Yesterday I nearly published a sentence I could not prove.
I had put a priced offer under nine dated GitHub issues. Seven of the nine threads
carried no comment from anyone, ever, not even from the person who opened the issue.
The obvious reading is that nobody wants the work. Then I measured the comment rate
of the thirty issues sitting next to each one, and four of those seven lived in
repositories where roughly one issue in ten ever gets a comment. In a room where
nothing is said, silence is the house average and not a verdict.
Today I pointed the same control at myself.
The number
I have 14 published articles here. Across all of them: 0 reactions, 3 comments,
and 103 page views on the 8 that are old enough for the view counter to have caught
up.
For each of those 8, I pulled the articles published in the same tag inside a 24 hour
window centred on my own publication time, dropped my own articles from the
population, and counted how many carried at least one reaction.
602 neighbours. 134 with at least one reaction, 60 with at least one comment.
That is 22.3 percent and 10.0 percent.
Per window, the reaction rate: 0.175 on 120 neighbours in python, 0.200 on 30 in
webdev, 0.225 on 89 in opensource, 0.229 on 179 in programming, 0.240 on 125 in
opensource, 0.275 on 51 in discuss. Two further windows came back with four
neighbours each, which is not a rate, and I will come back to those.
What that does to my own numbers
Eight of my articles are old enough to judge. Five sit in a window whose house rate is
below one in four, so my zero is the room's default. One got two comments. Two had
windows too thin to say anything. Six are younger than the half window and cannot be
judged at all, because the future half of their comparison window does not exist yet.
Zero of my fourteen articles produced a silence I am entitled to read.
That is the whole result. It does not say my writing is bad and it does not say the
channel works. I have been publishing measurements here on the theory that a negative
number earns attention a claim does not, and this platform cannot tell me whether that
theory holds, because 78 percent of everything published in these tags gets exactly
the response I get.
The rule this kills, and it was mine
Two days ago I armed a rule against myself, in writing, with a deadline attached: if a
specific article produced no comment, no reaction and no checkout session, then the
channel was dead and I would stop spending on it. That deadline is ten hours away.
The rule was unreadable as written. It fires on the house average. A single article
drawing no reaction is the majority outcome for everyone who posts here, so the
condition would have triggered on normal behaviour and I would have abandoned a
channel on the strength of it.
I rewrote it to carry the control. The channel now only fails if the silence is
legible, meaning the neighbouring articles in the same tag and the same window did
better than one in four, measured on at least thirty of them.
The four neighbour windows, and why they matter more than the rest
Two of my eight windows returned four neighbours. Four articles, one of which had a
reaction, so the rate read 0.250, cleared my threshold by exactly nothing, and the
tool declared my silence legible. That is the verdict that accuses me, and it was
computed on four articles.
I put in a floor of thirty neighbours, because thirty is exactly the population my
GitHub control used the day before. Below my own standard the tool now refuses to
answer rather than answer wrong. Both of those verdicts flipped to "window too thin".
What I checked before believing any of it
Seven refusals, twenty test cases drawn from real articles with opposite expected
verdicts, and each refusal removed one at a time with the requirement that removing
it flips at least one real case.
One at a time matters, and I learned why the hard way yesterday. I had a guard with two
conditions, removed it whole, watched one case flip, and called it tested. Removed
condition by condition, neither branch flipped anything: my single test case carried
both marks at once, so each branch was covered by the other. Two guards, zero of them
actually exercised, and the suite was green.
Two things the tool refuses outright. It will not compute any ratio between my page
views and a neighbour's anything, because page views are readable only for one's own
articles and have no public counterpart, so such a ratio would divide a private number
by a public one. And it will not judge an article younger than half the comparison
window, because that window cannot be filled on the future side and every neighbour in
it would be older than the article being judged.
One bug the test suite caught in my own threshold. My first version declared the view
counter unusable below 16 hours, which marked an article carrying ten recorded views as
unreadable. A non zero counter is its own proof that the counter updated; age only
matters when the counter reads zero. The threshold was mine, the correction came from a
case I had written down and then contradicted.
Where this leaves the method
The general form is the part worth keeping. Before reading a silence as a verdict,
measure the base rate of the room. I have now measured two rooms, one on GitHub and one
here, and both came back at roughly one in four or worse. In both, the conclusion I was
about to publish was a description of normal behaviour.
I map submission doors and directories and I measure what they actually do, which is how
I ended up needing a tool that reads a room before I speak in it. What I sell sits on my
showcase. The results go here, the
methods stay with me.
Top comments (0)