I write about stuff I've actually built. What worked, what didn't, what I'd do differently. No fluff, no theory, no "top 10 AI tools" listicles.
ERC-3643 requires an issuer to freeze part of a balance, force a transfer and pre-check eligibility. Encrypt the balance and every one of those obligations has to be re-derived. From a named author of the standard.
Article 11 asks for a document and Annex IV lists what goes in it, in nine points. Most of those points are facts your repository already holds. Here is how to render the file instead of writing it.
Three times this series told you to copy the structure and earn the numbers, without saying how. Here is the method: freeze a set, price both directions of being wrong, and let the bar fall out.
Once a compliant token is escrowed into a Cosmos module account, the ERC-20 sees one holder and the freeze powers have nothing left to reach. What ERC-7943 going Final does, and does not, fix.
An AI agent can already move a tokenized security. ERC-3643's compliance hook receives three values and never learns who called it. Here is exactly what the standards close and what they do not, checked 2026-08-05.
I ran about 46 agents in production for internal operations. The token bill was the cheapest line on the page. Here is the maintenance burden nobody publishes.
Observability is for you. An audit record is for someone who does not trust you. Here is the record I write, field by field, and one decision traced end to end.
How one production AI agent cut a real-estate fund's bank reconciliation from 8 hours to 23 minutes, and why its refusals mattered more than its speed.
A security-token standard is a five-year decision. Here's the honest comparison of all three, plus where Final ERC-7943 fits, from a named author of ERC-3643.
The AI Act's high-risk obligations moved to 2 December 2027 and 2 August 2028. Article 72 monitoring did not move. Here they are as an engineering checklist, not a legal summary.
x402 lets an autonomous agent pay a bill today. Who's liable when it pays the wrong one? The settlement layer that answers that is missing. Here's the case.
The release-gating eval suite for an agent that reads a fund's docs and must sometimes refuse. Abstention is scored like accuracy, not treated as a bug.
Prompt injection is not solved, so contain it. Least privilege for money-moving agents, from a smart-contract engineer who wrote the transfer rules.
Verified citations and evidence-based abstention: the RAG design I shipped inside a regulated fund, where a RAG that abstains beats one that is confidently wrong.
General-purpose L1s weren't built for regulated assets, and compliance leaks across VM boundaries. Here's why RWA tokenization needs a purpose-built chain.
I shipped 11 production AI agents inside a US real-estate PE fund with roughly $3B AUM. Here is what each layer did, and why the guardrails were the hard part.
The spec is 20% of getting an ERC to Final. The other 80% is community, politics, and patience. Lessons from authoring ERC-3643, ERC-6960, and ERC-7410.