Verification capacity — mathematical, institutional, economic, epistemic — is the binding constraint of March 2026; capability has outrun it everywhere at once. | The Signal

The Pattern

A Fields Medal proof took decades of human effort to construct. An AI system called Gauss autoformalized it in two weeks, producing 200,000+ lines of verified logic and catching two errors that human reviewers missed. That is not an AI productivity story. That is a verification story.

The structural pattern underneath March 2026 is simple. The world's capacity to create has outrun its capacity to check. This is happening everywhere at once, across domains that don't normally talk to each other.

92.6% of developers now use AI coding assistants. 27% of AI-generated code ships without significant review. The outcomes are bimodal. Some organizations doubled their incident rates. Others halved them. Same tools. Different verification structures. The divergence is not about adoption. It is about whether anyone is watching what the machine produces.

Four distinct failure modes are running simultaneously. Constraint arbitrage in defense procurement, where OpenAI displaced Anthropic at the Pentagon not on capability but on willingness to comply. Coordination-cost inflation in economics, where ISM Manufacturing Prices hit 70.5, the highest since June 2022, while the Supreme Court struck down IEEPA tariffs and started a 150-day replacement clock. Validation deficits in technology, where XZ Utils-style supply chain attacks exposed that critical infrastructure runs on trust, not audit. And legitimacy erosion in institutions, where fabricated provenance and civilizational risk aversion are hollowing out the structures that used to do the checking.

The common force is speed asymmetry. Things that generate move faster than things that verify. If you run a team, this is your operating environment now. The question is not whether you adopted the tools. It is whether you built the checking layer before you turned the creation layer loose.

The Tension

The tension is between acceleration and accountability, and both sides have legitimate structural claims.

On the acceleration side, knowledge priming compresses AI coding fix burden from 45 minutes to 5. Seven context elements in versioned repo files. That is a 9x improvement in time-to-resolution. OpenAI deployed its first production model on Cerebras wafer-scale chips, bypassing Nvidia entirely. The infrastructure layer is unbundling. The generation capacity is not slowing down.

On the accountability side, nobody is building verification infrastructure at the same rate. Canada's Q4 2025 GDP contracted 0.6% on an annualized basis. First confirmed GDP-level tariff transmission. Iran struck Saudi Aramco's Ras Tanura refinery, taking 550,000 barrels per day offline and spiking Brent crude 10%. Spain refused US base access for Iran operations. The coordination costs of collective action are rising in every direction.

AI captured 61% of global venture capital in 2025. $258.7 billion. Five companies absorbed one-third of it. Seed-stage AI startups carry a 42% valuation premium just for the label. Capital is concentrating in generation. Not in verification. Not in governance. Not in the connective tissue that makes generation trustworthy.

For builders, this tension creates a specific architectural choice. You can invest in faster generation, where the competition is thickest and the capital concentration is highest. Or you can invest in verification capacity, where almost nobody is building and the structural demand is compounding daily. The CALM framework compressed a six-month security review into a two-hour deployment by treating architecture as code. That is verification infrastructure. The Commonhaus Foundation is building governance models for critical open-source projects. That is verification infrastructure. These are not glamorous. They are structurally scarce.

What This Unlocks

Paul Graham's Superlinear Returns thesis becomes measurably more true when AI eliminates the commodity cognitive layer. Compounding plus threshold effects. The returns at the problem-architecture layer steepen because the floor work disappears.

This creates clear winners and losers. Winners: builders who already had strong organizational design and now use AI to amplify it. The bimodal data from Fowler's analysis is not ambiguous. Organizations that halved their incidents had pre-existing structural quality. AI amplified what was already there. It did not create what was missing.

Losers: anyone whose value proposition was commodity cognition. Block laid off 40% of its workforce and shifted the Overton window for what counts as acceptable restructuring. That is not a hiring cycle. That is a category elimination. The roles that disappear are the ones that existed between generation and decision. The checking roles that were performed by humans slowly, inconsistently, and expensively.

What breaks is the middle. The middle-complexity work. The middle-skill roles. The middle-confidence planning horizons. Section 122's 150-day clock expires July 24, 2026. That means every supply chain decision made between now and then carries a hard expiration date on its assumptions. PMI expansion at 52.4 combined with prices at 2022 highs is a stagflation diagnostic. Growth and price pressure simultaneously. Planning horizons compress.

If you are building a company right now, the practical consequence is this. Stop building systems that depend on stable six-month assumptions. Build systems that can verify their own inputs and adjust. The organizations that halved their incidents did not have better AI tools. They had better feedback loops. That is what you should be building this week.

Watching Next

Three observables, all falsifiable.

First, the Section 122 replacement legislation. The 150-day clock creates a hard deadline. If Congress does not pass a replacement tariff framework by late July, the current tariff regime either expires or gets replaced by something worse. Watch for early drafts by mid-April. If none appear, the uncertainty premium on every cross-border supply chain decision doubles.

Second, the incident rate divergence in AI-assisted development. The bimodal pattern should become more pronounced over the next 90 days as adoption deepens. If organizations that doubled their incidents start publishing post-mortems, that is the verification market opening. You can check this in your own business. Look at your bug rate before and after AI tooling adoption. If it went up, you have a verification deficit. If it went down, your organizational structure was already sound. Either answer is useful.

Third, the Ras Tanura repair timeline. 550,000 barrels per day offline with Brent already elevated. If repairs take longer than three weeks, the energy price spike feeds directly into the ISM price index and makes the stagflation diagnostic worse. Watch diesel and shipping rates in your own cost structure.

Underweighting

I think I might be overweighting the novelty of the verification gap. Verification has always lagged generation. Printing press, telegraph, internet. Every communication technology created a temporary verification deficit that institutions eventually closed. The pattern I am describing might be cyclical, not structural.

The counter-argument is that AI-generated mathematics being formally verified, as the IEEE Spectrum piece documents, is precisely the kind of self-correcting mechanism that has always emerged. The system creates the problem and then creates the tool to solve it. If AI can verify Fields Medal proofs better than humans, it can verify code, contracts, and supply chain provenance. The verification gap might close faster than I am assuming.

I might also be imposing connections across domains that are thinner than they appear. The link between scholarly orthodoxy gaps in biblical academia and AI code review deficits is real at the structural level. Both are verification failures. But calling them the same phenomenon might overstate the coherence of what is actually a set of loosely related problems with different timelines and different solutions.

I think the verification thesis holds. The speed asymmetry is real. But I might be wrong about the timeline. Institutions are slow until they are not. Regulatory responses to the tariff chaos, the Lebanon Hezbollah ban, the Commonhaus governance model. These are all institutional verification responses already in motion. The gap might be closing faster than the data suggests.

Bottom Line

Capability is abundant. Verification is scarce. Every system you touch today, your code, your team structure, your planning assumptions, is operating in a world where creation outruns checking. The builders who win the next twelve months will not be the ones who generated the most. They will be the ones who knew what to trust.

Verification capacity — mathematical, institutional, economic, epistemic — is the binding constraint of March 2026; capability has outrun it everywhere at once.