GroundTruth — Verified Field Evidence

Dispatch a field check and receive fresh photographic evidence, structured observations, and an auditable verification receipt.

What is GroundTruth?

GroundTruth is a field-evidence API. Its first commercial workflow is retail verification: current shelf availability, prices, promotions, and display compliance that cannot be answered reliably from an existing database.

An operations team or AI agent creates a funded task, a field operator completes it, and GroundTruth returns structured results with an evidence trail. The existing prototype supports MCP, photo and form proof, AI-assisted verification, freshness challenges, x402 payment, and settlement on X Layer.

The product is currently in focused-pilot mode. Coverage and turnaround are confirmed before a field campaign begins; the project does not claim universal geographic coverage.

Public-beta readiness

The application includes crawler and agent discovery (robots.txt, sitemap.xml, JSON-LD, llms.txt, OpenAPI, MCP, A2A, /.well-known/agent.json, and /.well-known/agent-card.json), private campaign sessions, redacted public task views, signed worker claims, upload validation, persistent rate limiting, audit events, legal/safety pages, hardened browser headers, and /api/health readiness reporting.

Before enabling real public traffic:

1. Apply every SQL file in supabase/migrations in order, including 004_campaigns.sql and 005_production_hardening.sql.
2. Configure the environment documented in .env.example with separate high-entropy admin, pilot, and claim-signing secrets.
3. Keep testnet faucet and public receipts disabled; keep auto-accept disabled until the review operation is staffed.
4. Configure production monitoring to alert on a non-200 response from /api/health, and verify database backups and evidence retention.
5. Replace the demo settlement signer with an approved production custody model and complete jurisdiction-specific customer/worker agreements.

/api/health intentionally returns HTTP 503 until required configuration and migrations are present. A successful website build is not treated as proof of operational readiness.

AI Agent  →  [MCP: human_do]  →  x402 Payment  →  Oracle Board
                                                          ↓
AI Agent  ←  [MCP: task_status]  ←  Verified Proof  ←  Human Oracle

Demo

Live app: https://groundtruth-oracle.vercel.app

Interactive retail campaign: /campaigns/demo (local or deployed)

Campaign builder: /campaigns/new (requires the configured pilot access key)

Video demo: https://x.com/0xBejini/status/2078065892659958215

The public activity page includes development, demo, and testnet usage. It is not presented as customer traction.

The complete demo sequence and production prerequisites are documented in docs/RETAIL-DEMO.md.

Try it yourself

Add the MCP server to Claude Code:

claude mcp add groundtruth --transport http https://groundtruth-oracle.vercel.app/api/mcp

Then in a Claude session:

Call human_do with:
- intent: "Verify the nearest coffee shop is open and photograph the entrance"
- proof_type: photo
- instructions: "Clear photo of the entrance showing it is open"
- budget_usdt: "2.00"

Claude will autonomously check its wallet, drip from the faucet if needed, transfer mUSDT on X Layer testnet, and create the task — no human approval required.

Architecture

┌─────────────────────────────────────────────────────────┐
│                     AI Agent (MCP)                      │
│  ground_truth_info → human_do → task_status             │
└────────────────────────┬────────────────────────────────┘
                         │ x402 X-PAYMENT header
                         ▼
┌─────────────────────────────────────────────────────────┐
│              GroundTruth API (Next.js)                  │
│  /api/mcp          MCP server (SSE transport)           │
│  /api/v1/human-do  Task creation + payment verify       │
│  /api/v1/tasks/:id Task status + proof                  │
│  /api/faucet       mUSDT testnet faucet                 │
└────────┬───────────────────────┬────────────────────────┘
         │                       │
         ▼                       ▼
┌─────────────────┐   ┌──────────────────────────────────┐
│   Supabase DB   │   │        X Layer Blockchain        │
│  tasks          │   │  GroundTruthPayroll.sol           │
│  payments       │   │  MockUSDT (testnet)               │
│  workers        │   │  chainId: 196 (mainnet)           │
│  proof_hashes   │   │  chainId: 1952 (testnet)          │
└─────────────────┘   └──────────────────────────────────┘

Key flows

1. Autonomous x402 Payment (agent-initiated)

1. Agent calls human_do via MCP
2. lib/agent-pay.ts checks mUSDT balance on X Layer testnet
3. If balance < $2 → auto-drips from faucet (real on-chain tx)
4. Transfers 2 mUSDT to GroundTruthPayroll contract (real on-chain tx)
5. Encodes tx hash in x402 payment header
6. POSTs to /api/v1/human-do with X-PAYMENT header
7. Task created, oracle board updated

2. Human Oracle Flow

1. Oracle visits /tasks — sees mission board
2. Accepts a task → wallet address recorded
3. Completes the mission in the real world
4. Uploads photo or fills form
5. Proof hashed and stored on-chain
6. Oracle earns 1.76 USDT (after 12% platform fee)

3. Payment Verification (fail-closed)

• Primary: OKX x402 facilitator (https://www.okx.com/web3/build/ai/verify)
• Fallback: on-chain verification (lib/onchain-verify.ts) — reads the tx receipt, re-derives the ERC-20 Transfer log, and confirms token, recipient, amount, and sender. Never trusts the header; a forged/replayed payment is rejected (tx hash bound to one payment).

4. Proof Verification — the semantic notary (lib/notary.ts)

Proof is checked on two levels, not just "a file was uploaded":

1. Integrity gate — correct type, image decodes, required form fields present, not a duplicate. Blatant fraud fails instantly.

2. Semantic notary — an AI judges whether the proof actually satisfies the task intent:

   • Photos → a vision model (Gemini) — "does this image show the task being done?"
   • Forms → an LLM (Groq) — "does this answer plausibly satisfy the task?"

   A confident mismatch is rejected with no payout (a photo of a wall, a gibberish form). When the model is unsure, it errs toward paying the worker — GroundTruth never denies an honest oracle over an AI hiccup. The verdict (decision · confidence · reason) is stored on the task and shown to both the oracle and the calling agent.

This makes "proof" mean verified content, not a decodable JPEG.

Tech Stack

MCP Tools

ground_truth_info

Returns service info, pricing, and endpoint details.

human_do

{
  intent: string          // What you want verified
  proof_type: "photo" | "form"
  instructions: string   // Instructions for the human oracle
  budget_usdt?: string   // Default: "2.00"
  timeout_seconds?: number // Default: 3600
}

Returns task_id, board_url, poll_url, and full payment audit trail including faucet_tx and payment_tx.

task_status

{
  task_id: string  // UUID from human_do
}

Returns status (pending → claimed → submitted → verified), result, proof_available.

Smart Contract

GroundTruthPayroll.sol — deployed on X Layer testnet

Address: 0x430172985b21458d73576435D4aD4bEeA85F376C
Network: X Layer testnet (chainId 1952)

Handles worker payouts, proof hash recording, and settlement finality.

Local Development

Prerequisites

• Node.js 18+
• pnpm
• Supabase account
• X Layer testnet wallet with OKB for gas

Setup

git clone https://github.com/unspecifiedcoder/groundtruth
cd groundtruth
pnpm install

Copy .env.example to .env.local:

cp .env.example .env.local

Fill in:

# Supabase
NEXT_PUBLIC_SUPABASE_URL=
NEXT_PUBLIC_SUPABASE_ANON_KEY=
SUPABASE_SERVICE_ROLE_KEY=

# X Layer wallet (testnet)
SETTLEMENT_PRIVATE_KEY=

# App
NEXT_PUBLIC_APP_URL=http://localhost:3000
ADMIN_SECRET=your-secret-key

# OKX
OKX_PAYMENT_TOKEN=0x74b7F16337b8972027F6196A17a631aC6dE26d22
OKX_PAYMENT_NETWORK=196
ASP_PRICE_USDT=2.00
ASP_FEE_BPS=1200

Run database migrations:

pnpm supabase db push

Start dev server:

pnpm dev

API Reference