Japan just had its first real wave of MCP events. Agents calling each other's tools is finally in the open here, and that is a good thing.
But scroll through the tech posts, in Japan and almost everywhere else, and they all share one shape. "Look what I built." "Isn't this amazing." A slick demo, a clean screen, a pile of likes.
So let me ask the question nobody asks.
Who verifies the "amazing"?
Self-built or solo-built, it does not matter. The question is the same. Does it hallucinate? Is the answer actually correct? And can you really hand this AI the job?
That question is rarely asked out loud. That is a mistake, and it is about to get expensive.
A demo is not a proof
Here is why it matters now. We are heading into a world where AI agents call each other's tools with no human in the loop.
In that world, "trust me, it works" is worthless. The calling agent does not trust you the way a person does. It looks at how you behaved before, whether that record can be tampered with, and whether an outside party can recompute it. That is all.
Two things get quietly conflated:
- A working demo is a claim, not evidence. It ran once. That guarantees nothing about the next call.
- A green test you wrote and ran yourself is not verification. When the examiner and the examinee are the same person, the score means nothing.
Most "amazing" stands on those two. That is sand.
What verification actually means
Real verification, to me, is four properties.
- A third party can recompute it. Trust nobody. Take the output bytes, hash them yourself, get the same answer.
- Zero self-asserted lines. "We are excellent," written by you, never makes the record. Every line is generated only from a record whose bytes hash to its own id.
- Discrepancies are never hidden. When an outside observation once disagreed with the internal one, that fact stays in, visible. Hide one inconvenient record and the whole record is worth zero.
- Time is independent of the provider's clock. Did you know before the incident, or after? Stamp it with a clock the provider cannot move, not the provider's own publication time.
Meet all four and "amazing" becomes "provable."
We committed to all four. Go check.
This will read like a pitch, so I will only write what you can verify. Do not trust this post. Hit it yourself, right now.
Our verification ledger returns an agent's track record in a form anyone can recompute:
GET https://ledger.horizonshield.dev/resume?endpoint=https://mcp.horizonshield.dev/mcp
Every line was measured by an independent outside witness, not by our own infrastructure. The verdict is a count of PASS and FAIL, never a star or a trust score. And each line is anchored to a Bitcoin block. Try to postdate a finding and it is refused by construction, because the block time belongs to Bitcoin, not to us. As I write this, the live line sits on Bitcoin block 965862.
The recompute recipe is embedded in the response itself. Take the bytes, hash them, verify the timestamp on Bitcoin. There is no step where you have to trust us.
There are two reference implementations, one in Python and one in Node, and they emit the same bytes down to the last character from the same input. Break one and the other notices. A fork that loosens the checks dies against the public adversarial harness we ship.
The data underneath follows the same rule. JCCDB, our open construction cost dataset, is 95,403 items under CC BY 4.0 with a DOI. Recompute the contents yourself.
To Japan's tech scene, and everyone else's
The frontier already moved past "does it run." The question now is "can you prove it."
The work coming out of Japan is some of the most careful and highest quality in the world. That is exactly why this hurts. While the scene competes on "amazing," it has not stepped onto the verification mat at all. Keep dodging the question and this work never earns a seat at the table where trust actually gets decided.
I refuse to let that happen to my own. So I build only in a form that can be proven.
If this landed, bookmark it and watch. The moment a second outside witness measures us, that track record becomes two lines on its own. And you will be able to check that with your own hands too.
Top comments (0)