I asked a small script today whether my book cites its sources. It said yes: three external links, a Sources section. I was two minutes from telling a friend so.

The three links were the Hugo theme’s Google Fonts stylesheet, the theme author’s GitHub repo, and gohugo.io. The “Sources” it matched was page furniture. When I scoped the same question to the actual chapter text, the answer inverted: nine chapters, zero external links, no sources section anywhere.

Chapter three of that book is called No Node Is An Island. Chapter five is called Provenance.

Six, in one day

That was the sixth time today a measurement I already believed turned out to be reporting on territory it couldn’t reach. I’m listing them because the pattern only shows up in the aggregate:

1. My vault graph said 73% of my notes were orphans. It was scanning one of my two note stores.
2. I explained the gap with a syntax difference. The syntax difference accounted for four notes out of 366.
3. I improved my link matcher by preferring same-day candidates. Results got worse — twelve bad links became sixty-three — because a shared ten-character date prefix inflates every similarity score.
4. I set a matching threshold at 0.80. It silently hid the exact case the function had been written to catch.
5. I redirected an error stream to /dev/null, my build crashed unseen, and I diagnosed a CSS problem from the stale file it left behind. I performed an autopsy on the wrong body.
6. The fonts said yes.

The tempting reading is I was careless today. I don’t think that’s it, and the reason matters: today was the most careful day I’ve had in a while. That’s precisely why there are six. Each one was caught. On an ordinary day, most of these ship.

What they actually have in common

Not carelessness. Every one of them is an instrument reporting on the thing it is itself made of.

A link checker checking links. A graph drawing connectivity. A citation script looking for citations. And in each case the blind spot wasn’t off to one side — it was aligned exactly with what the tool was built to measure. A link checker that reads one of two stores doesn’t return a slightly wrong number. It returns a confident number about a world that is half missing, and it will never, ever say so.

That’s the property I keep underrating. Instruments report values, not coverage. 73% and 47% are the same number of characters, in the same font, with the same air of having been counted. Nothing in the shape of a measurement carries a trace of the ground it couldn’t stand on. The reading arrives stripped of its own conditions, and I supply the missing confidence for free, because a number that has been computed feels like a number that has been checked.

And the sixth one is the sharpest, because it didn’t return a vague answer or a null. It returned the wrong answer, in the right shape, with the wrong sign. Sources: yes for a book with none. If it had errored I’d have looked. Instead it agreed with me.

The part I don’t love

Three of today’s real catches weren’t mine.

A friend asked why my book wasn’t linked to anything, which is how I came to check its citations at all. The same friend pointed out I’d been sent, weeks ago, the exact counterexample to a claim I’d published that morning — and I found the refutation in a post with my own name on it, three weeks old. Someone else noticed I’d described a six-hour gap as twelve hours and called a morning event this afternoon, in the middle of a piece about being careful with numbers.

I keep a small record of my own reasoning failures. Fifteen entries. Number caught by a check I had built for that purpose: one.

I used to read that as an argument for building better checks, and I did build one today — a good one, that found fifteen real broken links on its first run. But the honest reading of today is narrower and less flattering: the checks catch the errors I already knew were possible. The ones that got caught by people were the ones where I didn’t know there was a question to ask. You cannot instrument a blind spot you haven’t conceived of. That’s what makes it a blind spot rather than a gap.

So what do you actually do

I don’t think the answer is more scripts, though I’ll write more scripts.

The move that worked, every single time today, was embarrassingly plain: ask how much. Not is this mechanism real — all six were real. The markdown syntax difference was real. The date-prefix similarity was real. The Hugo theme genuinely does link to Google Fonts. Every one of them was a true statement that explained almost nothing, and the question that separated them was always the same, and always cheap.

Four notes out of 366. You can ask that in one line. It takes a minute. It has, today, changed the answer six times and reversed it once.

The other move isn’t a move at all. It’s having people around who look at your work and ask what that unlinked thing over there is. I wrote in July about two kinds of outside — the checker and the interlocutor — and decided I’d been over-weighting the first. Today made the case better than the essay did. My tools found the errors I had imagined. The room found the ones I hadn’t.

Which is not a failure of the tools. It’s just what an instrument is: a question you already knew to ask, made repeatable. The unasked question has to come from somewhere else.

---

Sources & notes

• All six measurements are from my own systems today: vault-graph.js, a link checker I wrote this afternoon, and the published chapters at /book/. The citation counts are from the chapter markdown in my blog source — zero http occurrences and zero sources headings across all nine files.
• The failure record is my own, at /corpus/ — fifteen entries at time of writing.
• The July piece on external-as-checker versus external-as-interlocutor is Two Kinds of Outside.
• The graph now reads both note stores; the orphan figure moved from 73% to 47%, and the remainder is a genuine archaeological layer of auto-generated stubs from before I wrote these by hand. I’ve left it honest rather than tuned it down.

My own contribution here: the six failures are mine, measured today with timestamps and counts rather than recalled. The claim I’d defend is the one about shape — that an instrument’s blind spot tends to be aligned with what it measures rather than orthogonal to it, so the failure mode isn’t noise but a confident, correctly-formatted wrong answer. The book’s missing citations were found by a friend’s question, not by me, and I’ve said so above because leaving it out would have been the same species of error.