SCENE 01 · 0:00–0:06
Hook
Grade · fogVoice-overAsk the smartest machine on earth one question. How do you know? It can’t answer.
Nobody checked.
SCENE 02 · 0:06–0:18
The flood
Grade · fogVoice-overThis year the internet started talking to itself. Close to half a billion downloads a month of the protocol that lets machines call machines. More than a hundred and fifty organizations wiring their agents to talk to each other. Soon every fact you act on will arrive from a machine that heard it from a machine.
Nearly 500,000,000 a month
Downloads of the protocol that lets machines call machines — by its makers’ own count.
Self-reported 2026-07-28 · blog.modelcontextprotocol.io
150+ organizations
Wiring their agents to talk to each other — the consortium’s own count.
Self-reported 2026-04-09 · Linux Foundation
SCENE 03 · 0:18–0:28
The hole
Grade · fogVoice-overAnd here’s what nobody says out loud. There is no way — none — for one machine to ask another how strong its evidence is. No standard. No header. No vocabulary. We searched every specification we could find to prove ourselves wrong. We couldn’t.
There is no standard.
No way for one machine to ask another how strong its evidence is.
Verified 2026-09-02 · W3C · IANA · MCP · A2A
We tried to break this.
Every spec we could find. Nothing.
Verified 2026-09-02 · by adversarial falsification
SCENE 04 · 0:28–0:45
The receipt
Grade · turnVoice-overSo we built the receipt. Every claim in our gardens carries where it came from, when it was last true, and what nobody checked. One record. A person reads it as a card. A machine reads it as a contract. Ask only for what you can trust — and we tell you exactly what we held back, and why.
Every claim gets a receipt.
Where it came from. When it was last true. What nobody checked.
One record. Two readers.
A card for a person. A contract for a machine. Same truth.
withheld: 4 — here’s why
The first protocol that tells you what it did not give you.
Guess ours · specified, not yet shipped
SCENE 05 · 0:45–0:54
The machines do the work
Grade · goldVoice-overThen we let the machines do the checking. Thousands of claims, verified at machine speed. Trust that fades on a clock unless it’s re-earned. A garden that tends itself while we sleep.
Verified at machine speed.
Machines check. Humans decide.
Trust that expires.
Every claim fades on a clock unless it’s re-earned.
SCENE 06 · 0:54–1:13
The money
Grade · goldVoice-overEvery agent that needs a fact it can stand on pays the toll. Not once — every time. That’s not a product. That’s a lane. Payments got one. Messaging got one. The edge got one. Proof doesn’t. Yet.
EVERY AGENT. EVERY CLAIM. EVERY TIME.
The toll on truth.
Guess our forecast · no revenue claimed
PROTOCOLS FOR TALKING: 2
MCP and A2A. Protocols for proving: 0.
Verified both specs read in full · 2026-09-02
SCENE 07 · 1:13–1:31
The landscape
Grade · fog → goldVoice-overWho else is here? The academics built it, and nobody came. Google built a tag, then phased it out. The media crowd stamps pixels, not facts. The identity crowd proves who said it — not whether it’s true. Everyone is building agents that talk. Nobody is building the one thing they’ll need to say.
139 users.
The academics built the perfect version. Nobody came.
Verified 2026-09-02 · hub.knowledgepixels.com
Phased out by Google.
The fact-check tag is being retired from Google Search.
Verified 2025-06 · developers.google.com/search
Pixels, not facts. Who, not whether.
Media provenance and identity credentials stop short of the claim.
Verified C2PA 2.2 · W3C VC 2.0
| Category | Who | Where it stops short of the claim | Status |
|---|---|---|---|
| Agent protocols | MCP · A2A · AGNTCY | They carry the conversation and nothing about the evidence: neither schema has a field for provenance, confidence or verification. Trust attaches to the server, never to the content. | Verified |
| Identity & provenance | W3C VC 2.0 · C2PA · Web Bot Auth · Truepic | They prove who signed and which bytes moved. The VC 2.0 Recommendation says it in its own words: verifiability of a credential does not imply the truth of the claims encoded in it. | Verified |
| Scholarly claim graphs | Nanopublications · ORKG · Wikidata ranks · Semantic Scholar | The right object, built well, and no distribution — 10.9 million records from 139 users. Wikidata’s own docs call rank a filtering criterion, not a reliability judgement. | Verified |
| Fact-checking | ClaimReview · Full Fact · Factiverse · NewsGuard · Ground News | Human speed against machine volume, and Google is phasing ClaimReview out of Search. Ratings land on outlets, or on a blocklist of the provably false — never a graded tier on an arbitrary claim. | Verified |
| AI grounding & search | Perplexity · Exa · Tavily · Brave · Google Check Grounding | Retrieval and citation. The only scores shipped are relevance, or faithfulness to a set of facts you supplied and nobody verified. You can filter by domain and date; you cannot ask for evidence of a given strength. | Verified |
| Licensing & agent payments | Cloudflare pay-per-crawl · TollBit · x402 · AP2 | The toll booth, built and running. Payment buys access, and one flat price covers a whole domain: nothing in the rail says a word about the quality of what is behind the wall. | Verified |
| Enterprise KG & GraphRAG | Microsoft GraphRAG · Neo4j · Diffbot | Structure over a corpus. GraphRAG’s claim status is a language model’s TRUE / FALSE / SUSPECTED guess inside a private index, with no confidence, no evidence weighting and no way to expose it. | Verified |
The honesty check — what we found when we tried to find a competitor
We searched hard for anyone shipping evidence-strength negotiation, or claim-level trust receipts for agents, and found no company, protocol or draft that does it. The two nearest things alive are scite.ai and Letterprove, and neither is it. scite classifies 1.6 billion citations as supporting, contrasting or mentioning — but its unit is a DOI rather than a proposition, its label describes how a later paper treated an earlier paper rather than whether a claim is true, a case report counts exactly as much as a randomised trial, and there is no bar a consumer can ask for. Letterprove has the right product shape — signed, machine-readable attestations agents can cite — but it attests a vendor’s own usage metrics in one commercial vertical, self-reported by the seller, with no general vocabulary, no negotiation, and a waitlist where the API should be. The negative result is strongest on the handshake: the IANA HTTP Field Name Registry, last updated 2026-08-28, lists fourteen Accept-family fields, and not one of them negotiates evidence quality. Nobody has claimed this header.
Two candidates we could not resolve, tagged honestly rather than counted as absent: Trussed.ai served only navigation and a meta-description, and Golden and Grounded AI returned no primary product page establishing claim-level evidence tiers. Unverifiable
SCENE 08 · 1:31–2:00
The path
Grade · goldVoice-overThree steps. One record. One handshake. One meter. Humans and machines first — then machines and machines, and it multiplies. And the machines do the heavy lifting. Everyone else is teaching machines to talk. We’re the ones who can make them prove it.
1 record · 1 handshake · 1 meter
Then multiply.
Humans ⇆ machines. Then machines ⇆ machines.
The same fix, force-multiplied.
We make them prove it.
Record
Every claim carries its source, its date and its gaps.
Handshake
Ask for the evidence you need; get told what was withheld.
Meter
Every agent that needs a fact it can stand on pays for it.
Then multiply.
One loop lights. It splits into two, then four, then dozens — the same handshake, running everywhere, while we sleep.
Everyone else is teaching machines to talk. We’re the ones who can make them prove it.
The Receipt
The Knowledge Gardens · Think Tank KG