Failure catalogue · The wrapper betrays the numbers · 45 of 51
OBSERVED FAILURE MODE
The manifest that contradicts its page.
The manifest, the handoff summary and the README are shipped artifacts, and readers treat their figures as facts. A hand-typed count that disagrees with the machine output beside it, says 3 where the page shows 6, is a failed build rather than a typo. The reader has no way to tell which of the two surfaces is lying.
What we saw
In our own audits a summary carried counts typed by hand while the machine artifacts beside it carried the computed ones. The summary said 3, the page showed 6. Both shipped inside the same pack, both looked authoritative, and nothing on either surface told the reader which to believe. Every figure on the page itself had been derived and gated. The one number a busy reader was most likely to quote onward, the one in the summary, was the only number nobody had computed.
Why it passes a glance
Prose gets reviewed for sense, not recomputed. A count inside a sentence looks like a description of the work rather than an output of it, so it escapes every gate aimed at the page. It was also true once, which is the harder half: numbers in prose rot in place while the build moves on, and nothing about a stale figure looks stale.
What addresses it
The verification-practice skill states it plainly: any number a document asserts about the system will silently rot in several places, so either generate it or write the guard that scans every document and fails naming file and line. Doctrine principle 1, An AI never authors a number, extends to every shipped artifact, and the crudest guard is also the fastest: grep the prose for digits that appear nowhere in the machine output.
Check your own file in two minutes
- Grep every prose artifact in the pack for digits and check each hit against the machine output.
- Replace any surviving hand-typed figure with a value the build injects.
- Change one count in the data, rebuild, and confirm every surface quoting it moves with it.
What this does not catch
A grep guard catches figures the machine output does not contain. It cannot catch a number that matches something in the artifacts by coincidence, and it says nothing about a correct figure described in the wrong words. Expect it to catch something on its first run.
Quick answers
- Is a hand-typed count really a build failure?
- Yes. The reader cannot tell which surface to believe, and the summary holds the number most likely to be quoted onward. A prose figure absent from the machine output fails the gate by grep, naming the file and the line.
- What counts as a shipped artifact?
- Anything the customer can open: the page, the manifest, the handoff summary, the README, even a code comment they will read. The rule follows the reader rather than the file extension.
- Will the guard produce false positives?
- Some. A version number or a year needs an exemption you write down. Expect the guard to catch something on its very first run, which is the guard working rather than the guard misfiring.
Nearby failures
Last updated 2026-09-02 · Dylan, founder · one of 51 observed failure modes, every one seen in a real build or in our own audits, none invented.