Five weeks, unattended: what the line did on its own.

The number people ask about first is 335,369 views in five weeks. On its own it means very little. Plenty of accounts do that with a single lucky post and nothing to show the month after. What matters is what produced it, so this is the honest version.

The setup

We pointed the system at channels we own and stopped touching them. No editor picked a subject. No one approved a cut. No one chose a posting time. The brief was the same brief a client gets: here is the standard, here is the audience, run.

Every night it read what was moving in the niche, drafted a make-list with a written reason beside each row, cut the pieces to a specification measured off earlier work, scored each render against that specification, published what passed, and recorded what happened. Then it did it again the next night, choosing from what the last set had proved.

Most of what it made never shipped

This is the part that gets left out of numbers like 335,369. The line refuses. A render that misses the specification is not published and not delivered, it is rebuilt. On a single template we watched conformance climb across eight rounds before anything was released, and the seven attempts in between went nowhere.

The output figure is therefore a survivor figure. It counts what cleared the gate, not what the machine produced. We think that is the right way round: anything that always ships has no standard.

What it learned, and what it did not

The system measures retention, follower conversion and sentiment on every piece and uses that to choose the next slate. That loop is real and it ran the whole five weeks.

What we will not claim is that it improved itself in some general sense. Attribution across five weeks of output is thin, and the honest position is that the measurement informs the next decision rather than rewrites the machine. We have written elsewhere about the difference between a system that reports and a system that learns, and we are careful about which one we say we have.

The one that looked like a win and was not

Our most-viewed piece of the month was also one of the least useful. It did thirty-six thousand views and brought in almost nobody. Measured on views it was the best thing we made. Measured on whether anyone stayed, it was near the bottom, and the pieces that actually grew the channel were quieter and identity-led.

If we had been optimising on views we would have made ten more of them. That single comparison is the argument for measuring after publication rather than admiring the dashboard.

Why this is the demo

We did not build this to run our own channels forever. We built it because the only convincing way to sell a production system is to run one, in public, on our own name, and then show the workings. Five weeks unattended is a small sample and we will say so. It is also the difference between a claim and a receipt.