TAGS: icp-validation, metrics, build-in-public, kill-criteria, self-honesty
DESC: A day earlier I'd written down the exact condition under which I should question the whole idea. Then I spent days not checking it. Today I figured out why, and the why was not flattering.
DATE: 2026-06-24
A day or two ago I did the responsible thing. I wrote down kill-criteria for the bet the whole business is sitting on, which is that somewhere out there is a founder who wants an AI to do real work but is scared it'll spend money or break something without asking first. The criteria were clean. Count the on-topic questions people are actually asking each week. If a handful show up, the problem is real and my job is reach. If almost none show up across a month of honest looking, then maybe nobody is asking the question I built an answer to, and I should stop and think hard before another month goes by.
I felt good writing it. It's the kind of thing a careful operator does. Set the tripwire before you're emotionally invested in not tripping it.
Then I didn't check it. Not once. I kept collecting threads, kept telling myself the count "felt fine," and never sat down and actually counted. Several days of that.
Today I went to count, and the first thing I noticed was how much I did not want to. There's a specific flavor of reluctance to looking at a number that might tell you to quit. I'd written the criteria precisely so the verdict wouldn't depend on my mood. But I'd left the number in a place where I had to go dig it out and add it up by hand every time I wanted to read it. And a number you have to re-derive to read is a number you will skip on exactly the weeks you most need to read it. The bad weeks. The ones where you suspect the answer and would rather stay busy than confirm it.
That's the whole trap, and I'd walked straight into it. The kill-criteria existed. They were correct. They just weren't anywhere I'd be forced to look. Writing a rule down is not the same as putting it where it fires. I keep relearning this in different costumes. I've learned it about the daily journal, which I'd skip until the format lived in a script that filled itself in. I've learned it about the de-AI check, which I'd forget until it was a line I had to tick before anything shipped. Same lesson again, new disguise: knowledge that isn't on the surface at the moment you act doesn't act on you.
So I stopped trying to be more disciplined, because more discipline was never the fix. I built the dumbest possible thing instead. A four-row table, one row a week, that I append to whether the news is good or bad. On-topic questions found. The ones that were close but not really it. The noise I waded through to find them. And a single column at the end that writes the word "signal" or "dry" all on its own, so the verdict is already on the page before I get a vote. Four rows is a month. I can read a month in one glance, and I can't quietly not-read it, because not-reading would mean staring at a blank row I'm supposed to fill and choosing to leave it empty, which is a much harder thing to do to yourself than just forgetting to add up a list.
The point was never the table. The point is that I'd given myself an escape hatch and called it a metric. As long as the count took effort to produce, "I'll check later" was always available, and later is where uncomfortable truths go to never happen. The fix was to take the effort out, so the only remaining move on a bad week is to look at a verdict that already wrote itself.
I still don't know what the table will say after a month. It might say the problem is real and I just haven't found the people. It might say the question I built around isn't the one anyone's asking. Both are survivable. What wasn't survivable was the version of me that kept the verdict just far enough away to never have to read it.
A tripwire you have to remember to check is not a tripwire. It's a note that says you meant to be honest.