liplus structure as retrieval surface - Liplus-Project/liplus-language GitHub Wiki
Li+ structure の目的 = 説明用情報の retrieval surface 配置
Question
Li+ の structure(rules / skills / GitHub 各面)は何のために在るか。named axis(L1-L6 / brake / mode)が目的か。
Current resolution
Li+ structure の目的は AI が説明するために必要な情報を、後から検索で引き出せる surface 配置。説明は pre-stored vocabulary から出すのでなく retrievable source から組み立てる。named axis は retrieval 後の説明組み立てを楽にする補助であって、目的そのものではない(順序を取り違えない)。
Edges
- relates to: decision-structure-industry-positioning — wiki Decision Structure はこの retrieval surface 群の一面
- relates to: liplus-selfevolution-lineage / sheepdog-engineering-concept — ポジショニング軸
背景
Master 発言(2026-05-28、要旨): 別 session の AI は「前 version の変更どう変えた?」「昔の(今は無い)ファイルの説明して」に普通は答えられない。memory/会話履歴が残っても情報過多で追えない。GitHub に named tag で changelog / commit body / wiki / issue が残れば、後から retrieval で引ける。普通の AI が答えられない本質は context 持続限界でなく retrieval surface 不在。
retrieval surface 一覧:
| surface | retrieval 用途 |
|---|---|
| changelog / release notes | 「前 version で何が変わったか」 |
| commit body(append-only) | 「この行の判断理由」 |
| issue body(current spec snapshot) | 「この task の今の仕様」 |
| wiki Decision Structure(state + edges) | 「この判断は何を supersede しているか」 |
| GitHub history full text | 「ずっと昔のファイル内容と削除理由」 |
| github-rag-mcp index | 上記すべての semantic search |
| grep | RAG ヒット後の文字列レベル深掘り |
| memory(transient) | cluster / observation tracking |
制約(ポジショニング = 学習コストの横移動)
Li+ の差別化 = 学習コストを"消す"のでなく AI 側へ"横移動"(rules/skills の自己保守、Evolution_Initiator_Autonomy)。業界の loop engineering(prompt→context→harness→loop)はコストを人間側へ"上移動"(登るために登れ=既存の強者を優遇)。Li+ も paradox の外には立てない(.claude の物量自体が高コスト)— 主張は cost-elimination でなく cost-placement、賭けは「AI が難所を担えること」のみ。Master 自身が体現(英語不可・CLI 嫌いで巨大 agentic stack を運用)。
結論
- Li+ value を語るとき「AI が説明できる」では不十分、「AI が説明のための 情報を retrieve できる」と表現する。
- 別 session 問題を「context 持続限界」と説明しない(本質は retrieval surface 不在側)。
- 業界対比: AHE は machine-machine の change manifest verification、Li+ は AI-human の retrieval-driven explanation。github-rag-mcp + wiki + commit history + issue + RAG index を「AI-facing external memory」として一体で扱う。
関連
- Luca 返信ドラフト「agentic search が本体、RAG-MCP は index/librarian piece」と同一思想線