The numbers that decide whether work ships

Since 12 March 2024, Interaction to Next Paint (INP) has been a Core Web Vital, replacing First Input Delay. FID only measured the wait before the first interaction was handled. INP observes every click, tap and key press in a visit and reports one of the slowest, so a page that feels quick on arrival and sluggish afterwards can no longer hide it.

The usual culprits are long tasks on the main thread: a tag manager loading a dozen scripts, a consent banner that blocks input while it decides, a carousel that re-renders on every tap. Breaking work into smaller tasks, and deferring whatever the first interaction does not need, moves INP more than a faster server does.

What we check: field data, not only a lab score. A page can pass Lighthouse on a fast laptop and fail INP on the mid-range phones most of its visitors use. We read the 75th percentile from real-user data before a launch is signed off, and we treat a third-party script that pushes INP past 200 ms as a defect, not a trade-off.

Gmail’s guidelines for anyone sending more than 5,000 messages a day to Gmail accounts set three conditions. Authenticate the domain with SPF, DKIM and DMARC. Support one-click unsubscribe, with a visible unsubscribe link in the message body. And keep the spam rate reported in Postmaster Tools below 0.3 per cent, ideally below 0.1.

What we check: that every marketing message carries the List-Unsubscribe and List-Unsubscribe-Post headers described in RFC 8058, that an unsubscribe takes effect without a login, and that suppression runs before a send rather than after it. This newsletter works the same way: double opt-in, an unsubscribe link in every issue, and the one-click headers for the mail apps that support them.

The 2025 edition of the OWASP Top 10 for LLM Applications puts prompt injection first, as LLM01, and lists Excessive Agency as LLM06: an AI system given more functions, permissions or autonomy than its task needs. Neither is fixed by a better prompt. Both are tested for.

• Write the evaluation set before building the agent, including attempts to slip instructions in through documents and tool results.
• Agree the pass mark in advance, and what happens below it, so that a good demo cannot move it.
• Give each tool the narrowest permission it needs, and put a person in front of anything that cannot be undone.

In practice that means a small, versioned set of test cases for each behaviour: answers the agent must give, actions it must refuse, data it must never reveal. The set runs on every change to the prompt, the model or the tools, and the results are kept with the release.

What we check: that the threshold was written down before anyone saw a result. A pass rate chosen after the fact is a description, not a test.

WCAG 2.2, a W3C Recommendation since 5 October 2023, added nine success criteria. One of them, Target Size (Minimum) at level AA, asks for pointer targets of at least 24 by 24 CSS pixels, with exceptions such as enough space around a smaller target. The contrast minimum is unchanged: 4.5:1 for body text and 3:1 for large text.

Two more of the new criteria are worth a look on any sign-up or checkout. Focus Not Obscured (Minimum) stops a sticky header or a cookie banner from hiding the field a keyboard user is on. Accessible Authentication (Minimum) means a log-in must not depend on remembering or transcribing something unless there is help, such as letting a password manager fill the field.

What we check: the smallest things on the page, such as close buttons, pagination and links packed into tables, on a phone, with a thumb. We design touch targets at 44 pixels, which clears the minimum with room to spare.