言語 / Language / 语言 / 언어: 日本語 | English | 中文 | 한국어

日本語

llive 完全解説 (2) — 「10 軸で考える AI」: 思考因子 × COG-MESH × 三重縞

コンセプト hook: 普通 AI agent は「思考」を 1 種類しか持たない. llive
は 10 種類の思考を同時に走らせ, それを互いに評価させ, 生き残った思考だけ
を集団へ取り込む. 10 種は「構造化」「再構成」「閉ループ」「自己拡張」
「不確実性」「探索」「整合」「来歴」「多視点」「現実接続」. これは認知科学
1990s〜2010s の主要 framework を 1 vector に圧縮したもの.

本日 (2026-05-21) marathon で 1881 PASS + v0.E 大規模前倒しが着地. 本記事は
その「思考因子側」 — COG-MESH-01〜10 と historical persona ontology (CE-19)
の交差点を辿る.

0. 連載中での位置づけ

#24-00 series index
#24-01 4 層メモリ
#24-02 思考因子 10 軸 + COG-MESH (← 本記事)
#24-03 構造進化 × TRIZ × Z3
#24-04 B-series (速い小脳)
#24-05 EvolutionLoop (遅い大脳)
#24-06 LLM backend non-transformer
#24-07 observability + governance
#24-08 lleval

10 思考因子 + COG-MESH は #24-05 の persona ontology (CE-19) と 1-N で結合.
本記事 #24-02 はそれを「何」と「なぜ」で説明する位置.

1. 10 思考因子の由来 — 6 framework の圧縮

ユーザー由来の 10 軸 (project_llive_cog_fx_factors). 元ネタは
「心理の深層」YouTube + 認知科学レビュー + Polya / Six Hats / Bayesian /
TRIZ / Provenance / Multimodal 系の 6 framework. それを 1 vector に圧縮した
結果:

これらは 直交ではない — 例えば factor_uncertainty と factor_exploration は
相関がある (UCB1 系). でも各々の 強さ を独立に持つことで, 集団内で
「同じ問題に 10 種類の見方で当たる」が可能になる.

2. なぜ 10 軸を 1 vector に持つか

LLM agent の文献では「思考は self-attention 1 種類」が主流. llive はそれを
vector に切り替え可能な multi-faceted thinking に拡張. これにより:

• persona との内積で「思考スタイル」が計算可能 — 例えば「岡潔 ベクトル」
  は (情緒) (国語力) (多変数) を高く持つ. 「ファインマン ベクトル」は
  factor_exploration + factor_reality_link を高く持つ.
• 集団内で同じ問題に 異なる持ち重みで 当たる派生個体を生成できる.
• 「この問題はどの軸が利くか」を fitness gradient で発見できる.

3. 主要因子 5 個の深掘り

3.1 factor_structurize — 「公理から積む」

axiomatic な思考. 数学者ガロア / グロタンディーク的. 抽象化階段を登る.
利点: 一般化能力. 欠点: 現実から離れる.

llive 内では BlockContainer の sub-block 順列が axiom 群に対応. factor_structurize
が高い派生は sub-block を 必須/任意 に分けてから再構成する mutation を好む.

3.2 factor_recompose — 「部品の入れ替え」

TRIZ Segmentation + 合成. 既存部品の組合せを書き換える. 利点: 局所探索高速.
欠点: 全く新しい構造は生まれない.

llive では PersonaImportAlgorithm (CE-20, 本日着地) がこの軸. 派生 A の persona
を派生 B が 部分採用する. 「ガロア + 岡潔」のような hybrid persona が
出現するのは factor_recompose を通る経路.

3.3 factor_closed_loop — 「自分を見て直す」

cybernetics の核. 自己観察 + 自己修正. llive では memory consolidation cycle
(海馬→皮質) と Approval Bus がこの軸. 集団内で評価 → 個体が結果を見て次世代に
反映する E.4 governance (CE-06/07/08, 本日着地) もここに乗る.

3.4 factor_uncertainty — 「分からないを定量する」

Bayesian / probability. 利点: 過剰自信を避ける. 欠点: 計算重い.
llive では Approval Bus の verdict 計算 + UCB1 exploration constant が代表.

3.5 factor_provenance — 「どこから来たか」

data lineage. Ed25519 sign + SHA-256 audit chain. llive Phase 4 (Production
Security MVR, v0.3.0) で着地. これは agent governance の 必須軸 で,
従来の LLM agent には欠けていた.

4. COG-MESH-01〜10 の対応

project_cog_mesh_implementation_2026_05_19. 10 因子に 1 機構ずつ 対応:

COG-MESH-10 は本日 marathon で CoevolutionGovernance として着地. これで
10 機構 → 10 因子 1-1 対応が完成. 集団内で どの因子が薄いか を機構の状態
から逆引きできるようになった.

5. 最新成果 (本日 2026-05-21 着地)

特に E.4 governance skeleton で COG-MESH-10 が closing できたのは本日の
最大成果. これにより 10 因子 ↔ 10 機構 1-1 対応が完成し, 派生集団の評価
→ 共謀検出 → Approval Bus 連携 が architecture level で繋がった.

6. 期待値 — 次に来るもの

6.1 CE-19 Historical Persona Ontology (短期)

既に 10 名 (岡潔 / グロタンディーク / ファインマン / ガロア / フォン・ノイマン
/ ニュートン / カント / ソクラテス / 老子 / 孫子) が PERSONA_ONTOLOGY として
着地済. 本日 CE-23 PersonaCorpusLoader skeleton が着地し, Raptor RAD コーパス
から persona を自動抽出して PERSONA_ONTOLOGY を拡張 する道が開けた. 次セッションで
LLM 抽出 + 実 RAD path 横断を実装し, persona 数を 30+ に拡大予定.

6.2 三重縞 (中期, ユーザー言語化)

「三重縞」 = 思考因子 / persona / 思考プロセス の 3 層が個体内で縞模様の
ように同時に走る状態. これは認知科学の 「並列認知」 仮説に着想を得たもの.
factor vector + persona composition + Six Hats / TRIZ / ARIZ をそれぞれ
別 layer で走らせ, 集団内 evaluation で互いを批評する. 着地時期未定.

6.3 神経インタフェース対応 (長期)

project_llmesh_neuro_long_term. Raptor RAD に bci / neuroscience /
neural_signal / prosthetic_neural / cognitive_ai / neuromorphic の 6 分野を
追加済. これは「脳 ↔ AI 直結インタフェース」が必要になったとき即座に
expand できるよう先回りで素材を集めている. 直接の実装は当面なし.

7. honest disclosure

• 「10 因子は overlap がある」 — factor_uncertainty と factor_exploration
  は相関 0.65 程度. 互いに直交ではない. 9 axis 化を検討した時期もあるが
  分かりやすさ優先で 10 のまま.
• 「factor_affinity の数値は heuristic」 — PERSONA_ONTOLOGY 10 名の
  factor_affinity vector は伝記 / 哲学史 ベースの人為的初期値. 後の
  PersonaCorpusLoader (CE-23) で コーパスベースに置換 されるが, 現状の
  数値は人による経験則.
• 「COG-MESH-10 は skeleton」 — 本日着地した E.4 governance は interface
  確立段階で, Quarantined Memory への 実書込み は別 module 委譲. 完成までは
  あと 1-2 セッションかかる.

8. Mermaid — 10 因子の構造

9. References (主要 20+ のうち抜粋)

• Polya, G. (1945). How to Solve It.
• Altshuller, G. (1971). TRIZ 40 inventive principles.
• Auer, P. et al. (2002). Finite-time analysis of the multiarmed bandit.
• Lehman, J. & Stanley, K. (2008). Exploiting novelty.
• Mouret, J.-B. & Clune, J. (2015). Illuminating search spaces by mapping elites.
• Hillis, W. D. (1990). Coevolving parasites improve simulated evolution.
• Constitutional AI (Anthropic 2022) — for HITL alternative.
• Six Thinking Hats (De Bono 1985).
• 岡潔『春宵十話』.
• ファインマン『ご冗談でしょう, ファインマンさん』.
• Maturana & Varela — Autopoiesis.
• Bayes — Essay towards solving a problem in the doctrine of chances.
• 完全リストは v0.6.0a1 リリース時に references.bib に同梱予定.

10. 2026-05-22 追記 — 10 因子 affinity vector の Rust 化 (RUST-15)

10 思考因子は派生個体の persona composition の effective_factor_affinity
として 10 次元 [0,1] vector で実装されている. 派生間の dissimilarity 計算は
本記事 #24-02 の中核機構と直結 — PersonaOverlapPenalty.apply (E.17) は
N×N pairs の persona_dissimilarity で 10 因子空間の距離を測る.

本日 (2026-05-22) RUST-15 として batch (NxN pair を 1 FFI call) Rust 化:

• single 1-pair: x0.80 (FAIL — FFI overhead で Python set 操作に負ける)
• batch N=64: x17.07 (PASS), 平均 x12.71

これにより「10 因子 vector の N×N pair 距離計算」が高速化され, 集団
N=64 で governance + diversity preservation を 64 Hz で回せる目処が立った.

10.1 思考因子側から見た意味

• factor_structurize (#0) と factor_exploration (#5) は TRIZ 系統で
  対立する 2 軸 だが, 10 次元 vector の L2 距離としては独立に効く
• PersonaOverlapPenalty (E.17 CE-25) で集団内 persona overlap を罰すると,
  派生集団は 10 因子空間で自然に散らばる
• MAP-Elites grid (E.17 CE-26) は persona 2 軸 × thought_factor 2 軸 の
  4 次元 grid なので, 上記の 10 因子 vector を 4 次元に marginalize して
  cell key とする

10.2 honest disclosure — 単発 Rust 化は逆効果

「思考因子 vector の距離計算を Rust 化」と聞くと「速くなる」と思いがちだが,
1-pair 計算では FFI overhead で Python の方が速い (x0.80). これは
feedback_rust_usage_matters 判定表の A パターン (純 Python ループ
1-pair). batch で N×N pair を 1 FFI に詰めて初めて x17.07 まで伸びる.

詳細は #24-05 と docs/perf_comparison/2026-05-22_kernel_implementation_comparison.md.

Series Navigation

• ← 前: llive 完全解説 (1) 「忘れない LLM」
• → 次: llive 完全解説 (3) 「矛盾は計算できる」
• 全体: llive 完全解説 (0) — series index
• repo: furuse-kazufumi/llive

English

llive Complete Guide (2) — "AI that Thinks in 10 Axes": Thought Factors × COG-MESH × Triple Stripes

Concept hook: An ordinary AI agent has only one kind of "thinking". llive
runs 10 kinds of thinking in parallel, makes them evaluate each other, and
takes only the surviving thoughts into the population. The 10 kinds are
"structurize", "recompose", "closed loop", "self-extend", "uncertainty",
"exploration", "consistency", "provenance", "multiview", and "reality link".
This compresses the major cognitive-science frameworks of the 1990s–2010s into
a single vector.

Today (2026-05-21) the marathon landed 1881 PASS + a large pull-forward of
v0.E. This article traces the "thought-factor side" of that — the intersection
of COG-MESH-01..10 and the historical persona ontology (CE-19).

0. Position within the series

#24-00 series index
#24-01 4-layer memory
#24-02 thought factors (10 axes) + COG-MESH (← this article)
#24-03 structural evolution × TRIZ × Z3
#24-04 B-series (fast cerebellum)
#24-05 EvolutionLoop (slow cerebrum)
#24-06 LLM backend non-transformer
#24-07 observability + governance
#24-08 lleval

The 10 thought factors + COG-MESH bind 1-to-N with the persona ontology (CE-19)
in #24-05. This article #24-02 sits at the position that explains them in terms
of "what" and "why".

1. Origin of the 10 thought factors — compression of 6 frameworks

A user-derived set of 10 axes (project_llive_cog_fx_factors). The source
material is the YouTube series "The Depths of Psychology" + cognitive-science
reviews + 6 frameworks from Polya / Six Hats / Bayesian / TRIZ / Provenance /
Multimodal. The result of compressing those into a single vector:

These are not orthogonal — for example, factor_uncertainty and
factor_exploration are correlated (UCB1 family). But by holding each one's
strength independently, the population can "attack the same problem with 10
different viewpoints".

2. Why hold 10 axes in a single vector?

In the LLM-agent literature, the mainstream view treats thinking as a single
kind of self-attention. llive extends that into multi-faceted thinking that is
switchable as a vector. This enables:

• "Thinking style" becomes computable via the inner product with a persona —
  for example, the "Oka Kiyoshi vector" holds (emotion) (Japanese-language
  ability) (multiple variables) high. The "Feynman vector" holds
  factor_exploration + factor_reality_link high.
• We can generate derived individuals that attack the same problem with
  different weightings.
• We can discover "which axis works for this problem" via the fitness
  gradient.

3. Deep dive into 5 major factors

3.1 factor_structurize — "Build up from axioms"

Axiomatic thinking. Mathematician-like (Galois / Grothendieck). Climbing the
abstraction ladder. Strength: generalization ability. Weakness: drifts away from
reality.

Within llive, the permutation of sub-blocks in BlockContainer corresponds to
a set of axioms. Derived individuals with high factor_structurize prefer
mutations that first split sub-blocks into required/optional and then
recompose them.

3.2 factor_recompose — "Swapping parts"

TRIZ Segmentation + synthesis. Rewrites the combination of existing parts.
Strength: fast local search. Weakness: no entirely new structure emerges.

In llive, PersonaImportAlgorithm (CE-20, landed today) is this axis. Derived
individual B partially adopts the persona of derived individual A. A hybrid
persona like "Galois + Oka Kiyoshi" emerges along the path that passes through
factor_recompose.

3.3 factor_closed_loop — "Watch yourself and fix yourself"

The core of cybernetics. Self-observation + self-correction. In llive, the memory
consolidation cycle (hippocampus → cortex) and the Approval Bus are this axis.
The E.4 governance (CE-06/07/08, landed today) — which evaluates within the
population so an individual sees the result and reflects it in the next
generation — also rides on this.

3.4 factor_uncertainty — "Quantify what you don't know"

Bayesian / probability. Strength: avoids overconfidence. Weakness:
computationally heavy. In llive, the verdict computation of the Approval Bus +
the UCB1 exploration constant are representative.

3.5 factor_provenance — "Where it came from"

Data lineage. Ed25519 sign + SHA-256 audit chain. Landed in llive Phase 4
(Production Security MVR, v0.3.0). This is a mandatory axis of agent
governance, and it was missing from conventional LLM agents.

4. Mapping to COG-MESH-01..10

project_cog_mesh_implementation_2026_05_19. Each of the 10 factors pairs with
one mechanism:

COG-MESH-10 landed today in the marathon as CoevolutionGovernance. This
completes the 10 mechanisms → 10 factors 1-1 mapping. We can now reverse-look-up
which factor is thin within the population from the state of the mechanisms.

5. Latest results (landed today, 2026-05-21)

In particular, finally being able to close COG-MESH-10 with the E.4 governance
skeleton was today's biggest landing. With this, the 10 factors ↔ 10 mechanisms
1-1 mapping is complete, and evaluation of the derived population → collusion
detection → Approval Bus integration is now connected at the architecture
level.

6. Expectations — what comes next

6.1 CE-19 Historical Persona Ontology (short term)

Already 10 names (Oka Kiyoshi / Grothendieck / Feynman / Galois / von Neumann /
Newton / Kant / Socrates / Lao Tzu / Sun Tzu) have landed as PERSONA_ONTOLOGY.
Today the CE-23 PersonaCorpusLoader skeleton landed, opening the way to
automatically extract personas from the Raptor RAD corpus to expand
PERSONA_ONTOLOGY. In the next session we plan to implement LLM extraction +
traversal of real RAD paths and expand the persona count to 30+.

6.2 Triple stripes (mid term, user-articulated)

"Triple stripes" = a state in which the 3 layers of thought factors / persona /
thinking process run in parallel within an individual like a striped pattern.
This was inspired by the "parallel cognition" hypothesis in cognitive
science. We run the factor vector + persona composition + Six Hats / TRIZ / ARIZ
each on a separate layer, and they critique each other in the within-population
evaluation. Landing time TBD.

6.3 Neural-interface support (long term)

project_llmesh_neuro_long_term. We have already added 6 fields to Raptor RAD:
bci / neuroscience / neural_signal / prosthetic_neural / cognitive_ai /
neuromorphic. This is preemptively gathering material so that we can expand
immediately when a "direct brain ↔ AI interface" becomes necessary. No direct
implementation for the time being.

7. Honest disclosure

• "The 10 factors overlap" — factor_uncertainty and factor_exploration
  correlate at about 0.65. They are not orthogonal to each other. At one point we
  considered collapsing to 9 axes, but we kept it at 10 for clarity.
• "The factor_affinity numbers are heuristics" — the factor_affinity vectors
  of the 10 PERSONA_ONTOLOGY names are artificial initial values based on
  biographies / the history of philosophy. They will later be replaced with
  corpus-based values by PersonaCorpusLoader (CE-23), but the current numbers
  are human rules of thumb.
• "COG-MESH-10 is a skeleton" — the E.4 governance that landed today is at
  the interface-establishment stage; the actual writing to Quarantined Memory
  is delegated to another module. It will take another 1-2 sessions to complete.

8. Mermaid — structure of the 10 factors

9. References (excerpted from 20+)

• Polya, G. (1945). How to Solve It.
• Altshuller, G. (1971). TRIZ 40 inventive principles.
• Auer, P. et al. (2002). Finite-time analysis of the multiarmed bandit.
• Lehman, J. & Stanley, K. (2008). Exploiting novelty.
• Mouret, J.-B. & Clune, J. (2015). Illuminating search spaces by mapping elites.
• Hillis, W. D. (1990). Coevolving parasites improve simulated evolution.
• Constitutional AI (Anthropic 2022) — for HITL alternative.
• Six Thinking Hats (De Bono 1985).
• 岡潔『春宵十話』.
• ファインマン『ご冗談でしょう, ファインマンさん』.
• Maturana & Varela — Autopoiesis.
• Bayes — Essay towards solving a problem in the doctrine of chances.
• The full list will be bundled in references.bib at the v0.6.0a1 release.

10. 2026-05-22 addendum — Rust port of the 10-factor affinity vector (RUST-15)

The 10 thought factors are implemented as a 10-dimensional [0,1] vector inside a
derived individual's persona composition's effective_factor_affinity. The
dissimilarity computation between derived individuals connects directly to the
core mechanism of this article #24-02 — PersonaOverlapPenalty.apply (E.17)
measures the distance in the 10-factor space via persona_dissimilarity over
N×N pairs.

Today (2026-05-22), as RUST-15, we did a batch (NxN pairs in a single FFI
call) Rust port:

• single 1-pair: x0.80 (FAIL — FFI overhead loses to Python set operations)
• batch N=64: x17.07 (PASS), average x12.71

This speeds up the "N×N pair distance computation of the 10-factor vector",
giving us a path to running governance + diversity preservation at 64 Hz for a
population of N=64.

10.1 Meaning seen from the thought-factor side

• factor_structurize (#0) and factor_exploration (#5) are two axes that
  conflict in the TRIZ family, but as an L2 distance in the 10-dimensional
  vector they take effect independently.
• When PersonaOverlapPenalty (E.17 CE-25) penalizes persona overlap within the
  population, the derived population naturally spreads out in the 10-factor
  space.
• The MAP-Elites grid (E.17 CE-26) is a 4-dimensional grid of persona 2 axes ×
  thought_factor 2 axes, so we marginalize the above 10-factor vector to 4
  dimensions and use it as the cell key.

10.2 Honest disclosure — a one-off Rust port backfires

When you hear "Rust-port the distance computation of the thought-factor vector",
you tend to think "it gets faster", but for a 1-pair computation Python is
faster due to FFI overhead (x0.80). This is pattern A in the
feedback_rust_usage_matters decision table (a pure-Python loop, 1-pair). Only by
packing N×N pairs into a single FFI in a batch does it stretch to x17.07.

For details see #24-05 and
docs/perf_comparison/2026-05-22_kernel_implementation_comparison.md.

Series Navigation

• ← Prev: llive Complete Guide (1) "The LLM that Never Forgets"
• → Next: llive Complete Guide (3) "Contradictions Can Be Computed"
• All: llive Complete Guide (0) — series index
• repo: furuse-kazufumi/llive

中文

llive 完全解说 (2) — "用 10 个轴思考的 AI": 思考因子 × COG-MESH × 三重条纹

概念 hook: 普通的 AI agent 只有 1 种"思考". llive 同时运行 10 种思考,
让它们相互评价, 只把存活下来的思考纳入群体. 这 10 种是"结构化""重组"
"闭环""自我扩展""不确定性""探索""一致性""来历""多视角""现实连接".
这是把认知科学 1990s〜2010s 的主要 framework 压缩到 1 个 vector 中的产物.

今天 (2026-05-21) 的 marathon 落地了 1881 PASS + v0.E 的大规模前置完成. 本文
追溯其"思考因子侧" — COG-MESH-01〜10 与 historical persona ontology (CE-19)
的交叉点.

0. 在连载中的定位

#24-00 series index
#24-01 4 层记忆
#24-02 思考因子 10 轴 + COG-MESH (← 本文)
#24-03 结构进化 × TRIZ × Z3
#24-04 B-series (快速小脑)
#24-05 EvolutionLoop (缓慢大脑)
#24-06 LLM backend non-transformer
#24-07 observability + governance
#24-08 lleval

10 思考因子 + COG-MESH 与 #24-05 的 persona ontology (CE-19) 以 1-N 方式结合.
本文 #24-02 处于用"是什么"和"为什么"来解释它的位置.

1. 10 思考因子的由来 — 6 个 framework 的压缩

源自用户的 10 个轴 (project_llive_cog_fx_factors). 原始素材是
"心理的深层" YouTube + 认知科学评论 + Polya / Six Hats / Bayesian / TRIZ /
Provenance / Multimodal 系的 6 个 framework. 将它们压缩到 1 个 vector 后的结果:

这些 并非正交 — 例如 factor_uncertainty 和 factor_exploration 是相关的
(UCB1 系). 但通过独立持有各自的 强度, 群体内就可以"用 10 种视角面对同一个
问题".

2. 为什么把 10 个轴放在 1 个 vector 中

在 LLM agent 的文献中,"思考是 1 种 self-attention"是主流. llive 将其扩展为
可作为 vector 切换的 multi-faceted thinking. 由此:

• 通过与 persona 的内积可以计算出"思考风格" — 例如"冈洁向量"在 (情绪)
  (国语能力) (多变量) 上取值较高."费曼向量"在 factor_exploration +
  factor_reality_link 上取值较高.
• 可以生成在群体内 以不同权重 面对同一个问题的派生个体.
• 可以通过 fitness gradient 发现"这个问题哪个轴有效".

3. 5 个主要因子的深入解读

3.1 factor_structurize — "从公理往上搭建"

axiomatic 的思考. 数学家式 (伽罗瓦 / 格罗滕迪克). 攀爬抽象阶梯.
优点: 一般化能力. 缺点: 脱离现实.

在 llive 内, BlockContainer 的 sub-block 排列对应公理群. factor_structurize
较高的派生个体偏好先把 sub-block 分为 必需/可选 然后再重组的 mutation.

3.2 factor_recompose — "部件的替换"

TRIZ Segmentation + 合成. 重写既有部件的组合. 优点: 局部搜索快速.
缺点: 不会产生全新的结构.

在 llive 中, PersonaImportAlgorithm (CE-20, 今天落地) 就是这个轴. 派生 B
部分采用 派生 A 的 persona. 像"伽罗瓦 + 冈洁"这样的 hybrid persona 出现的
路径正是经过 factor_recompose.

3.3 factor_closed_loop — "看着自己来修正"

cybernetics 的核心. 自我观察 + 自我修正. 在 llive 中, memory consolidation
cycle (海马体→皮质) 和 Approval Bus 就是这个轴. 在群体内评价 → 个体看到结果并
反映到下一代的 E.4 governance (CE-06/07/08, 今天落地) 也搭载在这里.

3.4 factor_uncertainty — "把不知道量化"

Bayesian / probability. 优点: 避免过度自信. 缺点: 计算量大.
在 llive 中, Approval Bus 的 verdict 计算 + UCB1 exploration constant 是代表.

3.5 factor_provenance — "从哪里来的"

data lineage. Ed25519 sign + SHA-256 audit chain. 在 llive Phase 4 (Production
Security MVR, v0.3.0) 落地. 这是 agent governance 的 必备轴, 而传统的
LLM agent 中是缺失的.

4. 与 COG-MESH-01〜10 的对应

project_cog_mesh_implementation_2026_05_19. 10 个因子各自与 1 个机制 对应:

COG-MESH-10 今天在 marathon 中作为 CoevolutionGovernance 落地. 由此,
10 机制 → 10 因子 1-1 对应完成. 现在可以从机制的状态反向查出群体内
哪个因子较薄弱.

5. 最新成果 (今天 2026-05-21 落地)

特别是用 E.4 governance skeleton 终于能够让 COG-MESH-10 收口, 是今天最大的
成果. 由此 10 因子 ↔ 10 机制 1-1 对应完成, 派生群体的评价 → 共谋检测 →
Approval Bus 联动 在 architecture level 连通了.

6. 期望值 — 接下来要做的

6.1 CE-19 Historical Persona Ontology (短期)

已经有 10 位 (冈洁 / 格罗滕迪克 / 费曼 / 伽罗瓦 / 冯·诺依曼 / 牛顿 / 康德 /
苏格拉底 / 老子 / 孙子) 作为 PERSONA_ONTOLOGY 落地. 今天 CE-23
PersonaCorpusLoader skeleton 落地, 开辟了 从 Raptor RAD 语料库自动抽取
persona 来扩展 PERSONA_ONTOLOGY 的道路. 下一个 session 将实现 LLM 抽取 +
真实 RAD path 横跨, 计划把 persona 数量扩大到 30+.

6.2 三重条纹 (中期, 用户语言化)

"三重条纹" = 思考因子 / persona / 思考过程 这 3 层在个体内像条纹一样同时
运行的状态. 这受到认知科学 "并行认知" 假说的启发. 把 factor vector +
persona composition + Six Hats / TRIZ / ARIZ 分别放在不同 layer 上运行, 在群体内
evaluation 中相互批评. 落地时间未定.

6.3 神经接口对应 (长期)

project_llmesh_neuro_long_term. 已经在 Raptor RAD 中追加了 bci / neuroscience
/ neural_signal / prosthetic_neural / cognitive_ai / neuromorphic 这 6 个领域.
这是为了当"脑 ↔ AI 直连接口"成为必要时能够立即 expand 而提前收集素材.
暂时没有直接的实现.

7. honest disclosure (诚实披露)

• "10 个因子存在 overlap" — factor_uncertainty 和 factor_exploration 的
  相关性约为 0.65. 彼此并非正交. 曾经也考虑过压缩到 9 个轴, 但出于易懂优先
  保持在 10 个.
• "factor_affinity 的数值是 heuristic" — PERSONA_ONTOLOGY 10 位的
  factor_affinity vector 是基于传记 / 哲学史 的人为初始值. 之后会由
  PersonaCorpusLoader (CE-23) 替换为基于语料库的值, 但目前的数值是人的
  经验法则.
• "COG-MESH-10 是 skeleton" — 今天落地的 E.4 governance 处于接口确立阶段,
  对 Quarantined Memory 的 实际写入 委托给另一个 module. 到完成还需要再
  1-2 个 session.

8. Mermaid — 10 个因子的结构

9. References (从主要 20+ 中精选)

• Polya, G. (1945). How to Solve It.
• Altshuller, G. (1971). TRIZ 40 inventive principles.
• Auer, P. et al. (2002). Finite-time analysis of the multiarmed bandit.
• Lehman, J. & Stanley, K. (2008). Exploiting novelty.
• Mouret, J.-B. & Clune, J. (2015). Illuminating search spaces by mapping elites.
• Hillis, W. D. (1990). Coevolving parasites improve simulated evolution.
• Constitutional AI (Anthropic 2022) — for HITL alternative.
• Six Thinking Hats (De Bono 1985).
• 岡潔『春宵十話』.
• ファインマン『ご冗談でしょう, ファインマンさん』.
• Maturana & Varela — Autopoiesis.
• Bayes — Essay towards solving a problem in the doctrine of chances.
• 完整列表将在 v0.6.0a1 发布时随 references.bib 一同提供.

10. 2026-05-22 追记 — 10 因子 affinity vector 的 Rust 化 (RUST-15)

10 个思考因子作为派生个体的 persona composition 的 effective_factor_affinity,
以 10 维 [0,1] vector 实现. 派生个体之间的 dissimilarity 计算与本文 #24-02 的
核心机制直接相连 — PersonaOverlapPenalty.apply (E.17) 通过 N×N pairs 的
persona_dissimilarity 测量 10 因子空间中的距离.

今天 (2026-05-22) 作为 RUST-15 进行了 batch (把 NxN pair 装进 1 次 FFI call)
的 Rust 化:

• single 1-pair: x0.80 (FAIL — FFI overhead 输给了 Python 的 set 操作)
• batch N=64: x17.07 (PASS), 平均 x12.71

由此"10 因子 vector 的 N×N pair 距离计算"得到加速, 为在群体 N=64 下以
64 Hz 运行 governance + diversity preservation 提供了可行的眉目.

10.1 从思考因子侧看到的意义

• factor_structurize (#0) 和 factor_exploration (#5) 是 在 TRIZ 系统中
  对立的 2 个轴, 但作为 10 维 vector 的 L2 距离则独立起作用.
• 用 PersonaOverlapPenalty (E.17 CE-25) 惩罚群体内的 persona overlap 时,
  派生群体会在 10 因子空间中自然地散开.
• MAP-Elites grid (E.17 CE-26) 是 persona 2 轴 × thought_factor 2 轴 的 4 维
  grid, 所以把上述 10 因子 vector marginalize 到 4 维并作为 cell key.

10.2 honest disclosure — 单次 Rust 化适得其反

听到"把思考因子 vector 的距离计算 Rust 化"时, 容易以为"会变快", 但
在 1-pair 计算中由于 FFI overhead Python 反而更快 (x0.80). 这是
feedback_rust_usage_matters 判定表中的 A 模式 (纯 Python 循环 1-pair).
只有用 batch 把 N×N pair 装进 1 次 FFI, 才会一路提升到 x17.07.

详情参见 #24-05 和
docs/perf_comparison/2026-05-22_kernel_implementation_comparison.md.

Series Navigation

• ← 上一篇: llive 完全解说 (1) 「不会遗忘的 LLM」
• → 下一篇: llive 完全解说 (3) 「矛盾是可以计算的」
• 全部: llive 完全解说 (0) — series index
• repo: furuse-kazufumi/llive

한국어

llive 완전 해설 (2) — "10개 축으로 사고하는 AI": 사고 인자 × COG-MESH × 삼중 줄무늬

콘셉트 hook: 보통의 AI agent는 "사고"를 1종류밖에 가지지 않는다. llive는
10종류의 사고를 동시에 실행시키고, 그것들을 서로 평가하게 하여,
살아남은 사고만을 집단에 받아들인다. 10종은 "구조화" "재구성" "폐루프"
"자기 확장" "불확실성" "탐색" "정합" "내력" "다관점" "현실 연결". 이는 인지
과학 1990s〜2010s의 주요 framework를 1개의 vector로 압축한 것이다.

오늘 (2026-05-21) marathon에서 1881 PASS + v0.E 대규모 앞당김이 착지했다. 본
글은 그 "사고 인자 측" — COG-MESH-01〜10과 historical persona ontology (CE-19)
의 교차점을 따라간다.

0. 연재에서의 위치

#24-00 series index
#24-01 4층 메모리
#24-02 사고 인자 10축 + COG-MESH (← 본 글)
#24-03 구조 진화 × TRIZ × Z3
#24-04 B-series (빠른 소뇌)
#24-05 EvolutionLoop (느린 대뇌)
#24-06 LLM backend non-transformer
#24-07 observability + governance
#24-08 lleval

10가지 사고 인자 + COG-MESH는 #24-05의 persona ontology (CE-19)와 1-N으로
결합한다. 본 글 #24-02는 그것을 "무엇"과 "왜"로 설명하는 위치다.

1. 10가지 사고 인자의 유래 — 6개 framework의 압축

사용자에게서 유래한 10개 축 (project_llive_cog_fx_factors). 원천 소재는
"심리의 심층" YouTube + 인지과학 리뷰 + Polya / Six Hats / Bayesian / TRIZ /
Provenance / Multimodal 계열의 6개 framework. 그것을 1개의 vector로 압축한 결과:

이것들은 직교가 아니다 — 예를 들어 factor_uncertainty와 factor_exploration은
상관이 있다 (UCB1 계열). 하지만 각각의 강도를 독립적으로 가짐으로써, 집단
내에서 "같은 문제에 10가지 관점으로 부딪힌다"가 가능해진다.

2. 왜 10개 축을 1개의 vector에 담는가

LLM agent 문헌에서는 "사고는 self-attention 1종류"가 주류다. llive는 그것을
vector로 전환 가능한 multi-faceted thinking으로 확장했다. 이로써:

• persona와의 내적으로 "사고 스타일"을 계산 가능 — 예를 들어 "오카 기요시
  벡터"는 (정서) (국어력) (다변수)를 높게 가진다. "파인만 벡터"는
  factor_exploration + factor_reality_link를 높게 가진다.
• 집단 내에서 같은 문제에 서로 다른 가중치로 부딪히는 파생 개체를 생성할 수
  있다.
• "이 문제는 어떤 축이 효과적인가"를 fitness gradient로 발견할 수 있다.

3. 주요 인자 5개의 심화

3.1 factor_structurize — "공리에서 쌓아 올린다"

axiomatic한 사고. 수학자 갈루아 / 그로텐디크 식. 추상화 계단을 오른다.
장점: 일반화 능력. 단점: 현실에서 멀어진다.

llive 내에서는 BlockContainer의 sub-block 순열이 공리군에 대응한다.
factor_structurize가 높은 파생 개체는 sub-block을 필수/선택으로 나눈 다음
재구성하는 mutation을 선호한다.

3.2 factor_recompose — "부품의 교체"

TRIZ Segmentation + 합성. 기존 부품의 조합을 다시 쓴다. 장점: 국소 탐색 고속.
단점: 완전히 새로운 구조는 생기지 않는다.

llive에서는 PersonaImportAlgorithm (CE-20, 오늘 착지)이 이 축이다. 파생 B가
파생 A의 persona를 부분 채용한다. "갈루아 + 오카 기요시" 같은 hybrid
persona가 출현하는 것은 factor_recompose를 거치는 경로다.

3.3 factor_closed_loop — "자신을 보고 고친다"

cybernetics의 핵심. 자기 관찰 + 자기 수정. llive에서는 memory consolidation
cycle (해마→피질)과 Approval Bus가 이 축이다. 집단 내에서 평가 → 개체가 결과를
보고 다음 세대에 반영하는 E.4 governance (CE-06/07/08, 오늘 착지)도 여기에
실린다.

3.4 factor_uncertainty — "모름을 정량화한다"

Bayesian / probability. 장점: 과신을 피한다. 단점: 계산이 무겁다.
llive에서는 Approval Bus의 verdict 계산 + UCB1 exploration constant가 대표적이다.

3.5 factor_provenance — "어디에서 왔는가"

data lineage. Ed25519 sign + SHA-256 audit chain. llive Phase 4 (Production
Security MVR, v0.3.0)에서 착지. 이는 agent governance의 필수 축이며, 기존의
LLM agent에는 결여되어 있었다.

4. COG-MESH-01〜10과의 대응

project_cog_mesh_implementation_2026_05_19. 10개 인자에 1개 기구씩 대응한다:

COG-MESH-10은 오늘 marathon에서 CoevolutionGovernance로 착지했다. 이로써
10 기구 → 10 인자 1-1 대응이 완성되었다. 이제 집단 내에서 어떤 인자가 얇은지를
기구의 상태로부터 역추적할 수 있게 되었다.

5. 최신 성과 (오늘 2026-05-21 착지)

특히 E.4 governance skeleton으로 COG-MESH-10을 closing할 수 있었던 것이
오늘의 최대 성과다. 이로써 10 인자 ↔ 10 기구 1-1 대응이 완성되어, 파생 집단의
평가 → 공모 탐지 → Approval Bus 연동이 architecture level에서 연결되었다.

6. 기대값 — 다음에 올 것

6.1 CE-19 Historical Persona Ontology (단기)

이미 10명 (오카 기요시 / 그로텐디크 / 파인만 / 갈루아 / 폰 노이만 / 뉴턴 / 칸트
/ 소크라테스 / 노자 / 손자)이 PERSONA_ONTOLOGY로 착지 완료. 오늘 CE-23
PersonaCorpusLoader skeleton이 착지하여, Raptor RAD 코퍼스에서 persona를 자동
추출해 PERSONA_ONTOLOGY를 확장하는 길이 열렸다. 다음 세션에서 LLM 추출 + 실제
RAD path 횡단을 구현해 persona 수를 30+로 확대할 예정.

6.2 삼중 줄무늬 (중기, 사용자 언어화)

"삼중 줄무늬" = 사고 인자 / persona / 사고 프로세스의 3개 층이 개체 내에서
줄무늬처럼 동시에 실행되는 상태. 이는 인지과학의 "병렬 인지" 가설에서 착상을
얻은 것이다. factor vector + persona composition + Six Hats / TRIZ / ARIZ를 각각
다른 layer에서 실행하고, 집단 내 evaluation에서 서로를 비평한다. 착지 시기 미정.

6.3 신경 인터페이스 대응 (장기)

project_llmesh_neuro_long_term. Raptor RAD에 bci / neuroscience /
neural_signal / prosthetic_neural / cognitive_ai / neuromorphic의 6개 분야를
추가 완료. 이는 "뇌 ↔ AI 직결 인터페이스"가 필요해졌을 때 즉시 expand할 수
있도록 미리 소재를 모아 두는 것이다. 직접적인 구현은 당분간 없다.

7. honest disclosure (정직한 공개)

• "10개 인자에는 overlap이 있다" — factor_uncertainty와 factor_exploration은
  상관이 0.65 정도. 서로 직교가 아니다. 9 axis화를 검토한 시기도 있었지만
  알기 쉬움을 우선하여 10개 그대로 유지.
• "factor_affinity의 수치는 heuristic" — PERSONA_ONTOLOGY 10명의
  factor_affinity vector는 전기 / 철학사 기반의 인위적 초기값. 이후
  PersonaCorpusLoader (CE-23)로 코퍼스 기반으로 치환되지만, 현재의 수치는
  사람에 의한 경험칙이다.
• "COG-MESH-10은 skeleton" — 오늘 착지한 E.4 governance는 interface 확립
  단계이며, Quarantined Memory로의 실제 기록은 다른 module에 위임. 완성까지는
  앞으로 1-2 세션 걸린다.

8. Mermaid — 10개 인자의 구조

9. References (주요 20+ 중 발췌)

• Polya, G. (1945). How to Solve It.
• Altshuller, G. (1971). TRIZ 40 inventive principles.
• Auer, P. et al. (2002). Finite-time analysis of the multiarmed bandit.
• Lehman, J. & Stanley, K. (2008). Exploiting novelty.
• Mouret, J.-B. & Clune, J. (2015). Illuminating search spaces by mapping elites.
• Hillis, W. D. (1990). Coevolving parasites improve simulated evolution.
• Constitutional AI (Anthropic 2022) — for HITL alternative.
• Six Thinking Hats (De Bono 1985).
• 岡潔『春宵十話』.
• ファインマン『ご冗談でしょう, ファインマンさん』.
• Maturana & Varela — Autopoiesis.
• Bayes — Essay towards solving a problem in the doctrine of chances.
• 완전한 목록은 v0.6.0a1 릴리스 시 references.bib에 동봉할 예정.

10. 2026-05-22 추기 — 10 인자 affinity vector의 Rust화 (RUST-15)

10가지 사고 인자는 파생 개체의 persona composition의 effective_factor_affinity
로서 10차원 [0,1] vector로 구현되어 있다. 파생 개체 간의 dissimilarity 계산은
본 글 #24-02의 핵심 기구와 직결된다 — PersonaOverlapPenalty.apply (E.17)는
N×N pairs의 persona_dissimilarity로 10 인자 공간의 거리를 측정한다.

오늘 (2026-05-22) RUST-15로서 batch (NxN pair를 1 FFI call로) Rust화:

• single 1-pair: x0.80 (FAIL — FFI overhead로 Python set 연산에 진다)
• batch N=64: x17.07 (PASS), 평균 x12.71

이로써 "10 인자 vector의 N×N pair 거리 계산"이 고속화되어, 집단 N=64에서
governance + diversity preservation을 64 Hz로 돌릴 수 있는 가닥이 잡혔다.

10.1 사고 인자 측에서 본 의미

• factor_structurize (#0)과 factor_exploration (#5)은 TRIZ 계통에서 대립하는
  2개 축이지만, 10차원 vector의 L2 거리로는 독립적으로 작용한다.
• PersonaOverlapPenalty (E.17 CE-25)로 집단 내 persona overlap을 벌하면,
  파생 집단은 10 인자 공간에서 자연스럽게 흩어진다.
• MAP-Elites grid (E.17 CE-26)는 persona 2축 × thought_factor 2축의 4차원
  grid이므로, 위의 10 인자 vector를 4차원으로 marginalize하여 cell key로
  삼는다.

10.2 honest disclosure — 단발 Rust화는 역효과

"사고 인자 vector의 거리 계산을 Rust화"라고 들으면 "빨라진다"고 생각하기 쉽지만,
1-pair 계산에서는 FFI overhead로 Python 쪽이 더 빠르다 (x0.80). 이는
feedback_rust_usage_matters 판정표의 A 패턴 (순수 Python 루프 1-pair)이다.
batch로 N×N pair를 1 FFI에 담아야 비로소 x17.07까지 늘어난다.

자세한 내용은 #24-05와
docs/perf_comparison/2026-05-22_kernel_implementation_comparison.md 참조.

Series Navigation

• ← 이전: llive 완전 해설 (1) 「잊지 않는 LLM」
• → 다음: llive 완전 해설 (3) 「모순은 계산할 수 있다」
• 전체: llive 완전 해설 (0) — series index
• repo: furuse-kazufumi/llive