Backend

LLM model switch

Google = Vertex. OpenAI / Anthropic / Meta = OpenRouter (`OPENROUTER_API_KEY`). Params + hints only; answers stay code-computed.

Google (Vertex)

OpenAI (OpenRouter)

Anthropic (OpenRouter)

Meta (Muse API · OpenRouter)

Live skeletons (Step 1–3)

JP grade 1–early 2 band. LLM fills params only.

Step 1arith.add.1dStep1 1桁の足し算
Step 1arith.sub.1dStep1 1桁の引き算
Step 2arith.add.2dStep2 2桁の足し算
Step 2arith.sub.2dStep2 2桁の引き算
Step 2arith.add.carryStep2 繰り上がり足し算
Step 2arith.sub.borrowStep2 繰り下がり引き算
Step 2arith.mul.2to5Step2 九九（2〜5の段）
Step 3word.change.rielStep3 お釣り（デイリー確認）

Template	Prompt	Skill	Grade	Status	Accuracy	Attempts
tpl-arith-01	12 + 15 = ?ai	arith.add.g3	3	Approved	88%	420
tpl-arith-02	8 × 7 = ?ai	arith.mul.g4	4	Approved	74%	310
tpl-arith-03	A market sells 24 mangoes in 3 bags. How many per bag?ai	arith.div.g4	4	Draft	—	0
tpl-arith-04	45 − 18 = ?ai	arith.sub.g3	3	Draft	—	0
tpl-word-01	Read: The rice field needs water. What does a farmer do first?ai	read.basic.g5	5	Rejected	41%	55
tpl-arith-05	36 ÷ 4 = ?human	arith.div.g4	4	Approved	81%	198
tpl-arith-06	9 + 14 = ?ai	arith.add.g3	3	Draft	—	0
tpl-arith-07	What is 1/2 of 16?ai	arith.frac.g5	5	Draft	—	0

Stage 0 pilot: approval gate skipped; live skeletons above are code-fixed. Payout logic stays outside this queue. Later stages re-enable draft → approved before learner fetch.