# trial run — implementation roadmap (pure vertical)

goal: english + italian nouns through the full pipeline on local models, emit
coverage report. build a thin end-to-end slice first, then widen each pass.

## phase 0 — foundation

- confirm @lila/shared types (lang codes, POS) exist and match
- package.json deps for what phase 1 needs
- one hardcoded test word to carry through the slice (e.g. "house")

## phase 1 — thin vertical slice (ONE word, 1a→3, crudest possible)

goal: prove a single card can travel the whole pipeline and come out the end.
allowed to be ugly — hardcode, skip voting, one model, fake CEFRLex.

- 1a: one lemma record, by hand or trivial read
- 2a: look it up in Kaikki, confirm it exists
- 2b: extract its senses → card(s) with gloss
- 2c: fill one missing translation with ONE local model, no voting
- 3: assign a difficulty crudely (even hardcoded "easy")
- OUT: one finished card. the pipeline has a shape.

## phase 2 — widen: real deterministic front (1a–2b, all words)

- 1a: real frequency list, verified
- 2a: real existence gate + miss-list triage
- 2b: real sense extraction, gender from tags, gloss-required drop
- still JSON output. inspect cards by hand.

## phase 3 — widen: real LLM back (2c + 3)

- storage seam: JSON → SQLite, schema, resumability
- 2c: real generate→vote, three families, all sub-passes
- 3: real CEFRLex single-sense + 3-model vote multi-sense + exclude-on-split

## phase 4 — coverage report + runs

- emit report
- run english, then italian
- decide: local good enough? rent vs API?