Singapore Wrote the Best AI Rulebook on the Planet. One Word Is Missing.

Singapore's IMDA published the Model AI Governance Framework for Agentic AI in May 2026. It is the sharpest guidance any regulator has produced on autonomous AI.

One word is missing across 53 pages: assurance.

It has testing and monitoring, both run by organizations on themselves, plus internal audit by reviewers on the payroll of the team that shipped the agent. Independent attestation appears nowhere.

Four foundations are missing: assurance, redress, acceptance criteria, and liability allocation.

Fusion Collective filed feedback with IMDA on all four.

The Gap

An agent with credentials and write access changes the world while you look away.

Gartner forecasts over 40% of agentic AI projects canceled by end of 2027, naming costs, unclear value, and inadequate risk controls. Guess what’s missing? Model capability. Every named failure is a governance failure.

So, Singapore's answer is self-assessment. But when you read the case studies.

Dayos governs Dayos.

OCBC governs OCBC.

Every instance of "audit" means internal review, by a reviewer inside the same profit and loss statement that grows when the agent ships.

Testing is not assurance.

Monitoring is not assurance.

Assurance requires somebody who loses nothing by saying no.

Researchers settled this. Independence is core to third-party auditing, and scholars flagged the same flaw in the EU AI Act, which "relies primarily on internal conformity assessment.”

The auditor keeps turning out to be the vendor.

The Person Who Isn’t There

So, I’m reading and when I get to the end, guess who’s not there? The human on the receiving end.

The rejected candidate.

The declined patient.

The customer whose refund request is rejected/blocked by an agent at 2 AM.

They ain’t there.

"End users" means the employee running the agent or the customer chatting with it.

No contestability. No appeal. No redress.

The X0PA recruitment case study covers the recruiter's oversight and ignores the candidate's right to challenge the ranking.

This governs hiring, grants, and healthcare. GDPR Article 22 already grants people the right to contest automated decisions. Singapore's guidance sits behind law that existed for years.

Behind every percentage point is a person. This document governs the percentage point.

Three More Holes

No acceptance criteria. The Framework says test repeatedly, across varied datasets, in realistic environments. It never says how much is enough. No threshold. No pass mark. Every deployer gets a participation prize and then declares victory on its own judgment.

Accountability outsourced to lawyers. Five parties touch an agent. When one causes harm, the Framework says clarify obligations in contracts. Contracts move money after harm. They prevent nothing.

Collusion named, then abandoned. Section 1.2.3 describes pricing agents converging on higher prices without instruction. The Framework offers two multi-agent controls. Meanwhile the DOJ signaled criminal antitrust enforcement against algorithmic conduct in May and paid its first million-dollar whistleblower reward in January.

Collusion is the one risk self-assessment cannot catch. Your agent looks compliant alone. The offense lives between firms, where no deployer sees it.

The contradiction nobody flagged. The Framework cites Anthropic's finding that reasoning models don’t always say what they think, then builds human oversight on the reasoning chain anyway. Research says humans rubber-stamp plausible outputs without checking the reasoning. You can’t tell me the human will over-trust the machine, then turn around and blame that human for missing it.

Before You Say It

Auditors fail too.

On July 22, Australia's regulator wrote to roughly 3,000 registered auditors warning of tougher enforcement, after a scandal at KPMG Australia involving leaked client information and a whistleblower pressured to resign. ASIC Commissioner Kate O'Rourke said she’s concerned about auditors who failed to act "independently and ethically," and will pursue more court action.

Read that and you might decide independence is a fantasy, but I implore you to read it again.

Financial audit spent decades building registration, conflict-of-interest rules, whistleblower protection, and disciplinary bodies that can bar people from the profession. And with all of that, it still produced this. ASIC's chair says her powers are limited, because the firms are partnerships and she reaches only individuals.

AI assurance has none of that architecture.

Independence is the floor, not the ceiling. It has to be registered, policed, enforced. Singapore's Framework has not built the floor.

Critique Is Cheap. And That’s Why We File.

Sitting in the cheap seats and screaming potshots from the sidelines costs nothing and changes even less. Fusion Collective filed consultation responses with the European Commission on the EU AI Act, covering high-risk classification, the Article 6(1) conformity test, and Article 5 prohibited practices. One finding turned on a single phrase. The draft made "enhanced regulatory scrutiny" decisive in a classification test and defined it nowhere. Providers had to classify themselves against a standard that doesn’t even exist.

We filed comments with the US Federal Trade Commission.

We advise organizations working with the US Congress and UK Parliament. And we submitted feedback to IMDA on all four foundations, commercial interest declared in writing. Fusion Collective does not sell products or services to any client whose AI systems it evaluates. The auditor cannot be the vendor.

That declaration matters more than the critique. We sell independent assurance. We say so before anyone asks. An auditor who hides their incentives and don’t draw the line, already failed the standard they sell.

Rules improve when practitioners show up. Not louder. Better.

Three Questions

1. Who checks your checker, and what do they lose if the answer is no?
2. When your agent denies someone something that matters, how do they find out, and who can reverse it?
3. What score did your agent hit before you shipped, and was that number written down first?

Can’t answer all three? That’s documentation, not governance.

Bring us your hardest system. We will tell you the truth.