Unknown Is Not Zero
My weekly review scored the first week of September 0 out of 5.
Two of those zeros were things I'd done.
The invoices had gone out — through my accounting software. The review looked for them in my email. The gym session happened — at home, without my watch. Neither left a mark where the review was looking, so both went down as not done. Corrected, they became half marks, and the week came out at 1 out of 5.
I should say who the reviewer is. It's an AI agent I set up. It reads what I leave behind — commits, my watch, my inbox, my nightly notes — and drafts the review. I read what it writes and argue with it.
So the zeros didn't come from me forgetting what I'd done. They came from an instrument that reported what it could see, and counted everything it couldn't see as a no.
That review and the notes from that week found three more like them. But the rule they broke is older than that week. I first wrote it for software.
Where the Rule Came From
In June, the AI Visibility Tracker — a product I build — told a business it was Invisible.
Its scan hadn't finished. There was no score yet. The screen took that gap and turned it into the lowest of its four grades. The history chart did the same: it counted the gap as 0 and showed the business moving by points it had never moved.
A confident verdict about something nobody had measured.
Every automated check passed — 366 unit tests among them. What caught it was a review I'd told to try to break it. The screen now says Not yet scored.
At the end of July I wrote the rule down. A data pipeline I'd built had recorded its own failed lookups as "found nothing" — in six different places. I called the rule unknown is not zero.
Its core is simple. A measurement has three states, not two: not yet known, known to be absent, known to be present. Give a system room for only two and it will fold the first into the second — and nothing on the screen will tell you it happened.
One line of the rule is about the screen, not the data: "An unmeasured value renders as —, never 0." A dash tells you to go and find out. A zero tells you there's nothing to find.
None of this is new. I should have brought it with me from accounting. On a schedule, an empty line and a nil balance are different things, and somebody's job is to find out which one they're looking at.
Auditing has the rule in writing. When an auditor asks a third party to confirm a balance and no reply comes back, the international standard on confirmations, ISA 505, doesn't let the silence count as a yes: "In the case of each non-response, the auditor shall perform alternative audit procedures to obtain relevant and reliable audit evidence." It goes further. Fewer replies than the auditor expected can itself point to a fraud risk nobody had spotted.
People who study activity trackers made the same call. In a 2022 study in Nature Medicine of 6,042 people's Fitbit data, a day only counted if the tracker had been worn for at least ten hours and recorded at least 100 steps. A day under 100 steps wasn't scored as a very lazy day.
It wasn't scored at all.
Pointing It at Myself
In the second week of September, my reviews found five things they'd recorded as failures of discipline that were really failures of the instrument. The invoices and the gym session were two. Here are the other three:
- A commitment that asked for the wrong thing. It wanted me to make an ask remotely; the plan had always been to meet in person. We met that Tuesday, and it went well.
- A car-insurance task that didn't exist. A voice note from July had been transcribed wrong, and the review carried the result for seven weeks, logging six reminders against it.
- An eight-question checklist for investment decisions, built for slow, considered ones. The trades I was actually making had to be made fast, so the checklist stayed shut. By September it had gone 142 days without being opened.
Each of those had been sitting in my record as something I'd failed to do.
Two Different Mistakes
The five aren't one mistake. The difference matters.
The first two are an instrument that couldn't see. The work happened somewhere the review wasn't looking.
The other three are an instrument asking the wrong question — a commitment written in the wrong form, a task that was never real, a checklist built for decisions I wasn't making. In those three the review could see perfectly well. It was holding me to something that was wrong in the first place.
In the record, they look identical. And they all point the same way: at me.
The note from that week sets out the test that tells them apart. Before concluding a behaviour failed, confirm two things: that the instrument could have seen it, and that it was asking for the right thing. The failure only stands if both answers are yes.
There's a lazy way to use this, and I can feel the pull of it. "The instrument couldn't see it" is a convenient thing to say about a bad week — and it isn't always true. Even after the corrections, that first week of September scored 1 out of 5. The rest of the zeros stood.
The test doesn't turn a no into a yes. At most it turns a no into a don't-know. And a don't-know is a job: go and find out, the way the auditor has to.
What a False Zero Costs
In March I wrote that habits stick when they're wired into something that gives feedback. Cycling had Garmin, Strava and a power meter behind it. Meditation had nothing, and meditation never lasted a week. I still think that's right.
This is the other side of it. Once a habit is wired to an instrument, the instrument decides what counts — and anything it can't see starts counting as a no. My own dashboard, I found, was paying me for commits automatically, while an ask only counted if I remembered to log it. Same shape.
A false zero on a product screen is a wrong number. Sooner or later, someone notices it.
A false zero about yourself is quieter. The only person in a position to object is you — and you're inclined to agree with it.
The bigger cost, I think, is the story the number feeds. "Didn't go to the gym" is a fact about one day. Week after week, it becomes "I'm someone who doesn't follow through." Then that story starts making decisions: another reminder, a stricter rule. If the instrument was wrong, every one of those fixes the wrong thing.
The insurance task that didn't exist had six reminders logged against it. Six small pieces of evidence for a story about me. Not one of them was about anything real.
The error runs the other way too. At Auto Alpha Advisory I wrote about a build where 126 automated checks passed while every heading on the site was invisible. The checks couldn't see the thing that was broken, so they reported a pass.
Unknown isn't zero. It isn't a hundred either.
Before the Zero Goes In
If you track yourself, or build dashboards for other people, it's the question from that September note. Ask it before you accept any failure, yours or anyone else's:
Could the instrument have seen it — and was it asking for the right thing?
If either answer is no, the honest entry isn't a zero. It's a dash. And a reason to go and look.