This week's best technical write-ups said exactly what their numbers do not show
Three technical write-ups from the past week come from different worlds: a Java performance article, a medical study and an agentic code reviewer. They share a habit worth copying. Each one states what its number covers and, in the same breath, what it does not.
Per-Ake Minborg's article on Inside Java reports that scanning a flat array of Valhalla value objects was 1.47 times faster than an identity-object array on his machine. The same paragraph says the benchmark measured only traversal, not construction, decoding, CharSequence calls or toString, that it ran on one internal JDK 28 build, and that scalarization is an optimisation rather than a guarantee. A reader cannot mistake the result for a promise about their application.
Google Research's smartwatch study does the equivalent in medicine. It deliberately does not estimate blood pressure readings, because single values are too noisy, and claims something narrower: whether a person shows a persistent hypertensive state over several days. It then measures that claim against 24-hour ambulatory monitoring, the clinical reference, and publishes sensitivity next to specificity, so the 65.8 percent figure for initial screening is visible beside the 90 percent figure that makes the watch useful.
The Quarkus post by Mario Fusco builds the same discipline into software. A finding from a model reviewer is only accepted if it points at a file in the diff and a line near a change; a non-AI calibrator can make the verdict stricter but never more lenient; and the report writer is not allowed to produce findings or scores, which are rendered from data. The system is designed so that a model's output cannot claim more than the evidence behind it.
Cloudflare's agent post, also from last week, argues from the opposite side: when an intermediary hides who is acting, a website loses the ability to tie a claim to an accountable party. It is the same problem in a different setting, a statement whose scope nobody can check.

What it means
Engineering teams can borrow the pattern directly. Benchmarks should carry their limits in the same sentence as their results; health or safety signals should be validated against the best available reference; and generated output should be constrained by code so it cannot exceed what was measured.