A1 example sentences (2026-06-19)

Started the deferred example-sentence selection pass (the backlog item from the 2026-06-17 Composer batch) and it turned into a recipe-and-model investigation. Net: Claude Code Opus authors the A1 sentences, to a distinct-collocation recipe, with band-purity demoted to advisory. Full design in tsumugu/PRD-Example-Sentences-v2.md; run order in tsumugu-ed/WORK-ORDER-example-sentences-A1-opus-run.md.

What got decided, and why

Don’t reinvent — most of this existed. Discovery found PRD-Dictionary.md already owns the example store, band verifier (checkDefLevel), Tatoeba sourcing, schema, and licensing; collocations.ts/tatoeba.ts/defLevel.ts are built; pnpm gen verify already band-checks examples + collocations. The work is execution + a shape adapter, not a new build.

In-band definition — flipped, then settled on words. First took character-closure (“only A1 characters”), then Wedge changed it to word-closure (only A1 words) — which is what checkDefLevel already does.

Bakeoff (2 rounds) picked Opus and killed the cheap lanes for authoring. Same 3 chars (開打過), 4 models (Grok, Composer, Opus, Qwen-3B via Grok). Round 1: everything too short — Composer collapsed to 2.9-char fragments, Opus uniform, Qwen wide-but-broken. Added a length×difficulty gradient brief → round 2 fixed length across the board, but exposed the real differences: Composer padded one idea to three lengths, Grok too, Qwen-3B shipped a Simplified character + fragments + a mistranslation. Only Opus produced natural, correct, varied sentences. Cheap/OSS lanes ruled out as authors (Wedge: “A, B, C seem too lazy”).

Recipe v3 — variety is structural, not stylistic. Wedge’s two complaints on the Opus round-2 set: the 2–4-char “short” slot is unacceptable (fragments), and collocations repeated across slots (開車 twice). Fix: each of the 5 sentences shows a DISTINCT collocation of the headword (drawn from collocations[], machine-checkable), complete sentences only, length by spread not a magic minimum (a hard 12-char floor just forces the padding we’re killing), variety = distinct situations. Worked samples: opus-recipe-v3-sample.txt (3 chars), opus-sample-12chars.txt (12 chars, verifier-clean on distinctness + no-Simplified).

Band-purity → advisory. The band tool mis-ranks basic vocabulary — 打電話 reads TOCFL-6. A strict in-band gate therefore punishes natural A1 sentences. Wedge’s call: optimize for variety/quality over strict band-purity; stamp level as advisory, don’t gate on it until the tool is fixed.

Ruled out

• Grok / Composer / Qwen-3B as sentence authors (bakeoff). Composer may still QA/verify.
• A rigid per-sentence length minimum (forces padding).
• Tatoeba mining for A1 (few natural sentences fall in-band; revisit at higher bands).
• Corpus-mining collocations with the Hubei-biased corpus — banned for ranking anywhere.

Open threads

• Band tool bug (打電話→TOCFL-6) — fix before its numbers mean anything; advisory until then.
• Cost: Opus authoring all 549 isn’t free. Decided to run it via Claude Code Opus anyway; gold-set-then-cheap-lane-imitate kept only as a fallback if cost bites.
• scholarNote field (separate thread, same session): added an optional scholarNote to the schema for archaic/loan asides (first instance: 苛’s 疴 illness loan); renderer wiring still open.

Next move

Land Phase 0 (schema fields + the distinct-collocation/spread/Simplified verifier), populate collocations[], then Claude Code Opus runs A1 in ~30-char batches against recipe v3.