Behind the scenes: scoring deck quality

Not every deck on Braineos is equally useful to study from, and up to now there was no way to tell the difference except by actually opening one up. We've started computing a hidden quality score for every public deck -- it's not shown to anyone yet; it's the groundwork for things like better recommendations and discovery down the line.

Six signals, each capped

The score is built from six independent, deterministic signals, each contributing points up to a cap once it hits a "full score" threshold -- more cards or votes past that point don't add more points, so a deck can't game the score by piling on far past what's actually needed to look complete:

• Cards (up to 25 points, full credit at 10 cards)
• Votes (up to 25 points, full credit at 5 votes)
• Study attempts (up to 15 points, full credit at 10 attempts)
• Tags (up to 10 points, full credit at 2 tags)
• Remixes -- other people copying the deck (up to 10 points, full credit at 2 remixes)
• A real description (15 points for one of real length, partial credit for a token one like "n/a")

An optional second opinion

On top of that deterministic floor, a free-tier LLM can optionally rate the actual content -- are the term/definition pairs accurate, clear, non-duplicative -- and nudge the score up or down by as much as 20 points. It's additive, not a replacement: a deck padded with junk cards to hit the card count can still get marked down by the content review, and a deck that's merely sparse can't be fully rescued by one good LLM rating alone.

What never gets scored

A deck stays unscored if its owner is suspended, if there's an open report against it, or if it's private -- a private deck's cards are never even sent anywhere for this review. And this isn't running on a schedule yet; it's a script a human runs by hand, not something wired into creating or editing a deck.