Mirror Ai | September 9, 2026
We Wrote a 30-Question Regulatory Submission in 3 Days. We're Not in Singapore. We're Not Lawyers.
Three weeks ago, MAS published its consultation on stablecoin regulation — 86 pages, 30 questions, draft legislation in the annex. We had until 16 October.
We wrote our response in three days.
We're not in Singapore. We're not a law firm. We're not stablecoin issuers. We're a compliance intelligence company — two people, Brisbane and Sydney, with a structured database of regulatory obligations and an MCP API.
Here's what happened when we used our own tools to do the work.
The Submission
It runs 1,750 lines, covers all 30 questions, and engages with the draft Bill down to the subsection level. We cross-referenced Circle's 2023 submission, the GENIUS Act, MiCA, the FCA's PS26/10, and the Hong Kong Stablecoins Ordinance — with specific section numbers, dates, and legislative references. We compared the Singapore approach to the EU's, the UK's, and the US's, and made concrete recommendations for interoperability. We caught a structural error in the draft legislation where sections 100B–100E didn't actually exist in the Bill.
We did not know any of this by heart. We didn't need to.
The Backend Took Months. The Writing Took Days.
That inversion is the part worth explaining.
We spent months building the infrastructure — ingesting regulatory data, structuring obligations into a queryable format, wiring up an MCP server that could fetch primary sources on demand, and building the review pipeline that checks citations against their source documents. The database, the API, the attestation workflows — all of that took time.
The actual writing, the part that looks like output, took three days. Not because the model is magic (though it helps) but because we had already done the work of making our data accessible to it.
The bottleneck in regulatory responses isn't writing. It's finding, verifying, and structuring the source material. When the source material is already structured, the writing becomes the easy part.
The Honest Bit
We also got things wrong.
Our first draft referenced draft amendments that don't exist — sections 100B through 100E of the PS Act. They're not in the Bill. The actual provisions are 20A, 20B, 21B, and 21C. We claimed "over 20,000 obligations across 50+ jurisdictions" in our Mirror database — neither number was correct. Both classic AI hallucinations, confidently stated and completely fabricated.
We caught every single one. Not because the model suddenly became honest, but because we built a process that assumes it will fabricate. Every citation gets checked against the source. Every legislative reference gets verified against the actual PDF. The MCP tool that helped write the draft can also query our database, fetch the Bill text, and validate the claim — and when it can't, we kill the claim.
The hallucinations aren't the problem. They're expected. The inability to catch them is the problem.
The Part Nobody Warned Us About
The hardest part wasn't writing the answers. It was getting them into MAS's submission system — 30 separate FormSG text fields that strip tables, formatting, bullet points, and numbered lists. Everything we worked hard to structure got flattened on paste.
So we built a conversion pipeline that turned every table into prose, stripped all formatting, and produced a plaintext version that survives the FormSG paste. Because the submission infrastructure MAS uses can't accept the structured data we spent months building.
That gap — between what regulators ask for and what their systems can receive — is its own market signal.
What We Learned
Three things worth sharing:
The data is the bottleneck, not the writing. A regulatory submission is only as good as the source material feeding it. Investing in structured, machine-queryable data pays for itself the first time you need to write a response under a deadline.
Assume the machine will lie. Hallucinations are a feature of the technology, not a bug in your process. Build verification into the workflow, not as a separate step at the end. Every claim should be checkable from the same interface that generated it.
The submission infrastructure is a decade behind the regulation. Regulators write sophisticated frameworks and then ask you to paste plain text into a web form. That disconnect creates a real compliance cost — and an opportunity for the tools that bridge it.
Two people in Australia. Three days. One credible submission to the Monetary Authority of Singapore. Not because we're the smartest people in the room. Because we had tools that could do the heavy lifting, and we knew when not to trust them.
