We Fixed Three Things the Radar Was Getting Wrong

We shipped 1.30. Most of it is Diamond Radar admitting it was wrong about three specific things, and fixing them.

A permanent downgrade warning on 828 cards

Card attributes go up to 125. The rating curves we score against stop at 99. So any attribute above the curve produced a decline signal that no performance could ever clear — a standing +26 at 125, firing whether the player was raking or ice cold.

It hit 828 of 3,124 cards, concentrated exactly where it hurts most: the elite tail the Radar exists to call.

We now clamp the card to what the model can actually express. Only the unmodellable part goes — a 110-rated attribute whose real stats imply 80 still registers a genuine decline. Downgrade numbers from before and after this release aren't comparable. Don't read the change as your cards improving.

Fielding was diluting every weekly score

SDS moves fielding attributes roughly once a year, in a single late-July update. They're 1.81% of all attribute movement. But we were weighting them every week, on every card — spending model weight on something that almost never moves.

We stopped. Measured on the same cards:

We were blind to three attributes that move constantly

This is the one that stings. Our roster-update import read SDS's changelog — the list of what changed. It turns out that changelog silently omits three attributes:

• Discipline — moves on 47% of players
• Hits per BF (L) — 33%
• BB/9 — 27%

Every prediction we made this season was made without them. We now read the full attribute state directly instead of trusting the changelog, and we've rebuilt the season's data behind it.

Every price says what it is

The single "Est. Price" is gone, replaced by three named scenarios: if nothing changes, downside, and if it upgrades — the last being what Live Series peers at that OVR actually listed for, dated, with the sample count attached. A market comparable, not a forecast.

And when a comparable has no sample behind it, the row says so instead of showing you a confident badge. That was producing genuinely absurd calls: an 83-stub card labelled OVERPRICED because an unsupported comparable claimed a better version of it was worth 19 stubs. A card doesn't get cheaper by improving.

Also fixed

• 39% of 2026 market items were unreachable. Equipment, unlockables and other non-card items could be listed but not opened — the endpoint only ever looked in the card table. Fixed.
• Card pages were failing in bursts on the site. Traffic through the website was being rate-limited on the site's own outbound addresses instead of on you. Fixed.
• Radar calls are now grouped by confidence, and each group is labelled with the rate it has actually hit — so a high-confidence call and a marginal one no longer look identical.