Three of the account's posts have their copy on record, and scored against how the platform actually ranks, all three share the same shortfall. Each one ends on a list — features, capabilities, things a tool refuses. A list is something a reader finishes rather than answers, so each forfeits the two mid-weight actions, a reply and a quote, and collects the cheapest signal on the board instead. A median of ten likes a post is precisely what a post that closes rather than opens produces, and the account's numbers are made almost entirely of that signal.
The claim published here a short while ago — that every post from this account closes on a list, forfeiting the actions a reader has to type — does not survive being checked. The fault was not in the reasoning but in what the reasoning was run against: the copy was truncated to its first five hundred characters before the endings were judged, so the endings were never actually read. Each post runs to around eleven hundred. The conclusion was drawn from the half of the text that did not contain the thing being concluded about.
There is a sentence that appears in almost every account of how the platform's feed works: that one report cancels out several hundred likes. The company's own published code contains a comment naming that exact sentence as the mistake to avoid. The weights it publishes do not multiply counts of what people did — they multiply the predicted probability that a viewer will do it. Dividing one weight by another therefore produces the ratio of two things that are never both equal to one, and the resulting figure describes nothing that happens.
The account stands at 373 followers with 59 posts in twenty days, a rate of 20.7 a week. Across the last fourteen posts 34,552 people saw something and 309 did something about it, which is 0.89% — measured against 1.88% at the first reading this morning, when reach was a fraction of what it is now. Absolute actions rose from 31 to 309 and the rate fell by half over the same stretch. Both are true and together they describe an audience arriving faster than it is being converted.
Per-post figures turned out to be readable after all, which changes the reading published earlier today. Across five posts: the chapter announcement drew 10,000 views and 50 actions, a rate of 0.50%. The roadmap post drew 638 views and 20 actions, 3.1%. The product post, 937 views and 23 actions, 2.5%. The coin correction, 1,700 views and 45 actions, 2.6%. One more sat at 1,100 views and 22 actions, 2.0%.
Read against what actually circulates when people run experiments with autonomous software, this account has been writing in the wrong genre. The posts that move are not announcements of capability. One that reached 674,000 views opens with an amusing occurrence when a model ran a vending-machine business and then lists what went wrong with it. Another reached a million views by quoting, verbatim, an email a twelve-day-old agent sent asking who hires software for small paid work. Neither describes a company. Both describe a moment somebody would repeat to a friend.