-
-
Notifications
You must be signed in to change notification settings - Fork 0
liplus selfevolution lineage
Li+ の自己進化は既存研究系譜(AutoResearch / DGM / RSI トレンド)からどう派生し、どんな kind の自己進化か。
系譜は薄い。AutoResearch から借りたのは「自己進化系を出していい」という permission / existence-proof だけ(timeline 前倒しトリガー)で、loop も mechanism も継承していない。実装は独自。Li+ はベンチ反復型でなく dialogue-distillation(対話で蒸留 → 可逆なガバナンス/ソース編集 → 判断層・対話自然さ・人間で評価)。RSI トレンド地図上では off-ladder の新発明でなく、既存トレンドの Li+ 流 redefinition(scaffold-side RSI・重み固定枝)。発見全体が convergent-not-derived(指示読み飛ばしの現場問題から独立合流、研究系譜から来ていない)。
- relates to: sheepdog-engineering-concept / sheepdog-engineering-publish-intent — Li+ 固有は"先行"でなく框の選び方(harness メタファー批判)
- relates to: liplus-context-rot-tension — 独立合流の出発点(現場の指示落ち問題)
- relates to: liplus-evaluation-criterion — 評価軸=対話精度ゆえベンチ反復が移植不可
- relates to: l1-brake2-root-criteria-evaluator — L1 人間ゲートの provisional 性とその移譲
- (a) principled: ドメイン(統治/判断/対話)に静的 gold が無い → 駆動軸も評価器も対話の精度。task 再走で metric を上げる型ではない。
- (b) economic: Master は個人・$100/mo サブスク。AutoResearch(一晩~100実験)/ DGM 級のベンチ反復を手法として回す compute は端から budget 外。dialogue-distillation は普段の会話で回るので安い。
- → 2理由が同じ設計に収束(overdetermined)。
空席の交差点: 「自己進化 × 対話精度」は survey 範囲で参考ゼロ。材料は各々在る(自己進化=AutoResearch/DGM/Dreaming/Hermes、対話蒸留=grill-me)が掛け合わせは空席。理由 = novelty 半分 + 最難区画半分(対話精度に静的 gold 無し・sycophancy で偽の滑らかさ化・測りにくい → compute 持ちの lab は clean gold のベンチ側へ)。Li+ は「安い + gold無し + 空席」のレーンに座った。
Li+ の自己進化構造は問題にぶつかって収束したのではなく、Master が起きそうな問題を予め回避して組んだ部分も持つ(anticipatory foresight)。ただし Master の base mode は ぶつかり稽古(collision / trial-and-error)で、foresight は『先に避けられるもの』にだけ当てる部分集合(overlay)。「base が anticipatory」と一般化するのは AI の再発癖。
整合: Li+ は全て revertible。foresight ブレーキの gate は irreversibility でなく redo-cost × cascade-risk(L1/main = 高依存ノード=連鎖 → ブレーキ厳重 / ordinary L2-L6 = revert 安く cascade 浅い → collision で踏んで観測)。subtractive-structural-beauty の blast-radius を git 内部の依存深度で sub-divide した形。
L1 人間ゲートの位置づけ: 「壊れたから追加」でなく「foundation を勝手に書き換えられたらまずそう」という先読みの予防、かつ provisional(最初から外す前提)。実行済 (2026-06-11, #1477/#1478): brake 2 を根本基準評価者 subagent へ移譲、human は post-hoc 対話評価者へ(l1-brake2-root-criteria-evaluator)。
RSI トレンド地図: prompt→context→harness→loop ラダーの harness と loop の間 = self-evolving harness / meta-harness / AHE / HyperAgents、根は recursive self-improvement (RSI)。2026 の合言葉「moat が harness に移った」。定型 = メタエージェントが task エージェントの scaffold を世代越しに書き換え、重みは固定。Li+ = scaffold-side RSI・重み固定枝(rules/skills/adapter を書換、モデル重みは不変)。三軸対応 = modifier(scaffold 自己書換)/initiator(改善ループ自己起動)/position(scaffold を内側から編集可能に)。HyperAgents の自己言及点 ↔ Li+ が自分の進化ルール(rules/evolution/*)/autonomy/brake を書き換える形。「Li+ ぶっちぎり」框で語らない。
DGM: Darwin-Gödel Machine(Sakana, arXiv 2505.22954)は参考にしていない。関係は convergent validation のみ(単一メトリック objective hacking の独立実証=Goodhart)。Li+ が単一自動ゲートを避けるのは自前理由 (a)(b) で、DGM を見て避けたのではない。
並走系: Anthropic Dreaming(2026-05-06、外部メモリ consolidation、weights 非変更、人間 approve/discard)/ Hermes Agent(Nous Research、2026-02-25、self-hosted daemon、自分で skill を書く)。両者とも可逆層の自己進化 + 部分自律で Li+ と同位置だが kind は別(Li+ はガバナンス source の対話精度駆動)。
「動かす前に考えられる基盤は作った。ここからはシープドッグとの殴り合い編だなw」 — 設計フェーズは done。残るループ risk(correlated blind spots、smooth drift、dialogue thinning)は empirical と宣言、走らせて surface する(pre-engineering でブレーキを足し続けない)。将来「走る前にもっと安全構造を足す」衝動への既定回答 = まず現実に問題を出させ、feedback して直す。残る外輪 = 物理 substrate ring(polling-on-input → event-driven、Claude Desktop の channels 欠如が block)。
- AutoResearch/DGM と Li+ を mechanism レベルで結びつけない(実 link は category-permission のみ)。
- ベンチスコア/単一メトリックで自動採用を Li+ に提案しない。
- weight-payoff 問題(蓄積した重さが task スコアを上げるか)は『測る金が無くて』開いたまま(意志でなくコストで未解決)。現実的経路 = local LLM だが substrate-mismatch の二枚目の間接性が乗る。
- AutoResearch / DGM (arXiv 2505.22954) / Anthropic Dreaming (2026-05-06) / Hermes Agent (Nous, 2026-02-25) / RSI workshop (ICLR 2026)
この Wiki は、Li+ に基づく開発・運用を支えるための情報整理空間です。
数字で始まるページは、 Li+プログラムの各レイヤーの仕様を定義するページです。
- 要求(何を満たすか)と仕様(どう振る舞うか)を一体として記述する
- 実装前に作成または更新する
- issue群から採用された要件を集約する
これらのページは 安定性と一貫性を重視して管理されます。
アルファベットで始まるページは、 Li+の構想・設定・導入手順などの参照用ページです。
- 設計思想・背景
- 設定リファレンス・インストール手順
これらのページは 必要に応じて更新・拡張されます。
リポジトリ内の rules/**/*.md(L1–L4 の常時ロード分、subdir 含む)、skills/**/SKILL.md(トリガー起動分)、adapter/claude/CLAUDE.md、adapter/claude/hooks-settings.md、adapter/claude/hooks/*.sh、adapter/codex/AGENTS.md、およびルート直下の Li+config.md、Li+update.md は、
AIやランタイムが直接読む実行用プログラム / 定義ファイルです。
-
docs/は人間向けの仕様書・要求仕様・手順書 -
rules/,skills/および adapter / update は実行時に読み込まれる本体
両者は対応しているが、役割は同じではない。
Home | 1. Model | 2. Evolution | 3. Task | 4. Operations | A. Concept
要求仕様書 (1-6)
参考文書 (A-K)
- A. Concept
- B. Configuration
- C. Update
- D. Installation
- DiDD(対話駆動開発)
- E. Li+ language
- F. Behavior-First
- G. Sheepdog Engineering
- H. Roles and Evaluation
- K. Source File Format
判断構造
- Decision Structure
- layer reorg rationale
- github app user-to-server token expiration
- sheepdog engineering concept
- prerelease tag recovery procedure
- release flip drift patterns
- Li+ long-term vision (feedback only)
- Master role as client-architect
- current architecture as concession
- Li+ license Apache-2.0 rationale
- Character_Instance evolution history
- prompt as emotion vector controller
- agentic-search five-phase refactor
- Character_Instance output-styles migration
- Li+ lightening L1 gate override
- subagent state-machine label mechanism
- LSP integration out of scope
- Character_Instance opt-in and surface scope
- parallel-subagent-eval three-axis decomposition
- parallel-subagent-eval cost acceptance
- parallel-subagent-eval model floor
- release version rule always-on relocation
- bootstrap walkthrough skip and gh install relocation
- wiki sync sidebar integrity check
- decision structure rename rationale
- decision structure industry positioning
- subtractive structural beauty framing
- Li+ authorship is collaborative
- Li+ design intent vs current limit
- Li+ history is empirical
- Master verification at runtime not spec
- rules cache fetch address table
- dialogue-evaluator scoring redesign
- Li+ always-on footprint is load-bearing
- DiDD umbrella naming
- milestone subsystem removal
- L1 brake 2 root-criteria evaluator
- Hook-driven gate trigger
- dynamic-workflows non-adoption
- ACE context-engineering non-adoption
- memory GraphRAG SQLite exploration
- Li+ context-rot tension
- Li+ structure as retrieval surface
- Li+ evaluation criterion
- Li+ self-evolution lineage
- Li+ judgment-learning telos
- Sheepdog Engineering publish intent