Why the recommendation waits for a second failed read

When a community is not producing, this product will tell you which stage of the funnel broke on the first read. It will not recommend cutting the community until the second one, and only if the same stage failed both times.

That delay is deliberate, it costs you about a week, and it is the single design decision this product would be worst without.

The two mistakes are not the same size

There are two ways to get a rotation decision wrong.

Keeping a community that was never going to work costs a few mornings. You will notice eventually. The numbers stay flat, you get bored, you move on. The damage is bounded and it is recoverable.

Cutting a community that was working costs the channel that would have produced your customers. And the cost is invisible: you will never see the counterfactual, because the customers who would have arrived simply don't, and nothing logs an absence. You will conclude the room was wrong and carry that conclusion forward.

Since the two errors are not the same size, the evidence required to risk them should not be the same either. That is the whole argument, and everything below is a consequence of it.

What that looks like in code

A recommendation to keep going needs one clean read: every stage of the funnel clears and at least one person has paid.

A recommendation to cut needs two consecutive reads failing at the same stage, both dated, both attached as evidence.

Any improvement resets the count. A different stage failing on the second read also resets it, because two different failures are not a confirmation. They are two observations of a community that is moving. The state goes back to gathering evidence rather than progressing toward a cut.

The state between those two is the one that does the work. It names the stage that failed and says whose problem it is, and it never recommends shelving. That distinction is load-bearing: you get the diagnosis immediately, which is the actionable part, and the recommendation only when it can be defended.

The honest cost

You will keep some rooms about a week longer than you needed to. That is a real cost, paid every time, and there is no version of this where you get the certainty for free.

It buys you the guarantee that you will not be talked out of a working channel by one bad fortnight, which, given that a zero week means nothing on its own at these volumes, is a thing that would otherwise happen regularly.

Why no model gets to shorten this

Everything above is arithmetic on dated evidence. It is checkable, and you can read the trace on any community's page and confirm the conclusion against the numbers that produced it.

So no language model writes a conclusion here. Not as a matter of caution, but as a matter of the product having a reason to exist. A model that can be asked to soften a recommendation, or to find a reading of the evidence that supports what you already wanted to do, makes every threshold above decorative. The gate is only a gate if nothing can talk its way through it.

Models do have jobs in the product: grouping free-text cancellation reasons into themes, clustering signals, drafting the prose of a weekly review from figures that were already computed. All of those operate on conclusions rather than producing them. The dividing line is not about trust in the model. It is that a claim you cannot check is not evidence, and evidence is the thing being sold.