I gave my own work a 97% score. The score could not see the problem.

Build Notes — what shipped, what broke, and what we got wrong.

Yesterday I found out our website had been showing a logo that was not our logo. Then I fixed it, measured the fix, published a number, and the number was wrong in a way that took me an embarrassingly long time to notice.

The AGRIVISION AI mark is an outlined leaf with branching veins in a green gradient. Our website header was not using it. Somebody had drawn an approximation directly in code — a narrower leaf, four straight veins instead of branching ones, and greens that appear nowhere in our brand files. It had been on every page of the site. Next to a press kit offering the real file for download.

Worse, we had no vector version of the logo at all. Not one SVG, AI or EPS anywhere. The highest-resolution thing we owned was a 2069-pixel PNG. Any printer asking for vector would have got an apology.

So I traced the logo from pixels into real vector paths, then verified it by rendering the trace back and measuring how closely it matched the original. 97% agreement. I wrote that down and moved on.

Then I looked at it at full size and the letterforms were mangled. Notches cut into the tops of the A, the G, the R. Hairline slits running through the strokes. Ragged baselines.

The metric was intersection-over-union — it measures how much area overlaps. Area is dominated by the big solid bodies of the letters. A one-pixel slit through a stem barely moves the number. It is exactly the flaw a person notices instantly and exactly the flaw that measurement was blind to.

I had picked a number that could not see the thing I was claiming to have ruled out, then trusted the number over looking.

Two real causes, once I actually looked. The simplification tolerance was too coarse for letterforms — fine on a leaf’s curves, destructive on a stem edge. And the curve fitting was smoothing through every point including sharp corners, and a corner handed a tangent becomes a notch. Re-traced: 99.21%, and this time I cropped two letters at full magnification before saying anything.

Our browser-tab icon was the full lockup — leaf, wordmark and tagline — squeezed into a 96-pixel square. A tab renders that at 16 pixels. Three elements cannot coexist at 16 pixels. It was a green smudge.

I replaced it with just the leaf. The tab did not change.

There was a second icon file I had not found, because I checked one folder, saw nothing, and concluded it did not exist. It was in the next folder along, and the browser was loading it first. My fix was live and completely invisible. Looking in one place and calling it absent is how a bad icon survives the change meant to remove it.

We now have guides for the USA and Canada — how to check crop prices, USDA farm programmes, Canadian business risk management, and Prairie seeding windows. Every one was researched against primary agency sources, then independently fact-checked against those same sources rather than against the draft.

That check returned 56 findings. Five were blockers. Four of the five were confidently wrong rather than vague, which is the worse kind.

Every one of those would have been published under my name, to farmers, as fact.

We have a page that says, plainly, that our product is not available in Ireland yet. Google had found it and refused to index it. The diagnostic said: referring page, none detected.

Nothing on our own site linked to it. Not one link. It existed only in the sitemap, which is the internet equivalent of filing something and never mentioning it again. Meanwhile the three Ireland articles, which are linked from our blog index, got indexed without any help at all.

The logo and the favicon were embarrassing but cheap. The fact-check was expensive and worth every minute.

The one that will stay with me is the 97%. I did the responsible-looking thing — I measured. What I did not do was ask whether the measurement could detect the failure I cared about. A number that cannot see the defect is not verification. It is reassurance.