初始化项目，由ModelHub XC社区提供模型

Model: rpant/iolai26-solve
Source: Original Platform

prompts/proposer.md

@@ -0,0 +1,32 @@
+You are a linguist solving a puzzle about {task_lang}, an unfamiliar language. You are given attested sentence pairs. Induce the grammar and output it as JSON — the grammar will be EXECUTED BY A MACHINE, so it must be complete and mechanical.
+
+## Attested data ({task_lang} = {work_lang})
+{pairs_block}
+
+## Hints from the puzzle
+{hints_block}
+
+## Automatic analysis (may contain errors — trust the data over this)
+Morpheme segmentation of {task_lang} words:
+{segmentation_block}
+
+Word alignments (task-language token -> likely meaning):
+{alignment_block}
+
+## Output format
+Emit ONLY a JSON object:
+{{
+  "lexicon":  [{{"morph": "<task-lang morpheme>", "gloss": "<{work_lang} word(s)>", "pos": "<N|V|ADJ|...>"}}],
+  "affixes":  [{{"position": "prefix|suffix|circumfix", "form": "<morph>", "form2": "<circumfix 2nd part>", "feature": "<the {work_lang} word or marker this realizes, e.g. 'the', 'plural'>", "trigger": "<pos it attaches to, or empty>"}}],
+  "rewrites": [{{"pattern": "<regex>", "repl": "<replacement>", "context": ""}}],
+  "order":    ["<pos tags in {task_lang} constituent order, e.g. V, N>"],
+  "agree":    [],
+  "redup":    []
+}}
+
+Rules:
+- Every {task_lang} morpheme in the data must be accounted for (lexicon or affix).
+- Prefer affix `feature` values that are actual {work_lang} function words ("the", "will", "not") — they are emitted verbatim when translating into {work_lang}.
+- rewrites capture sound changes at morpheme boundaries (e.g. "aa" -> "a").
+- The grammar must reproduce EVERY attested pair exactly when executed.
+- Simplest grammar that fits wins. No prose, JSON only.