Sorting ideas by score works until the scores are close, and then it quietly stops working. Here are three real cards from one market, 1 point apart, carrying three different verdicts. The ordering function our Compare mode runs puts Payment reconciliation timeline first — and the idea you should spend this week on is Overdue invoice follow-up, which it puts second. Both statements are correct, and the gap between them is the whole skill.
Why the score stops picking for you
A weighted score compresses seven judgements into one number so that a list can be sorted. That is useful at the extremes: an idea at the top of the range and an idea at the bottom are genuinely different animals, and you do not need help telling them apart.
In the middle it is doing something else. A point of difference between two ideas is not a fact about the ideas — it is a fact about how a rating landed on a scale, and those ratings are model output that moves between runs. Treating that point as the tiebreak is how a founder with ten hours a week ends up spending them on the wrong thing while believing they were being rigorous.
What does not move is what the checks said. An idea can be told to stop because it failed a deal-breaker outright, and a failed deal-breaker does not care how well the rest of the card scored. So the number and the verdict answer two different questions: how promising does this look and what are you allowed to do next.
Three ideas one point apart
These are published cards from our idea library — finalists of Discovery sweeps on time-tracking & invoicing for freelancers, first swept 2026-06-03. They were not chosen for the argument: they are the three whose scores sit closest together in the whole library, which is the only case where picking is hard.
Scores and verdicts verbatim from the published cards, bars relative to the highest, with each card's own leading risk underneath. Read the bars and then read the verdicts: the second column does not follow the first.
Three near-identical numbers, and one of these ideas is being told to stop. If you had ranked them in a spreadsheet you would have picked the top row, and the top row is the one card here that is not asking to be worked on.
What the ranking actually does with them
The Compare mode does not re-argue any of this. It takes results that already exist and puts them in a fixed order: score first, then the decision band when scores tie, then the identifier so the order is stable. This page imports that same function, so the table below is what the product would produce for these three cards.
| # | Idea | Score | Verdict | Its own next move |
|---|---|---|---|---|
| 1 | Payment reconciliation timeline | 61 | HOLD | Offer a reconciliation-dashboard waitlist with screenshots and collect signups before any build. |
| 2 | Overdue invoice follow-up | 60 | VALIDATE | Test a landing page with a waitlist and payment-ready CTA for automated overdue reminders. |
| 3 | Retainer prepaid-hours tracker | 60 | KILL | Killed: a 60-score with a spreadsheet-replicable wedge is a trap — the score alone doesn't save an idea. |
Produced by calling the Compare mode's own ordering function on the three cards above, not written out by hand. The tie between rows two and three is broken by the decision band, which is why the card being told to stop sorts last despite matching the score of the card above it.
And here is the sentence that function writes, unedited:
#1 (Payment reconciliation timeline) ranks highest at 61/100 — HOLD. Validate next: #2. Hold: #1. Kill: #3.
It names a rank-one idea and then tells you to act on a different one. That is not a defect in the summary — it is the honest shape of the situation, printed instead of smoothed over. A tool that only handed you the ordered list would have handed you the half that cannot say what to do.
The rule this implies when you have ten hours a week
Ten hours is roughly one experiment. Not one experiment per idea — one experiment, total, for however many weeks it runs. So the job is not to rank ideas, it is to spend a single budget.
- Filter before you sort. Anything carrying a stop verdict leaves the list, whatever it scored. Ranking it against what is left is arithmetic on a decision that has already been made.
- Among what is left, take the cheapest decisive test, not the highest score. Each card here carries its own next move, and they are not equally expensive. The one you can settle in a week is worth more than the one that scores a point higher and needs a month.
- Bench the rest in writing. Not discarded, benched — with the reason and the condition that would bring it back. The ideas you did not pick are the ones you will re-litigate at midnight in three weeks unless the reason is written down.
What a ranking cannot do for you
It cannot tell you the scores are close enough to be a coin flip — it will order them anyway, because ordering is what it does. The spread is yours to read, and 1 point is noise.
It cannot compare ideas that were never scored the same way. The ordering assumes every row came through the same rubric; three ideas rated by three different means produce a table that looks authoritative and means nothing.
And it cannot want anything. If you will not actually do the async outreach that Overdue invoice follow-up needs, its rank is irrelevant, and the honest comparison is between the ideas you would really work on.
Comparing three ideas in the product costs 1 credit, because comparison is priced per 3 ideas — each idea is a full check, and the ranking on top of them is arithmetic we do not charge for.
The short version
- Three published cards, 1 point apart, with three different verdicts. The score column and the verdict column disagree, and the verdict column is the one with a decision in it.
- The ordering is fixed in code — score, then decision band, then id — while the ratings it sorts are model output that moves. Argue with the order, not with the point.
- Rank one here is a hold. The card to spend the week on is rank two, and the tool says so in its own summary.
- With ten hours a week you are choosing one experiment, not one idea. Filter on verdicts, then pick the cheapest decisive test among the ideas still standing.