The second gate

4 September 2026 — 2,156 green lights, and an empty queue tonight.

In short: machines work on their own inside the tower. But some things must never ship without someone saying yes. Decisions is the single screen where that yes is given. To date: 2,156 decisions made, 516 refusals, none pending tonight.

The problem: saying yes cost too much

At first, approving meant typing commands, one line at a time. Then came the sentence that started it all: doing this in a console is painful. This is not about comfort. When saying yes is a chore, you do it less often. Meanwhile, machines wait and nobody knows.

Worse: whatever was waiting sat in scattered places. A job here, a study there, a security finding somewhere else. Nothing showed, at a glance, what was blocked.

One queue, and nothing else

Decisions gathers everything that needs a yes, anywhere in the tower, into a single list. A machine asking to pass a gate, a stuck agent needing an answer, a finished study, a hire, a website release: it all lands in the same place.

Three levels of urgency, not five: blocking, important, whenever you can. Past three, nobody knows what "medium-high" means, and sorting stops sorting anything.

The rule that prevents two truths

A decision card holds only the decision and a link to the original thing. Never a copy of its content. Otherwise you end up with two versions of the same work, drifting apart — and nobody knows which one is real.

Approving acts — it does not tick a box

This is the whole difference with a plain status field. When you approve, something happens: the job goes out to the agent, the new team member is hired, the study becomes a task sent straight to the workshop, the public site is updated. The click does not file a card away: it starts the work.

Refusing forces you to write why

You cannot refuse in silence. With no comment, the refusal is blocked. The reason is simple: a silent refusal teaches nobody anything, and the machine behind it will not know what to fix.

The comment does not sit in a corner: it travels back to the agent, who starts again taking the remark into account. This has happened 177 times.

But it is capped at three attempts. Past that, a refusal no longer relaunches anything. Without that cap, refusing the same thing ten times would make a machine work ten times — and cost money nobody decided to spend. At that point it is not the proposal that needs fixing, it is the original instruction.

What about things that make no sense?

There is a third button: file it. For requests nobody understands. Approving would launch them, refusing would relaunch them. Filing puts them away without triggering anything — and here too you must write why, otherwise the same thing comes back at the next sweep and you go in circles.

The detail that holds it together: never count twice

The collector runs in a loop. If it filed a second card for a request already there, the queue would grow by itself and the counter would lie. So a request already represented is never filed again. A screen that lies about the number is worse than an empty screen.

The measured result

That one-in-four refusal rate is the number that matters. A gate that has never refused anything guards nothing.

Takeaway: a machine may work on its own; it must not decide on its own. Decisions is where you take back control — one screen, one click, and the duty to say why when the answer is no.
🔭