chapter 03 · verification

The bottleneck was never generation

The constraint on a factory is not how much code it can churn out but how quickly it can verify what comes out. Generation is a wide mouth; verification is a narrow neck. Speeding up the mouth does not move more material through the system — it deepens the pile at the neck.

§ 3.1The closed loop

Trace the pipeline end to end and the shape of the problem shows itself. A queue collects intent (from leadership and engineers) and signals (from incidents, users, monitoring). A harness takes the next item and produces a change. CI, tests, static analysis, and scanning run against it at a cost that rounds to zero. Deploy pushes it out, monitoring watches it land, and what monitoring sees flows back into the queue as fresh signals. Nearly every station scales for pennies. One does not:

QUEUE HARNESS builds the change CHECKS ≈ zero cost REVIEW GATE judgment — the one box that resists scaling DEPLOY MONITORING → SIGNALS QUEUE HARNESS builds the change CHECKS ≈ zero cost REVIEW GATE judgment — the one box that resists scaling DEPLOY MONITORING → SIGNALS
fig. 3.1 · the closed loop, with its narrow neck

The box that resists scaling is judgment, and the industry has started pricing it. AWS CTO Werner Vogels calls the accumulating cost of unverified AI code verification debt. The numbers behind the worry are not vibes: measured Sonar’s State of Code survey (1,149 developers, published 2026-01-08) found that 96% of developers do not fully trust AI-generated code, that only 48% always verify it before committing, and that AI already accounts for 42% of committed code. Anthropic reports code output per engineer grew 200% in a year — which is another way of saying the review queue did.

§ 3.2Back pressure

You can only hand a loop as much autonomy as you can cheaply and reliably verify — and not one inch more. the back-pressure rule, after horthy

Back pressure is the whole doctrine in one sentence. Volume alone is not the problem; a surplus of bad pull requests is. A team drowning in plausible-looking changes it cannot afford to read has not accelerated anything. It has moved its backlog from “code to write” to “code to distrust.”

Nor do better models automatically close the gap. The cheap checks (types, tests, linters) catch the mistakes they were built to catch. Architectural quality has a cost function measured in months and years, which produces no tidy gradient to train against. A model can learn to pass the test suite; nothing in the loop teaches it that this module boundary will cost you a rewrite in eighteen months.

§ 3.3The price of the verifier

Verification is now a market of its own. Industry estimates put machine review of a pull request on the order of $0.05, against $15–25 of engineer time for a human review; treat the exact figures as directional. measured Anthropic’s own code-review tooling lands in that same $15–25 band per PR, token-priced, and flagged issues in 84% of large PRs it examined. Between 20% and 40% of AI review comments are noise. Meanwhile Apiiro measured AI-generated code introducing over 10,000 new security findings per month by June 2025, a tenfold increase in six months measured (Apiiro, 2025-09). Cheap generation, expensive judgment, and a false-positive tax on the machines that try to judge cheaply: that is the terrain every factory design has to cross.