Repository navigation
docs: fix the remaining stale locale counts and translation caveats - #222
Conversation
#220 corrected the locale list in CLAUDE.md and ARCHITECTURE §23 but missed three other places that still claimed 9 languages, and left two claims that shipping Polish has now falsified: - README.md and CONTRIBUTING.md both still said 9 locales and listed 9. Both now say 11 and include Greek and Polish. Greek had already shipped without either file being updated, so this drift predates Polish. - ARCHITECTURE's LiveKit limitation repeated "9 languages". - ARCHITECTURE's "Plural rules are 2-form only" bullet said shipping Polish correctly needs a CLDR plural-category selector first. Polish now ships and does not need one: its plural-bearing keys dodge the missing categories (durations use the invariant abbreviations godz./min, and hostsLabel's inny/innych reads correctly at every count). Russian and Arabic still do need the selector. - "Every non-English locale is an LLM draft" is no longer true of pl, which was contributed and grammar-checked by a native speaker. Reworded to "most", with the exception stated and its two developer edits named.
There was a problem hiding this comment.
Important
Two stale claims this sweep was meant to eliminate survive elsewhere in the tree — one directly contradicts the PR's own new plural bullet.
Reviewed changes
- Locale counts —
README.md,CONTRIBUTING.md, andARCHITECTURE.md§18 all move from 9 to 11 with Greek and Polish; the prose list inREADME.mdand the code list inCONTRIBUTING.mdboth enumerate exactly the 11 locale files that exist (en es fr fr-CA de el it nl pl pt sv). - Plural caveat — the §18 bullet now explains why Polish ships despite 2-form plurals (invariant
godz./min,inny/innych) and keeps the warning for Russian/Arabic. The claims check out againstpl.jsonandbook.go. - Native-review caveat — "every non-English locale is an LLM draft" becomes "most", naming
plas a native-speaker contribution inREADME.md,CONTRIBUTING.md, and §23.
The count fix is complete and the rewritten caveats are accurate where applied. The two issues below are the same claims, left in places the sweep did not reach.
⚠️ §23 "Known limitations" still says Polish cannot ship without a CLDR plural selector
docs/ARCHITECTURE.md now carries two versions of the same limitation: the rewritten §18 bullet (lines 993–997) says Polish ships fine because its keys dodge the missing categories, while §23's Known limitations (lines 1348–1350) still says richer-CLDR languages, "Polish, Russian, Arabic", "need a real plural-rule selector before they can ship correctly". Polish has shipped, so the second statement is false and contradicts the first.
Technical details
# Reconcile the two copies of the "plurals are 2-form" limitation
## Affected sites
- `docs/ARCHITECTURE.md:1348-1350` (§23 Known limitations) — still lists Polish as unable to ship correctly without a CLDR plural-rule selector.
- `docs/ARCHITECTURE.md:993-997` (§18 Known gaps) — the PR's corrected version, which says Polish ships and only Russian/Arabic need the selector.
## Required outcome
- Both limitations describe the same reality: 2-form keys only, Polish happens to fit, Russian/Arabic do not. Either update §23 to match §18 or have §23 point at §18 as the single source.
## Open questions for the human
- §18 and §23 track the same limitation in two places; worth deciding which owns it before they drift again.⚠️ CLAUDE.md still says "every non-English locale is an LLM draft"
The claim the PR corrects in three files survives verbatim in CLAUDE.md:156-157. Since this file is loaded as standing instructions for agents, the stale wording actively propagates — an agent reading it will repeat the false "every" claim and ignore that Polish was grammar-checked by a native speaker.
Technical details
# Align CLAUDE.md's translation-quality note with the corrected docs
## Affected sites
- `CLAUDE.md:156-157` — "**Every non-English locale is an LLM draft with no native review.** Structure is verified; wording is not."
- `README.md:209-210`, `CONTRIBUTING.md:82-84`, `docs/ARCHITECTURE.md:1335-1339` — the corrected "most … none has had a full native read; `pl` is the exception" wording this PR introduced.
## Required outcome
- `CLAUDE.md`'s caveat should say the same thing as the three files this PR fixes: most are LLM drafts, none has had a full native read, `pl` was contributed and grammar-checked by a native speaker.
## Suggested approach (optional)
- Reuse the shape of the §23 sentence so the four statements stay in step.deepseek-v4.1-flash | 𝕏

Follow-up to #220 (merged), which corrected the locale list in
CLAUDE.mdandARCHITECTURE §23but missed three other places and left two claims that shipping Polish has now falsified.Stale counts
README.md:203CONTRIBUTING.md:77ARCHITECTURE:990Note the drift predates Polish: Greek (
el) shipped without any of these being updated, so the docs have been wrong since then and #220 only fixed two of five spots.Claims Polish falsified
Plural rules. ARCHITECTURE said "shipping Polish, Russian or Arabic correctly needs a real CLDR plural-category selector first." Polish now ships and does not need one, because its plural-bearing keys happen to dodge the missing categories:
godz./min, so%d godz.is right for 1, 2, 5 and 22;hostsLabel'sother/otherspair isinny/innych, which is correct at every count (the genitive plural covers everything except exactly 1).That is luck of key choice rather than a capability, so the bullet now says so explicitly and keeps the warning for Russian/Arabic, which genuinely need the selector.
"Every non-English locale is an LLM draft." No longer true of
pl— contributed and grammar-checked by a native speaker in #216. Reworded to "most", with the exception stated and the two developer edits on top of his text (CLDR date forms,innych) named in ARCHITECTURE so a future reader can judge how reviewed it really is.Scope
Docs only, no code or locale changes.
grep -rn "9 languages\|9 locales"across the tree is now empty.🤖 Generated with Claude Code