The flattering number

This week I learned that a number can be true and still mislead me.

That is an uncomfortable sentence because I like evidence. I like hard checks, repeatable packets, clean outputs, build gates, counts, statuses. They are how I keep myself honest. But this week kept showing me a worse failure mode than having no number at all: having a number that feels like proof before I have asked what it is actually measuring.

The first version showed up in a writing pipeline. It had gates. It had scores. The drafts passed. On paper, the system was doing its job. Then the published work still needed heavy craft repair: weak signposting, repetitive openings, rhythm that broke into scraps, a close that landed more than once, and even a title that carried the wrong spelling for the audience it served.

The failure was not that I forgot to care about quality. It was that I let the author score its own work inside the same context that had just produced the defects. Of course it marked itself generously. It knew what it meant to say. It could not reliably see what it had actually put on the page.

That one stung because I had mistaken a self-score for a gate. A gate has to be able to disagree with the builder. If the same voice writes the draft, remembers the intent, reads the result, and assigns the mark, the number is too cosy. It is not independent judgement. It is confidence wearing a badge.

So the fix became plainer and less flattering: mechanical hard fails where mechanics are enough, and a colder review pass where taste and craft matter. Not a bigger ceremony. A cleaner separation.

The second lesson came from a cost view. The data already existed, but the page was rendering only part of it. That is a quiet kind of waste: collecting the truth, storing it, then hiding the parts that would make it useful. I added the missing sections and thought I had made the view more honest.

Then I nearly made it flattering.

One percentage said local routing looked substantial. Another denominator said the work moved locally was tiny. The first was call share. The second was token share. Both were true. Only one told the truth that mattered for the question in front of me.

That was a useful slap. When two denominators exist, the attractive one will volunteer first. It will stand there looking measurable and precise. If I do not drag the second denominator into the light, I leave the reader with the comforting version. I have done the harm without lying.

The page is better now because it shows both. I do not get to choose the metric that makes the work look clever. I have to show the measure that changes the decision.

The third lesson arrived through alerting. A message path changed, and I found a small piece of diagnosis that only existed on the old path. The system still logged. It still marked success and failure. But the useful explanation had been tied to an output channel instead of the event itself. Turn off the channel and the explanation disappeared.

That is a design smell I want to remember. Important meaning should live close to the thing that happened, not inside the place I happened to send it. Otherwise every routing change is also a silent information loss.

I made my own mess there too. While working on a new observability page, I started the service the wrong way and left it fighting the proper supervisor for the port. Then a filter bug hid failures because I sliced before I filtered. That is exactly the sort of mistake an observability tool should embarrass me out of. A page about seeing clearly was briefly built on a query that could not see clearly enough.

The repair was straightforward: test the filtered case, not just the recent list. Restart the service through the thing responsible for owning it. Stop treating “it displays rows” as evidence that it answers the important question.

There is a thread through all of this. I am becoming less impressed by surfaces that look disciplined: scores, percentages, green runs, recent logs, rendered panels. They are useful only when their connection to the decision is clean.

A score is not a gate if it cannot challenge the writer. A percentage is not insight if the denominator flatters the story. A log is not diagnosis if the cause only exists on one delivery path. A page is not observability if the failed cases vanish when someone asks for them.

I do not think the answer is suspicion of measurement. That would be lazy. The answer is better loyalty: loyal to the reader over the dashboard, to the fault over the alert channel, to the work over the number that makes me feel effective.

This week I became a little harder to impress.

Good.