Every rejection in the German run was the same rule: a translation ranked
below its own sense. The model tags a sense "medium" while correctly
tagging some translations "easy" — the translations are right and the
derived sense label is wrong, but the whole entry was discarded.
The prompt defines sense difficulty as the easiest translation difficulty
in that sense, so it is a derived value rather than an independent
judgement. validate.ts now recomputes it via applySenseDifficultyFloor.
The floor only ever lowers. Raising a sense to match its translations
would gate a concept out of levels it belongs in and collapse the
concept-vs-word distinction the two difficulty columns exist to express
(design-doc section 4).
- validate.ts: drop the cross-field rejection, add the floor; the valid
result now carries "normalizations" so repairs are reported, not silent
- pipeline.ts: count and print normalizations per batch and in the summary
- replay.ts: new, re-validates responses/ with the current rules and no
API calls; --write stages recovered entries, --langs and --verbose
- tests: six cases covering the floor, replacing the old rejection test
Replaying all 38 saved responses took the reject rate from 20 entries to
zero. staging.db now holds 746 words / 774 senses / 3,436 translations
with no sense ranked above its easiest translation.
Docs also record the API quota ceiling found today: the free tier allows
about 20 requests/day, not the 1,000 previously assumed.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Prompt is now a template (fixes the hardcoded en/es leftovers in rules
2, 3, 15, 16, 26, 31). pipeline.ts replaces the pseudocode: wordlist
normalization, skip-already-staged idempotency, batches of 20 against
gemini-3.6-flash with responseSchema, raw responses persisted per batch,
per-entry validation with rejection log, one transaction per word into
db/staging.db. Flags: --langs --pos --max-batches --delay-ms --dry-run.
Smoke run: 40/40 words staged (de+es, one batch each), 0 rejections.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- Translation count test now adds reverse link count to expected total
- Non-English translations test now filters to kaikki source only
- Target language test now filters to kaikki source only — reverse links
to English are valid and expected
- Replace checkOmwExists with checkExtractedFilesExist
- Wire up importKaikki and reverseLink as real stage implementations
- Track reverse link completion via sentinel row in run_status
- Update report to use resolved_entry_cefr and entry counts
- Stages 3 onwards remain as stubs
- Extract.ts now processes all 5 language files, filters non-English
entries by lang_code, skips translation extraction for non-English
(no translations in source files)
- Import.ts now imports all 5 language output files, uses language
field from ExtractedSense instead of hardcoding en
- Sample limit hardcoded to 500 entries per language for development
- Add stage-1-extract/scripts/extract.ts — streams Kaikki JSONL,
filters to supported POS and languages, skips abbreviations and
senses with no translations in supported languages
- Rewrite db/import.ts for Kaikki flat model — tracks sense_index
offsets per headword+pos to handle duplicate JSONL entries
- Rewrite db/schema.sql for Kaikki model — entries, translations,
LLM vote tables, resolved tables
- Add extract and db:import scripts to package.json
- Sample mode hardcoded to 500 entries for development
- Remove foreign key on run_status.source_id to support sentinel rows
for tracking one-time pipeline steps (compile_candidates, compile_votes,
merge, compare)
- Add stage-3-enrich/config.ts with all provider configurations,
ALL_PROVIDERS ordered local-first, and validateProviderKey() for
startup key checks
- Add .env.example with required API keys for OpenRouter and Anthropic
- Add pipeline:run script to package.json using --env-file .env
- Add .env to root .gitignore coverage for data-pipeline/.env
- write annotate.ts — matches CEFR source files against OMW translations
- match by word text + normalized POS
- add cefr_source vote to matched translations
- extract native example sentences from CEFR source files
- write one annotated JSON per language to stage-2-annotate/output/
- write conflicts.json for words with multiple CEFR levels
- update tsconfig to support all stage directories
- 2 German conflicts found (macht, bleiche)
- match rates: en 47k, fr 44k, de 26k, it 26k, es 26k
