shapeshift Translation Quality Benchmark
Measures the quality of the /translate skill by comparing its output against existing human translations. Uses stratified key selection with a fixed/rotating split, LLM judges, and programmatic validation to produce a comprehensive quality report with regression tracking across all 9 supported locales.
Data Artifacts
All benchmark data lives in scripts/translations/benchmark/ (gitignored):
| File | Purpose |
|---|---|
testKeys.json |
Selected test keys with categories and fixed flag |
coreKeys.json |
Persistent core key set (stable across runs) |
ground-truth.json |
Captured human translations before removal |
report.json |
Latest benchmark report (becomes baseline on next run) |
baseline.json |
Previous report (auto-copied by setup.js) |
Pipeline (7 Steps)
Step 1: Select Keys
node .claude/skills/benchmark-translate/scripts/select-keys.js [--count N] [--core N]
Selects N keys (default 150) stratified across 6 categories: glossary-term, financial-error, single-word, interpolation, defi-jargon, general. Validates all selected keys exist in en + all 9 locales.
Fixed/rotating split:
--core N(default 100): Number of fixed core keys for stable regression tracking--count N(default 150): Total keys (core + rotating)- If
coreKeys.jsonexists: loads it, validates keys still exist in all locales, tops up if needed - If
coreKeys.jsondoesn't exist: selects core keys via stratified sampling and saves them - Remaining keys (default 50) are randomly selected as rotating keys from the non-core pool
- Each entry in
testKeys.jsonhas"fixed": true(core) or"fixed": false(rotating)
Outputs scripts/translations/benchmark/testKeys.json.
Step 2: Setup
node .claude/skills/benchmark-translate/scripts/setup.js
- If
report.jsonexists from a previous run, copies it tobaseline.json - Reads
testKeys.json, captures ground truth translations for all 9 locales - Writes
ground-truth.json - Removes test keys from locale files so
/translatecan regenerate them
Step 3: Translate
Invoke the /translate skill using the Skill tool. This regenerates the removed keys through the full translate-review-refine pipeline.
Step 4: Judge (Sub-Agents)
Launch 9 sub-agents in 3 waves of 3 (matching /translate's wave structure) using the Task tool. Each sub-agent receives the locale info, all key triplets, and glossary terms.
Wave 1: de, es, fr Wave 2: pt, ru, tr Wave 3: ja, uk, zh
For each locale, use this prompt:
You are an expert multilingual localization quality assessor for a cryptocurrency/DeFi application.
Rate translations from English into {LANGUAGE_NAME} on a 1-5 scale.
1 = Wrong/misleading meaning
2 = Significant issues (wrong register, missing nuance)
3 = Acceptable but could be more natural
4 = Good, natural, accurate
5 = Excellent, indistinguishable from professional native translation
Check: meaning preservation, naturalness, register ({REGISTER}), UI conciseness,
glossary compliance (these stay English: {NEVER_TRANSLATE_TERMS}),
placeholder integrity (%{...} preserved), DeFi terminology conventions.
Rate each translation INDEPENDENTLY. Community translations can contain errors.
Input: JSON array of {key, english, human, skill}
{ITEMS_JSON}
Output: Return ONLY a JSON array of objects with these exact fields:
{key, humanScore, skillScore, humanJustification, skillJustification, preferenceNote}
Scores must be integers 1-5. Justifications should be 1-2 sentences. preferenceNote should say which is better and why, or "tie" if equal.
Locale info for prompt substitution:
| Locale | Language | Register |
|---|---|---|
de |
German | Formal (Sie) |
es |
Spanish | Informal (tú) |
fr |
French | Formal (vous) |
ja |
Japanese | Polite (です/ます) |
pt |
Portuguese | Informal (você) |
ru |
Russian | Formal (вы) |
tr |
Turkish | Formal (siz) |
uk |
Ukrainian | Formal (ви) |
zh |
Chinese (Simplified) | Neutral/formal |
Building the items array for each locale:
- Read
scripts/translations/benchmark/ground-truth.json - Read the current (post-translate)
src/assets/translations/{locale}/main.json - For each test key, build:
{ key: dottedPath, english: groundTruth.english[key], human: groundTruth.groundTruth[locale][key], skill: getValueFromLocaleFile(key) }
Getting never-translate terms: Read src/assets/translations/glossary.json, collect all keys where value is null (excluding _meta).
Each sub-agent must write its output to /tmp/{locale}-judge-scores.json. Parse the JSON array from the sub-agent's response and write it to that path.
Step 5: Compile Report
node .claude/skills/benchmark-translate/scripts/compile-report.js
Loads judge scores from /tmp/{locale}-judge-scores.json, runs programmatic validation (including Cyrillic script check for ru/uk), computes summary stats, and writes scripts/translations/benchmark/report.json. If baseline.json exists, includes regression deltas. Report includes coreSummary and rotatingSummary alongside the overall summary.
Step 6: Restore
node .claude/skills/benchmark-translate/scripts/restore.js
Restores locale files via git checkout --, verifies no diff remains.
Step 7: Present Results
Read the compile output (printed to stdout) and present to the user:
- Overall score summary with baseline regression (if available)
- Core vs rotating stats (divergence suggests overfitting)
- Notable improvements/regressions
- Per-locale and per-category highlights
- Any items needing attention (low scores, validation failures, glossary issues)