Shift left assumed humans wrote every line

GuruPrakash SampathKumar · 11 September 2026 ·Testing what a model wrote

A model writes code faster than most teams can test it. That is not a complaint about the quality of the code. It is a statement about volume.

Shift left was a good answer to a real problem. Find a defect early, because it costs more the later you find it. It was designed for a world where a person wrote every line, and where the rate lines arrived was the rate people could type them.

That is no longer the constraint. Pushing testing earlier does not help much when the thing you are trying to keep up with is generation itself.

What holds instead

The answer is not an earlier test. It is a check that runs on every change and cannot be skipped or argued with. Testing before release and watching after release stop being two phases and become one loop.

This site is a small example of it. npm run build runs five things in order, and any one of them failing stops the build.

astro check type checks the components and the templates. A content entry that breaks its schema fails too, so a note that carries a picture with no description cannot be published here at all.

scripts/contrast.mjs reads the stylesheet and checks the low end of every font size clamp and every text colour pair. The floor is 7

rather than the 4.5 the standard asks for, because we want this readable by people in their sixties and seventies. A 13px clamp floor fails. So does a 6.4 ratio. No reviewer catches either by eye.

scripts/host.mjs cross-checks the host named in src/lib/site.ts against the adapter in astro.config.mjs. A half-finished move from one host to another fails the build instead of deploying.

Then the site builds, and scripts/weight.mjs weighs every page. The budget is 250,000 bytes uncompressed, fonts counted on every page, and a srcset measured at its largest candidate rather than its first, because a budget is a worst case.

None of that is clever. It is four scripts and a type checker. What makes it work is that it runs every time, on a machine, and the answer is pass or fail.

What the gates do not do

They do not know whether a sentence is true. They do not know whether a page was worth publishing. They will catch a font size and a byte count, and they will never catch the paragraph that is confidently wrong.

That part is still a person reading it, slowly.

For what happens once something is live, how we work puts it plainly: we keep it running and we tell you when it stops being right.

All notes