MFDS Advertisement Review
“Did the brand do due diligence before publishing?”
Public verifier
See an AI decision prove itself.
Pick a real scenario below. Watch what the agent considered, what it chose, and what we'd show a regulator or a judge if anyone challenged it. Every example is anchored on Base Sepolia right now — clickable, checkable, can't be edited.
Bloom Co. · scenario №02
Ad copy — verified live
Anchored on Base Sepolia · Apr 26, 2026, 6:36 AM · tx 0x2235d540…
✓ all 6 checks passOpen on easscan ↗
What Bloom Co. actually got
Scenario No.02
Bloom Co. · marketing agent · regulator-watched ad copy
Marketing AI shipped an ad that meets MFDS rules.
GPT-5 drafted three headlines. Claude reviewed each against MFDS guidelines. The founder signed off at 10:09 KST.
What the AI considered
- "Cures insomnia" — rejected: treatment claim.
- "Fights sleep disorder" — rejected: implies medical efficacy.
- "Supports relaxation before sleep" — permissible structure-function claim.
What it chose
Compliant variant. Drafted on Meta Ads. Sent for human approval.
Why this matters
If MFDS opens a post-publication audit, the chain is right there: two LLM calls, the rejected high-risk options, the policy refs, the founder signature.
Who can ask, and what they ask
The same record satisfies very different questioners. Each gets the same answer — and none of them have to trust the other party's word for it.
Meta Ads policy team
“Was the copy reviewed by a human?”
A future ad-fraud lawyer
“Was this AI alone or with sign-off?”
What this proves — in plain English
This decision happened.
Recorded on Base Sepolia at April 26, 2026 at 3:36 PM KST. The transaction is permanent and publicly visible — try editing it.
These were the inputs.
Every prompt, every retrieved context, every alternative the AI considered — fingerprinted into the same Merkle tree. Change one byte and the entire root invalidates.
This was the chosen output.
Not just what shipped — what was rejected, scored, and overridden. A future investigator can see the full reasoning, not just the final answer.
A neutral third party witnessed it.
trace.ai anchored the root with 0x4eC5A687…; the operator signed the payload with 0x1e6d0f5e…. Neither party can produce this evidence alone.
tx · 0x2235d540633130…5e8ecd5a
Run a simulation
Watch a decision earn its receipt — across 7 industries.
Pick a scenario. Press play. Each step shows a different artifact — the ticket, the score bars, the signature, the Merkle tree, the block, the verifiers — so you see the proof being built. Same protocol works for e-commerce, finance, healthcare, HR, insurance, and legal.
Currently simulating
CS bot processes a refund
Next
Want to put your own AI agent on the record?
One npm install. One SDK call. Every decision gets a permanent, independently verifiable receipt.