add-rbnf

Add RBNF Rules

You are adding or expanding Rule Based Number Format (RBNF) rules for a language in the CLDR repository. RBNF rules turn the binary form of a number into words (e.g. 42 → "forty-two"). The user will specify a BCP 47 language tag (e.g. sw, de_CH, sr_Latn) — either in their request or, if not given, ask which language to work on.

Prerequisites

This skill assumes a usable CLDR build environment: Maven, Java JDK, and a built tools/cldr-code/target/cldr-code.jar (the Phase 3 validation steps run this jar and mvn test). If the jar hasn't been built yet, run mvn -B package -DskipTests --file tools/pom.xml -pl cldr-code (or an equivalent build) before starting Phase 3.

Phase 1 — Language Analysis

1. Parse the language code from the user's request. If none was provided, ask the user which language to work on.

2. Check if common/rbnf/<lang>.xml already exists.

   • If it exists, read it. You are expanding existing rules — do not modify working rulesets, only add missing ones.
   • If it does not exist, you are creating rules for a new language.

3. Identify the language's grammatical properties relevant to number formatting using your linguistic knowledge:

   • Grammatical genders (e.g. masculine, feminine, neuter) — do numbers inflect for gender?
   • Grammatical cases (e.g. nominative, genitive, dative, accusative) — do numbers inflect for case?
   • Number word irregularities (e.g. French 70 = soixante-dix, German compound ordering with units before tens)
   • Plural categories that affect large number words (million, billion, etc.)
   • The word used for the decimal separator (e.g. "point", "Komma", "virgule")
   • Whether the language has ordinal concepts that need custom digit-ordinal rules beyond what root.xml provides

4. Find the most structurally similar existing RBNF language file to use as a template:

   • Romance languages → fr.xml, es.xml, it.xml
   • Slavic languages → ru.xml, sr.xml, uk.xml
   • Germanic languages → de.xml, nl.xml
   • Other families → find the closest match among existing files in common/rbnf/
   • Read the chosen template file to understand the structural pattern.

5. Present your findings to the user:

   • Which grammatical properties are relevant
   • Which template language you'll use
   • Which rulesets you plan to create (list them)
   • Ask the user to confirm or correct your analysis before proceeding.

Do not proceed to Phase 2 until the user confirms.

Phase 2 — Incremental Rule Authoring

Build rulesets incrementally in complexity order. After writing each ruleset (or small group of closely related rulesets), run the full validation loop in Phase 3 before adding more.

Ruleset build order

1. %spellout-numbering — basic counting form (typically delegates to a base cardinal ruleset).
2. %spellout-numbering-year — year formatting.
3. %spellout-cardinal-* — cardinal forms:
   • Start with the base cardinal (often %spellout-cardinal-masculine or %spellout-cardinal depending on whether the language has grammatical gender for numbers).
   • Then add gender variants (feminine, neuter) that delegate to the base for most numbers but differ at key points (typically "1" and sometimes "2").
   • Then add case variants (genitive, dative, accusative, etc.) if the language inflects numbers for case.
4. %spellout-ordinal-* — ordinal forms, same progression (base, then gender/case variants).
5. %digits-ordinal-* — digit-based ordinals (e.g. "1st", "2e") in an OrdinalRules grouping — only if the language needs custom rules beyond root.xml.
6. Private helper rulesets (%% prefix) as needed — add these alongside the public rulesets that use them.

Rule authoring standards

Every public ruleset (% prefix) must include these special rules:

• -x: — negative number handling (e.g. minus >> or moins >>)
• x.x: — fractional number handling with the language-appropriate decimal word (e.g. << point >> or << Komma >>)
• Inf: — infinity handling (e.g. infinity or infini) — only include if the language has a word for infinity; check the template language for guidance

Termination rules:

• Cardinal/numbering rulesets: 1000000000000000000: =#,##0=;
• Ordinal spell-out rulesets: 1000000000000000000: =#,##0=.; (or equivalent digit fallback like =%%digits-ordinal=)

For large number words (million, billion, trillion, quadrillion), use the plural syntax when the language has plural forms:

1000000: << $(cardinal,one{million}other{millions})$[ >>];

When a rule needs different text for standalone (zero remainder) vs. compound (non-zero remainder), use the bracket-pipe syntax [ >>|text] instead of splitting into two rules. The text before | is used when the remainder is non-zero; the text after | is used when it is zero:

// English ordinal: "two hundredth" (standalone) vs "two hundred first" (compound)
// <%spellout-numbering< formats the quotient through the cardinal ruleset
100: <%spellout-numbering< hundred[ >>|th];

// Telugu ordinal: "వెయ్యవ" (standalone 1000th) vs "వెయ్యి మొదటి" (1001st)
// Common prefix "వెయ్య" gets "ి >>" when compound or "వ" when standalone
1000: వెయ్య[ి >>|వ];

This eliminates the need for separate standalone/combining rule pairs (e.g. 1000: వెయ్యవ; / 1001: వెయ్యి >>;).

Naming conventions:

• Public rulesets: %spellout-cardinal-<gender>[-<case>] (e.g. %spellout-cardinal-masculine-genitive)
• Private helpers: %% prefix (e.g. %%et-un, %%spellout-leading)
• Ordinal rulesets: %spellout-ordinal-<gender>[-<case>]
• Digit ordinals: %digits-ordinal-<gender> (in OrdinalRules grouping)

XML file format

For a new language, create common/rbnf/<lang>.xml with this structure:

<?xml version="1.0" encoding="UTF-8" ?>
<!DOCTYPE ldml SYSTEM "../../common/dtd/ldml.dtd">
<!--
Copyright © <this year>-<this year> Unicode, Inc.
CLDR data files are interpreted according to the LDML specification (http://unicode.org/reports/tr35/)
For terms of use, see http://www.unicode.org/copyright.html
-->
<ldml>
    <identity>
        <version number="$Revision
quot;/>
        <language type="<lang>"/>
    </identity>
    <rbnf>
        <rulesetGrouping type="SpelloutRules">
            <rbnfRules><![CDATA[
... rules go here ...
]]></rbnfRules>
        </rulesetGrouping>
    </rbnf>
</ldml>

For locale variants (e.g. de_CH, fr_BE), add <territory> or <script> elements in the identity block as appropriate.

Add an OrdinalRules grouping only when the language needs custom digit-ordinal rules beyond root.xml:

        <rulesetGrouping type="OrdinalRules">
            <rbnfRules><![CDATA[
... digit ordinal rules ...
]]></rbnfRules>
        </rulesetGrouping>

For an existing language, read the existing common/rbnf/<lang>.xml file and add new rulesets inside the existing <rbnfRules> CDATA section without modifying rulesets that already work.

Phase 3 — Validation

Run this validation loop after each ruleset (or small group of related rulesets) is added or modified. Track how many times you have attempted validation for the current ruleset — if you exceed 5 attempts, stop and ask the user for guidance.

Validation steps

1. Delete the existing .ssv file (so the generator will create a fresh one):

   rm -f common/testData/rbnf/<lang>.ssv
   

2. Generate test data from the current rules:

   java -DCLDR_DIR=$(pwd) -Xmx6g -jar tools/cldr-code/target/cldr-code.jar generate-rbnf-ssv
   

3. Run the RBNF tests:

   mvn -DHAS_CLDR_ARCHIVE=false -Dtest="TestRBNF" --file tools/pom.xml -pl cldr-code test
   

4. Read the generated .ssv file and review it for linguistic correctness:

   • Read common/testData/rbnf/<lang>.ssv
   • Spot-check these key numbers: 0, 1, 2, 3, 10, 11, 12, 13, 20, 21, 100, 101, 1000, 1000000
   • Verify gender agreement for "1" across all gendered rulesets
   • Verify correct plural forms for large number words
   • Check ordinals for 1-10 (often irregular)
   • Verify the decimal separator word matches the language convention
   • Verify the negative number prefix matches the language convention

5. If any issues are found (test failures OR linguistic errors):

   • Fix the rules in the XML file
   • Go back to step 1 and repeat

6. If everything passes, proceed to the next ruleset in the Phase 2 build order, or to Phase 4 if all rulesets are complete.

Error diagnosis

Phase 4 — Final Review

Once all rulesets have been written and individually validated:

1. Run the full validation loop one final time (Phase 3 steps 1-4) to ensure everything still works together.

2. Read the complete .ssv file and review all number words across all rulesets. Pay special attention to consistency between related rulesets (e.g. cardinal masculine vs. cardinal feminine should only differ where gender matters).

3. Present a summary to the user of what was created or modified:

   • List all rulesets added or changed
   • Note any known limitations or areas where the user should double-check
   • Show a few example outputs (e.g. how 1, 21, 100, and 1000000 render in each ruleset)

4. Apply code formatting if any Java files were touched:

   mvn --file=tools/pom.xml spotless:apply
   

5. The user can then commit the changes. The commit message should follow the format: CLDR-NNNN Brief description of change (the user will provide the JIRA ticket number).