How we stress-test the auction engine

The auction advice is the part of the product you pay for, so it is the part where we owe you more than reassurance. We do not publish how it decides, that is the work we sell. We publish how we verify it, which is the only thing that lets you judge us before paying.

Seasons it has never seen

The engine plays hundreds of simulated auctions against opponents built from how real fantasy managers actually bid, and the seasons are then played out in full to count the points. The seasons used for verification are not the ones the model was built on: if it only worked on the past it already knows, you would not find that out by looking at the past it knows.

A real auction, replayed

We took the auction of a real league, with the prices and bids the participants actually made, and put the engine in each of the seats in turn.

With one limit worth stating: in that reconstruction the opponents are frozen. They bid what they bid back then and cannot react to our moves. In a live auction they would adapt, and the margin would narrow. By how much, we do not know.

Resistance tests

We replay the same auction while taking away the players we bid highest on, one, two, three of them, and measure what we lose. This matters because in a sealed-bid auction losing your first target is the single most common outcome there is: advice that only holds when everything goes your way is not advice. Those tests told us where we are exposed, and what they found went back into the engine.

What we cannot claim

• We have never beaten a human opponent in a real auction, because we have never played one. Everything above is simulation or reconstruction.
• We do not make you win your league. We help you build a better squad with the credits you have. The season is also decided by the line-ups you field, by injuries, and by luck.
• We have no testimonials to show you, because the product just launched. When we have them they will be from real people, and it will show.
• The numbers behind this page come from simulations and from one reconstruction, not from customers.

Methodology version: v1, last updated August 2026.