Goal

Every chat catalog id should have an Elo so the settings list can sort and badge models. Only some of those numbers are leaderboard facts.

Attested

A score is attested when the Arena text / overall row maps to the catalog id by the rules below. Store the integer before ±.

Name matching (in order)

1. Exact lowercase match of the arena name to a catalog id.
2. Display-form normalize — treat (xHigh), (High), (Max), (thinking), (non-thinking), (thinking-minimal) as -xhigh / -high / -max / -thinking / -non-thinking / -thinking-minimal suffixes; collapse spaces/underscores to -.
3. Punctuation aliases of that same token — . vs - in version numbers (claude-opus-4.6 ↔ claude-opus-4-6), plus :free / -free on the same id.
4. Same-model alias lines already grouped in Kotlin (several ids on one line that already share one score) inherit together when one of them is attested.

Must not copy

• A dated snapshot onto a different dated snapshot (gpt-4o-2024-05-13 is not gpt-4o-2024-08-06).
• A dated snapshot onto a generic id that is already bound to another snapshot (gpt-4o stays with gpt-4o-2024-08-06, not May).
• A quality / thinking tier onto the base id (muse-spark-1.2 (xHigh) updates muse-spark-1.2-xhigh only; mimo-v2-flash (thinking) is not mimo-v2-flash).
• Unrelated ids that only share a source line for formatting (gpt-4o-mini and gpt-oss-20b are not aliases).
• A score from WebDev / Vision / Agent / Search boards onto this catalog. Text / overall only.

When two rules conflict, keep the more specific arena row.

Estimated

If a chat catalog id still has no attested score:

1. Keep the existing estimate if one is already in arenaScores (default on refresh).
2. Only fill a missing score:
3. -latest / -preview of an attested base → that base score
4. Fine-tune / quant / :free of an attested base → that base score
5. Else the closest same-family, similar-size sibling that is attested
6. Never invent a score for a non-chat id (isChatModel() false: embed, tts, guard, image, veo, …).

Mark estimates in the auto-fill section of arenaScores, never in the attested snapshot.

New board models

Add a baseEntries row only for a newly seen frontier / currently shipped chat model that Kai users can select. Do not backfill historical arena rows (early Llama, Dolly, Vicuna, …).

New rows need display name, context window (from the leaderboard when listed, else a documented family default), release month, and the attested Elo. Do not remove existing catalog entries.

Misses

Runtime logs ModelCatalog miss: once per unknown id. A miss is a candidate for a later baseEntries add — not a reason to fuzzy-match Elo.