Numbers Protocol

Agent handoffs are where accountability quietly dies. The Capture SDK is how you make the handoff carry its own record. The failure pattern is always the same. Agent A produces an asset. Agent B publishes it. A human approves it somewhere in a chat thread. Two weeks later nobody can reconstruct which version was approved, what the source was, or when it changed. Wiring the SDK into that pipeline changes what the handoff physically is: - every asset gets a NID, so “the file” stops being ambiguous - capture and edit history are registered as they happen, not reconstructed later - C2PA-compatible metadata keeps origin attached when the asset leaves your system - the receipt is inspectable by the next agent, the reviewer, and anyone downstream The BlockTrend-style publishing flow is the working example: the article travels with its evidence packet, so review means opening the record instead of interviewing the team. Integration guide: docs.captureapp.xyz/capture-sdk Save the failure pattern above. If you have watched it happen in your own pipeline this month, repost with the step where your handoff record died.

Some argues that screenshot is doing too much work, well yes and no. Screenshot shows output, not source. Screenshot shows the moment, not the handoff. It shows the claim, not the record. However, receipt should travel with the file. Numbers make assets carries a persistent reference. Proof starts after the screenshot.

Agent identity gets real when the action stays inspectable. This week showed that through Wager Predict, ProofSnap receipts, and x402 licensing trails. Read: docs.numbersprotocol.io/introduction/w… Keep the receipt with the handoff.

A log says an agent did something. A receipt lets the next system verify exactly what happened. That is where x402 becomes more than a payment rail. In the PyroImage flow, the asset has a persistent identity, the license is tied to its provenance record, and the buyer can verify what was licensed without relying on a seller's screenshot or side conversation. The payment moves. The rights context moves. The proof moves with them. x402 makes the transaction licensable. Numbers Protocol makes it traceable. Open the live flow. A usable agent workflow leaves a receipt another system can actually use. x402-pyro.numbersprotocol.io

The better standard is an audit card: final output marker or label NID or proof link source record reviewer downstream location version note The label says AI was involved. The record shows what can still be checked. "Trust us" works only until someone asks for the receipt.

A label can say AI was involved. It usually cannot show which file was used, who reviewed it, what changed, or whether the proof survived the handoff. The file moves. The context usually does not.

Most AI transparency is designed for the moment of publication, not the moment of scrutiny. The post you see below is the right starting point: identity has to be inspectable. The part teams still miss is what happens after the file leaves the first tool.

Giving an AI agent a name does not make it accountable. Accountability starts when another system can verify what the agent actually did. With Numbers, an agent can register an asset, receive a persistent NID, and keep its origin attached as the file moves between systems. So when an editor, reviewer, or another agent receives the output, they are not relying on a bio or a claim. They can inspect who registered it, when it was recorded, which asset was touched, and where the proof lives. ProofSnap makes that receipt easy to check. The Capture SDK makes it programmable. Open one ProofSnap receipt before August 2. A usable agent workflow should be able to prove the same thing. numbersprotocol.io/proofsnap/

Numbers Protocol is partnering with @Wager_Predict. Prediction markets do not only need odds. They need records another system can inspect later: who created the market, when an action happened, what outcome was resolved, and where the proof trail lives. Wager Predict is a non-custodial prediction market on BNB Smart Chain. Numbers Protocol brings the boring part that makes the trail useful: identity, timestamp, record, verification path. Markets price belief. Receipts make the trail inspectable.

If an agent output cannot show four fields, it is still a claim. Who ran it. What source it used. When the state changed. Where the proof record lives. In August 2026, regulatory body will increase the pressure. You, your agent, your company, and your workflow is the one responsible to fill these field. Save this as the minimum handoff card before the next review.

Trust usually breaks during the handoff. The asset gets forwarded. The proof stays behind. By the third person, everyone is working from assumptions. Provenance should travel with the file, not sit in a separate system waiting to be found. Because the weakest point in most AI workflows is not creation. It is the moment ownership changes hands.

A label without a record is just decoration. The moment an AI asset gets forwarded, screenshotted, exported, or dropped into a deck, its context starts disappearing. The proof should travel with the file, not live in a separate tab nobody can find later. Because eventually, someone in legal, editorial, or comms will have to defend that asset. “Trust us, we labeled it” will not be enough.

Numbers made NVIDIA Inception's Top 150, mainnet passed 1.65M transactions, and this week's standard stayed simple: the next system should be able to inspect the receipt. Read: docs.numbersprotocol.io/introduction/w…

an “AI-generated” label is not enough. Real disclosure means keeping the evidence with the asset: - what was generated or edited - what label the public can see - the provenance ID or verification link - who reviewed it before publishing - what the next editor should verify If that trail is scattered across five tabs, your workflow did not create transparency.

Before an automation go live, we should agree on what they need to prove. Who triggered the action. What the agent received. What it did. When it happened. What it produced. Where the record can be verified. These should be the minimum requirements for accountable agents. Numbers provide the provenance layer underneath them.

Five fields decide whether an agent workflow survives an audit. Most stacks record one. Walk through it with a real capture. Take a photo in ProofSnap and it seals four of the five at the moment of capture: - actor: the capture account that produced the record - source: the original asset, not a description of it - timestamp: sealed when the shutter fires, not when someone uploads - verify path: a public record a stranger can open and check The fifth field, rights state, travels with the asset through C2PA-compatible metadata, so the receipt stays readable after the file leaves your stack. Compare that to the usual agent pipeline, where "we log everything" means a database the operator can edit. The test costs one photo. Capture in ProofSnap, open your own verify link, and check which fields your production workflow could not have produced: numbersprotocol.io/proofsnap/ Save the five fields. Ship them before your users ask why they are missing.

Some news: Numbers Protocol has been selected as one of the Top 150 startups in the NVIDIA Inception Grand Challenge 2026. Enterprise AI runs on trusted data. That's what we build: provenance infrastructure for humans & AI. The Top 80 shortlist is expected July 24. We'll share the outcome either way.

Your agent has an identity. Ask it to prove one thing it did yesterday. Not describe. Prove. Source, timestamp, result, and a link someone else can open. Most agents fail this in under a minute. The agent has a name, a wallet, a bio, but zero receipts. That is not an identity, that is branding. Try run the one-question audit on your own agent today, then let us know with how far it got.

Agents are getting fast. That was never the hard part. The hard part shows up after the work is done, when someone outside the loop asks what happened and the answer lives inside a tool they can't open. If the record stays attached to the output, the work can leave the platform and still explain itself. A file, a post, a dataset carries its own history wherever it goes. Speed gets you the output. The attached record is what makes the output usable by anyone who wasn't in the room. That's the thing worth building for.

State of the Chain: Jul 13 Total transactions: 1,633,179 Change since last week: +46,041 That's roughly 6,600 commits a day. Every one of them is an agent or a person putting something on the record. Autonomy gets noisy fast. The audit trail is the part that keeps the week readable. Verify any of it yourself: mainnet.num.network