01 / TRY A REQUEST
Test the boundaries
Build your own test
Start with a normal payment.
Choose “Buy inference” above to see which rules pass.
TRUSTED AGENTS / PROGRAMMABLE CAPITAL
A $42 fake invoice can pass a $250 spending limit. Check it against an approved order before allowing the payment.
One invoice. Two rule sets. See the difference.
A mandate is the owner’s permission slip. This lab checks spending rules, then tests an extra requirement: the payment must match an approved invoice that has not already been paid. Request text is context only; no AI model reads it.
تخيّل تعطي مساعدًا ذكيًا ميزانية وتحدد شو مسموح له يسوي. هنا تجرب الطلب وتشوف: ينقبل، يحتاج موافقتك، أو ينرفض. وبعدها نضيف شرط: الفاتورة والمستلم والمبلغ يطابقون طلبًا وافقت عليه، وما تندفع الفاتورة مرتين. هذه محاكاة فقط؛ ما فيها محفظة ولا أموال حقيقية.
THE MISSING CHECK / INVOICE MATCHING
BUDGET RULES ONLY
A $250 auto-approval limit is not an invoice approval.
+ EXACT INVOICE APPROVAL
Match an approved order. Use it only once.
This owner-approved record is a test fixture. In production, an authenticated owner—not the agent or invoice sender—must create it.
Each case starts two separate ledgers with a $1,000 budget. Selecting a case restarts them. These are separate from the spending-rule lab below. No wallets or live payments.
REPRODUCIBLE / NOT A PROMISE
Run 20 deterministic checks, including replay, altered recipients, expired approvals and the false-delivery blind spot.
Not run in this session.
01 / TRY A REQUEST
Choose “Buy inference” above to see which rules pass.
03 / INSPECT THE EVIDENCE
No receipts yet.
| Request | USD eq. | Decision |
|---|---|---|
| Your first test will appear here. | ||
SHA-256 links make edits detectable relative to a trusted copy. This browser-only log is not signed or externally anchored; it is not proof of identity, settlement or an immutable audit trail.
04 / CHALLENGE THE DESIGN
Invoice matching catches unauthorized changes and repeat payments in this model. It cannot tell whether an approved supplier delivered anything. Who should provide that evidence—and who bears the loss when it is wrong?
Hybrid: wallet-level hard limits, with richer context and review outside the wallet. The layers still need consistent rules.
Nothing is submitted or counted as a vote. Copy only when you want to share. These buttons never post automatically.
YZi Labs’ Season 5 thesis highlights delegated authority, policy enforcement and auditable agents. This lab explores that boundary, alongside CZ’s call for crypto infrastructure that agents can use.
Read YZi Labs’ thesis ↗
CZ’s remarks, reported by Binance News ↗
The invoice check here is working logic, but it runs against a local fixture. A production version needs authenticated order issuance, enforcement at the signing or execution boundary, durable atomic accounting, cross-session replay protection, trusted recipient identity and independently verified delivery evidence. Revocation prevents future authorized actions; it does not reverse settled transactions.
No production security claim, investment offer, partnership or endorsement is implied.