Every game has its own vocabulary — and why we don't flatten them

A database covering sixteen games faces one early decision: force every game into a shared schema, or keep each game's own shape. Almost every cross-game site takes the first option, because it makes the code simple. It also destroys most of what makes each game legible.

Here is the concrete case for the second option.

Nobody agrees on what to call a unit

The level axis is the range our stat calculators run over. Where a game layers promotion or Insight phases on top of levels, the axis runs continuously across them so the phases can be compared — Arknights' 220 is E0 Lv.1–50, E1 Lv.1–80 and E2 Lv.1–90 laid end to end, not a level cap of 220. A dash means we do not yet publish a growth-curve calculator for that game, or its calculator runs on something that is not a character level: Umamusume's is a 1–5★ axis and Star Resonance's is a 1–30 skill track, and printing either in this column would invite a comparison it cannot support.
^∗ Hololive Dreams ranks its gacha cards, not its talents, and Star Resonance ships classes rather than a rostered cast — neither has a rarity tier to report, which is itself the point of this page.
^† Duet Night Abyss reads 100% at 5★, and we are not printing that as a fact: the game's own rate table lists a 4★ character tier at 5.1% per pull, so the share is unreportable until that tier is carried. Flagged in our own build checks rather than smoothed over.

Sixteen games. Fifteen different words for the same concept, and not one shared rarity scale.

Three things the table shows that a flattened schema would hide

1. The top rarity is not a fixed number of stars

Blue Archive's best students are 3★. Arknights' are 6★. Dokkan's apex is LR, above UR, SSR and SR. Zenless does not use stars at all — its agents carry letter ranks, S and A, in a field the game calls rank rather than rarity. And two games here do not rank their roster at all.

A database that normalised all of these to "rarity 1–6" would have to invent a mapping. Is Blue Archive's 3★ equivalent to Arknights' 6★, because both are the top? Or to Arknights' 3★, because both are three? Both answers are defensible and both are wrong, because the games are not measuring the same thing. So each game keeps its own field, its own scale and its own label, and its pages say so in its own words.

2. "Top rarity" means something completely different per game

Look at the share column. In Wuthering Waves 79% of the roster is top-rarity; in CookieRun: Crumble it is 3% — two Cookies out of sixty-nine.

That is not a difference in generosity — it is a difference in what the category is for. In the small-roster games, top rarity is simply the normal state of a playable character and the lower tier is a starter pool. In Dokkan, with 4,348 cards accumulated over a decade, LR is a genuine apex that most of the catalogue never reaches.

The practical consequence: "it's a 5★" carries almost no information in Wuthering Waves and a great deal in Dokkan. Any cross-game tier list built on rarity is comparing categories that are not comparable.

3. A level is not a unit of anything

Limbus tops out at 45. Star Rail at 80. Dokkan runs to 150. And Arknights' "Lv.90" is meaningless on its own, because it is the ninetieth level of the third track — an operator climbs 1–50 at Elite 0, then 1–80 at Elite 1, then 1–90 at Elite 2. Three separate ladders, each restarting at one. Reverse: 1999 does the same thing with Insight phases.

So "Lv.60" describes a wildly different fraction of a finished unit depending on which game you are in, and in two of them it does not identify a state at all without naming the phase. This is why our calculators run a continuous axis across the phases rather than a per-phase level: it is the only way to ask "what is this unit worth at the point I will actually stop" and get a comparable answer. The consequences for comparing units are in why max-level stats mislead.

What we normalise, and what we deliberately don't

Only two things are made consistent across games:

• The record shape — every unit has an id, a name, official Japanese and Korean names, a set of facets and a set of stats. That is a container, not a schema; the facets inside it are the game's own.
• Derived comparisons — percentiles, cohort medians and nearest alternatives are computed the same way everywhere, but always within one game and one rarity. That is what makes them meaningful. See how to read a stat percentile.

Everything else stays native. Fate has Noble Phantasms; Arknights has Modules and Base Skills; Wuthering Waves has Echoes; Blue Archive has terrain grades and an armour table; Dokkan has Links and Categories; Umamusume has aptitude grades against a race calendar. None of these have equivalents in the other fifteen games, and a shared schema would either drop them or bury them in a generic "extra attributes" bag.

Why this costs more and is worth it

Keeping sixteen schemas means sixteen extraction paths, sixteen sets of facets and sixteen ways for something to break. A flattened model would be a fraction of the work.

It would also mean that a Blue Archive player looking for terrain grades, or an Arknights player looking for module types, would find neither — and would go back to the wiki, which is where the game-specific detail actually lives. The whole reason to build a database rather than read a card is that it holds the things the card leaves out. Flattening throws exactly those away.