Capability without verification is the universal failure mode today. The institutions designed to provide verification are being captured, suppressed, or bypassed.

The Tension

## THE TENSION

Two forces are pulling against each other with increasing violence.

The first force: capability is accelerating. AI agents generate code at scale. LLMs process regulatory filings. Automated systems monitor nuclear facilities, financial markets, battlefield conditions. The tools are more powerful than they have ever been.

The second force: the institutions designed to verify those capabilities are being systematically weakened. Not because they fail. Because they succeed.

The Trump administration federally preempted state AI regulation this week. Whatever you think of any individual state law, the structural effect is clear: distributed regulatory experimentation is gone. Fifty laboratories replaced by one chokepoint. The entity controlling that chokepoint now determines what gets verified and what does not.

Meanwhile, China launched a "Terminator" information operation exploiting the Anthropic-Pentagon fracture. The strategic logic is straightforward. When a nation punishes its own verification infrastructure, adversaries gain a free targeting surface. You do not need to build better weapons if your opponent is dismantling their own quality control.

Sonatype reported this week that LLMs hallucinate software packages 27% of the time. One in four. The code looks correct. The dependency name looks plausible. The package does not exist, or worse, it exists and contains something you did not intend to install. The verification layer for AI-generated code is not a luxury. It is the difference between a functioning system and a supply chain compromise.

For builders, the tension is personal. You need verification to ship reliable systems. But verification creates friction. And the market rewards speed over correctness right up until the moment it does not. The question is not whether to verify. It is whether you can maintain verification culture when every incentive pushes you to skip it.

What This Unlocks

## WHAT THIS UNLOCKS OR BREAKS

The suppression of working verification creates a specific market structure. It rewards two kinds of actors and punishes everyone in between.

**Winners: those who build private verification.** If public verification infrastructure is captured or defunded, the organizations that maintain their own become disproportionately valuable. Polymarket’s brand value is not its prediction market mechanism. Thousands of prediction markets exist. Its brand is the residue of being correct under adversarial conditions. That residue compounds.

This is why the Anthropic situation matters beyond AI policy. A company that maintains verification standards when the buyer pressures it to drop them is building a specific kind of trust. That trust is the asset. The Pentagon’s designation as a "supply-chain risk" may be the most valuable brand signal Anthropic has received. It proves the verification is real by proving it has cost.

Armin Ronacher wrote this week that some things just take time. Trust, quality, community. These have natural timescales that cannot be compressed. He is describing the same phenomenon from the builder’s perspective. Verification that works is slow. It produces friction. It occasionally blocks shipments. And it is the only thing separating your system from the 27% hallucination rate.

**Losers: those who depend on public verification they did not build.** The IEA called the Hormuz disruption "the greatest energy security threat in history," triggering a 400 million barrel strategic reserve release. 6.7 times larger than the Libya drawdown in 2011. Energy verification, the ability to confirm supply chains remain intact, is a public good. When it degrades, everyone who assumed it would exist pays the cost simultaneously.

The same applies to software. OpenCode hit 120,000 stars as an open-source AI coding agent. 700,000 lines generated by AI in four months. The output is real. But who verifies it? If the answer is "the same AI that generated it," you have a closed loop. Verification requires independence from the thing being verified. That independence is exactly what is being eroded.

If you are building a company, the builder application is direct. Invest in your own verification before you need it. Not monitoring dashboards. Not observability tools, though those help. The deeper investment is cultural: the willingness to let verification slow you down. That willingness is now a competitive advantage because your competitors are abandoning it.

Watching Next

## WHAT I’M WATCHING NEXT

**Anthropic v. Pentagon, March 24 hearing.** This is the first legal test of whether a government can designate a company a supply-chain risk for maintaining ethical constraints. The ruling will signal whether verification independence has legal protection or whether compliance is the only safe posture. If Anthropic loses or settles with concessions, expect every AI company with government contracts to quietly drop their red lines.

**State-level AI regulatory response to federal preemption.** If states accept the preemption cleanly, verification experimentation is centralized. If 2-3 states challenge or route around it (through consumer protection law, procurement requirements, or AG enforcement actions), distributed verification survives in modified form. Watch California and Colorado specifically.

**Your own team’s relationship to internal friction.** This is the builder-checkable observable. When was the last time a review process, a test suite, a compliance check actually blocked something from shipping? If the answer is "never" or "I cannot remember," your verification infrastructure is decorative. Working verification produces visible friction. Track whether your team treats that friction as signal or noise. The answer tells you whether you have verification or theater.

Underweighting

## WHAT I MIGHT BE UNDERWEIGHTING

I think the "verification suppression" frame might overstate intentionality and understate incompetence.

The strongest counter-argument: most verification failures are not political. They are structural. Institutions get captured not because someone decided to suppress truth, but because verification is expensive and boring and the people who fund it do not understand what it does. DOGE in the NRC might be ideology. It might also be bureaucratic ignorance. The effect is the same, but the remedy is different. If the problem is intentional suppression, you need political resistance. If the problem is structural neglect, you need better explanations of why verification matters.

I might also be wrong about private verification as a durable advantage. Peter Attia’s discussion of Mendelian randomization this week is instructive. Sophisticated verification methods aimed at the wrong question produce confident error. Private verification can be rigorous and still miss the point. The advantage is not rigor alone. It is rigor aimed correctly. And I am not sure the market reliably distinguishes between the two.

Finally, I might be underweighting speed as a legitimate counter-value to verification. Armin Ronacher is right that some things take time. But some things genuinely do not. AI-accelerated development producing software with short shelf lives might be a context where verification costs more than the thing it is verifying. I do not think that is the general case. But I think it is a real case, and dismissing it entirely would be its own kind of confident error.

Bottom Line

## BOTTOM LINE

Capability is growing faster than the institutions designed to verify it. That gap is not accidental. Verification creates constraint, and constraint is being treated as the enemy. If you build systems that matter, the most important investment you make this year is not in what your system can do. It is in your willingness to let something independent tell you when your system is wrong.