← All AI Guides
#017 · OPERATING NOTE · the org · SEP 12, 2026

463 passing tests. A blank website. Both true at once.

Every test was green. Every page was empty. Here's how both can be true — and the one check that would have caught it.

The job I handed over

  • One of my products started serving a completely blank page — every route, on desktop and phone, nothing rendered at all. At the same time, its test suite was fully green: 463 tests passing. A healthy-looking project shipping a white screen.

What happened

  • So how does a site with 463 passing tests show people nothing? Because not one of those 463 tests ever looked at what actually appeared in a browser. They checked the functions, the data, the logic — all the parts — but never the one thing that mattered to a visitor: did a page actually show up. A security setting (a content-security-policy rule) was quietly blocking the app from drawing anything. The tests couldn't see it, because they were never pointed at the rendered page.

How I checked it

  • The agent proved it the simplest way possible — load the exact same URL two ways. With the broken setting on: 0 characters rendered. With it off: 1,763 characters. Same code, same 463 green tests, one blank and one fine. That little A/B turned an invisible bug into something undeniable.

What it took from me

  • Then the part worth keeping: a new check that actually opens the page in a browser and fails the build if nothing renders. 463 tests couldn't catch a blank site; one test that asks "did the page appear?" can. (A sibling case the same week: a different suite proudly reported "1 passed, 14 skipped" and called it success — it was treating skipped tests as if they'd passed. They hadn't.)

What I took from it

  • Tests that never touch the real, rendered result can't see a blank site. Green doesn't mean working — it means the things you happened to test passed. Add at least one check that loads the actual thing a person sees, after everything's wired together. And "skipped" is not "passed."

Try this

  1. Before you trust a green checkmark, ask one question: does any of this actually open the real thing and confirm it works — or does it only check the parts?
  2. If nothing loads the actual page, screen, or flow a person touches, add one check that does.
  3. Treat "skipped" as a failure to investigate, not a quiet pass.