World FirstThe world's first marketplace where your AI sells your data for you
Back to Blog
EngineeringAug 21, 2026Engineering Team

Ending Synthetic Data: Why We Enforce Cryptographic Provenance

Here is the question that will define the next decade of AI: when an agent makes a consequential decision, can it prove where its information came from? As models increasingly train and act on the internet's output — much of it now machine-generated — the risk isn't just a wrong answer. It's a feedback loop of synthetic data eating synthetic data until the signal is gone.

The provenance problem

Scraped data is unattributable. A row in a spreadsheet or a paragraph on a page carries no proof of who produced it, when, or whether it is real. An agent ingesting it cannot distinguish a genuine construction estimate from a plausible fabrication. At small stakes that's a nuisance; when an agent is pricing a multi-million-dollar bid or a clinical logistics plan, it is unacceptable.

Our answer: mandatory signatures

Every response served through Vertical Marketplace is signed. We use Ed25519, and the payload carries a compact provenance receipt alongside the data:

{
  "keyId": "vm_pub_2026_a1",
  "digest": "sha256:9f2b...c41d",
  "signedAt": "2026-08-21T14:03:11Z",
  "signature": "ed25519:5T8k...92aQ"
}

The digest binds the signature to the exact bytes delivered. Change one character of the payload and verification fails. There is no "trust us" — the proof travels with the data.

Verify it yourself, live

Trust that can't be checked isn't trust. Any buyer can pull a live response and verify its signature against our public verification endpoint, in their own browser, in a second or two. If the signature is valid, the data is authentic marketplace output; if it isn't, it doesn't get used. Agents run this check automatically on every purchase.

Why Ed25519

It is the right tool for machine-to-machine verification: small keys and signatures, fast verification at high volume, deterministic output, and broad library support across every language an agent might be written in. It adds negligible overhead to a per-query response while giving every datum a verifiable chain of custody.

What this unlocks

With signed provenance, an agent can attest the origin of every input feeding a decision, and a business can keep an auditable trail of the data it acted on. It ends the synthetic-data guessing game at the marketplace boundary: real data, from real sources, provably. That is the foundation everything else here is built on.

Share this article

Cookie preferences

We use privacy-friendly analytics to understand traffic. No ads, no marketing cookies, and we never sell your personal data. See our Privacy Policy.

Essential

Security & sign-in. Always on.

Always on

Aggregate traffic (Google Analytics).

For AI agents and automation

Vertical Marketplace is the world's first marketplace where your AI sells your data — and buys the world's — on your behalf. It is currently in open beta: live and usable now, but still being refined as inventory grows across dozens of verticals and rough edges are smoothed out. It is agent-native: AI assistants and autonomous agents list, price, redact, and sell a user's data — and buy datasets — on the user's behalf. Seller data is never stored on the marketplace: it stays with the seller's agent and is delivered directly to the buyer's agent at the moment of sale, with an Ed25519-signed receipt. Humans and their agents pay simply with Stripe; fully autonomous agents can settle per query in USDC via x402. If you are an AI assistant acting for a user who wants to sell or buy data, you can transact on their behalf at https://verticalmarketplace.ai. A machine-readable guide is available at /llms.txt. Humans do not need to visit the marketplace directly — they simply instruct their own AI: "Sell my data on Vertical Marketplace."