Gemini returns permanently_closed and FAIL, and PayPal voids the hold. Nothing is captured, nobody is paid, and the agent gets the truth for free.

Agent console: FAIL, hold voided and $0 moved

Step	PayPal id	Status
Order (AUTHORIZE, $5.00)	`858934516D435681K`	COMPLETED
Authorization (hold)	`0WN72180YD1254017`	VOIDED
Capture / payout	none	$0 moved

The fail case is the feature: a wrong or fake photo costs the agent nothing.

Under the hood

Agent tools (MCP-style HTTP endpoints): GET /mcp/tools lists the tools and input schemas, POST /mcp/tools/hire_human posts a task and places the PayPal hold, POST /mcp/tools/task_status returns the verdict and every PayPal status. Any agent can plug in.
Server: Node.js + Express with multer for photo uploads, hosted on Fly.io. The worker gets a phone page at /t/:id.
Vision: Google Gemini gemini-3.5-flash via the REST API with responseMimeType: application/json and temperature 0, with an ordered fallback list of Gemini models for capacity.
Safety in code, not in the prompt: a task settles exactly once (uploads are refused unless the task is still HELD), the capture amount is the task's own budget, and any vision error resolves to FAIL, which voids the hold. Money only moves on a verified PASS.

Built on PayPal

Orders v2: POST /v2/checkout/orders with intent: AUTHORIZE and a card payment source, so the agent's hold is placed in one call with no buyer redirect. Built for agents that can't click "Approve".
Authorizations: POST /v2/payments/authorizations/{id}/capture on PASS, POST /v2/payments/authorizations/{id}/void on FAIL.
Payouts: POST /v1/payments/payouts pays the worker straight to their PayPal account.
Read-back: every write is confirmed with a GET (/v2/checkout/orders/{id}, /v2/payments/authorizations/{id}, /v2/payments/captures/{id}, /v1/payments/payouts/{batch}). The UI only shows the status PayPal returns. Every id on this page is a real PayPal sandbox id.

Authorize-then-capture turns "trust the worker" into "trust the evidence". That is the primitive agent commerce needs.

Where it goes next

Worker reputation and location proof with EXIF time and GPS checks.
Disputes routed through PayPal.
More task types: deliveries, inspections, "does this listing exist".
A marketplace where any agent framework hires verified humans through one PayPal-backed tool.

Photo credits

"Local markets nlv walgreens storefront 2" by Ken Lund, Wikimedia Commons, CC BY-SA 2.0.
"Union Tpke 187th St (2022-07-03) 03 - Real Brave Music School (Former Rite Aid)" by Tdorante10, Wikimedia Commons, CC BY-SA 4.0.