Before the team decides on a proposed deal, two simulated readers look at it and say whether it is worth sending. One is Brand buyer, a skeptical buyer who spends real marketing money. The other is Talent manager, a protective manager who guards their client's rate card and reputation.
They are a cheap first filter, not real people. The real judges come later: a talent manager who reviews samples, and a brand that says yes and pays. Until enough real answers exist, treat the two scores as a hint.
| Reader | What it asks itself |
|---|---|
| Brand buyer | Would I pay this price for this talent for this deal? Is it clear what we get, where it runs, and for how long? Does this person fit my brand? |
| Talent manager | Would I let my client do this at this price? Are the terms vague, too broad, or unpaid? Does the brand clash with my client's niche or values? |
Both are told to be hard to impress. A seven out of ten means they would really sign it or recommend it.
Each reader returns four things:
A seed carries the worse of the two verdicts and the lower of the two scores, so a single drop is not hidden by a friendlier second opinion.
The two readers get the same facts and nothing else: the brand and its categories, a short talent card, the kind of deal, the proposed terms, the price, and the reasons the pair was matched.
They never see each other's answer. They are asked separately and at the same time, so one cannot lean on the other.
They never see protected traits. Gender, age, race, religion, looks, and similar personal details are kept out of what they are shown, and they are told never to judge by them.
If a reader fails or is skipped, the seed simply shows as not judged. It does not stop the seed from being made or seen. A failed re-judge never wipes out an earlier answer, so a drop already on a seed stays there.
Each seed shows a small panel with both readers side by side:
The seed's own badge shows the combined verdict and score. Above the list, the page says whether the judges are on, how much of this month's budget is used, and that a seed a judge said drop is never sent automatically.
A warning appears when the price has moved since the score was made: the old price is named and the panel asks for a re-judge.
Filters narrow the list to what needs attention: needs a fix, judge said drop, judges say send, or not judged, each with a count. Re-judge sits next to Edit price on any seed nobody has been asked about yet.
When a seed has a drop, Approve becomes Approve over drop, and a confirmation says plainly that the asks will not go out by themselves. The seed's timeline records who approved it and that it was over a judge's drop.
A drop is the one verdict that changes what the system may do on its own:
Fix, send, and not judged do not block anything. A seed the judges did not score — judges off, budget used up, or the call failed — can still be approved automatically. Once both real people have said yes, the introduction is not blocked either.
The judges can be switched off. When they are off, no calls are made and each new seed records that it was not judged.
There is a monthly budget, ten dollars by default. Spend is counted from the first of each month (UTC). When one more round would go over, new seeds are left unjudged with a note, and the page says the cap has been reached. Nothing else about seeding changes.
Show calibration puts the scores next to what really happened to each seed:
For each reader it reports how often send lined up with a yes and drop lined up with a no. Fix is counted separately, because it means "not like this", not a no. It also counts the seeds the team approved over a drop.
Read small numbers with care: under about twenty decided seeds the rates are noise, and the report says so.