We Were Publishing A Number That Undersold Us

We shipped 1.32. It is a small release, and the part worth your time is that we were publishing a number about ourselves that undersold what the Radar does.

The track record was answering a different question than you ask

The Radar's public track record has always shown a hit rate: of every card the Radar called, how many actually upgraded.

That is an honest number. It is just not the one you live.

You do not read every call. You open the board, look at the top of the list, and work down a few. So the number that matters to you is: of the top 10, how many were right?

Those are different questions, and they have different answers.

We now publish precision at the top 5, 10 and 20, per roster update and as a season total, alongside the hit rate rather than instead of it.

Across 44 major roster updates:

About 11 points, in our favour. We had been publishing the lower number and describing it as our accuracy. Both are now there, and now you can tell which one you are reading.

Why the top of the list does better

Because ranking is the thing the model is actually good at. It is more confident about the cards it puts first, and that confidence turns out to be earned. The hit rate flattens that out by counting a marginal call at the bottom of the board the same as the pick at the top.

When should you sell after an upgrade?

This came out of a question we get a lot: the night of, next day, or wait?

We measured it against real market data on every 2026 major — every card that actually upgraded, priced along the whole ladder afterwards.

The first day is not where the money is. The edge an upgrade gives a card keeps building for roughly two weeks, then flattens out. Selling into the first 24 or 48 hours captures about a third of what the upgrade ends up being worth.

Two honest caveats, because they matter more than the headline:

• The "typical +60% after an upgrade" figure you may have seen from us was mostly an artefact. If you take the biggest mover out of a large group, it will look spectacular whether or not anything real happened. Measured against matched control cards that did not upgrade, the real edge is a lot smaller — real, but smaller.
• We do not yet trust our own day-one number. The way that measurement is anchored means the baseline is taken after the first day has already started, which is the exact interval the question is about. We are fixing it. Until then, treat the shape of the curve as the finding and the first-day figure as unsettled.

This did not change anything the Radar serves you. It changed what our own internal screens are allowed to claim, and it is the reason the price columns there now carry a caveat instead of a green number with no explanation.

The rest

Most of 1.32 is groundwork you will not see: measurement scaffolding that lets us tell, after the next major roster update, whether recent changes to the model actually helped — separately, one at a time, rather than as one pooled number that describes no change anyone made.

That sounds like housekeeping. It is the difference between knowing a change worked and believing it did.