170 lines
5.5 KiB
Markdown
170 lines
5.5 KiB
Markdown
|
|
---
|
|||
|
|
license: apache-2.0
|
|||
|
|
language:
|
|||
|
|
- ja
|
|||
|
|
- en
|
|||
|
|
base_model:
|
|||
|
|
- google/gemma-2-2b-it
|
|||
|
|
pipeline_tag: text-generation
|
|||
|
|
tags:
|
|||
|
|
- axis
|
|||
|
|
- sovereign-logic
|
|||
|
|
- logic-engine
|
|||
|
|
- determinism
|
|||
|
|
---
|
|||
|
|
---
|
|||
|
|
language:
|
|||
|
|
- ja
|
|||
|
|
- en
|
|||
|
|
license: other
|
|||
|
|
library_name: transformers
|
|||
|
|
tags:
|
|||
|
|
- axis
|
|||
|
|
- gemma-2
|
|||
|
|
- logic-engine
|
|||
|
|
- sovereign-ai
|
|||
|
|
datasets:
|
|||
|
|
- custom
|
|||
|
|
metrics:
|
|||
|
|
- logical_consistency
|
|||
|
|
---
|
|||
|
|
App Store
|
|||
|
|
https://apps.apple.com/jp/app/verantyx-logic/id6757994077
|
|||
|
|
💠 AXIS: Advanced Cross-Integrated System (V1.6)
|
|||
|
|
── 知能主権の確立と決定論的演算のための統治エンジン ──
|
|||
|
|
|
|||
|
|
🧩 AXIS の工学的定義
|
|||
|
|
AXISは、AIを「非決定的な出力を生成するブラックボックス(旋盤)」として扱い、その外側に「決定論的な検証器(Verifier)」を置くことで、出力を完全に統治するアーキテクチャです。
|
|||
|
|
|
|||
|
|
1. 旋盤アーキテクチャと検証プロトコル
|
|||
|
|
|
|||
|
|
AIユニットは、高次元データから論理パーツを削り出すための**「旋盤(Lathe)」**です。
|
|||
|
|
|
|||
|
|
リジェクト・ループ: AIが提案した解は、外部検証器(Python/Sympy等)が制約式(Constraints)に基づき判定。1bitでも矛盾があれば即座に棄却(Reject)し、Session IDを更新して再生成を強制します。
|
|||
|
|
|
|||
|
|
物理パージ (Context Reset): torch.mps.empty_cache() を実行し、直前の「失敗した思考」というキャッシュを物理的に消去。各試行を統計的に独立させ、ハルシネーションの連鎖(Context Drift)を断ち切ります。
|
|||
|
|
|
|||
|
|
2. 立体十字(3D Semantic Lattice)の座標管理
|
|||
|
|
|
|||
|
|
各ノードは、相互に独立(直交)することを目指した 5 次元軸 (s
|
|||
|
|
1
|
|||
|
|
|
|||
|
|
…s
|
|||
|
|
5
|
|||
|
|
|
|||
|
|
) で管理される SemanticNode クラスとして実装されます。
|
|||
|
|
|
|||
|
|
s
|
|||
|
|
1
|
|||
|
|
|
|||
|
|
: 物理的実体性(数値・定数との整合性)
|
|||
|
|
|
|||
|
|
s
|
|||
|
|
2
|
|||
|
|
|
|||
|
|
: 論理的必然性(公理系からの導出可能性)
|
|||
|
|
|
|||
|
|
s
|
|||
|
|
3
|
|||
|
|
|
|||
|
|
: 文脈依存性(Context Stackとの一致率)
|
|||
|
|
|
|||
|
|
s
|
|||
|
|
4
|
|||
|
|
|
|||
|
|
: 倫理性スコア(安全規約への適合度)
|
|||
|
|
|
|||
|
|
s
|
|||
|
|
5
|
|||
|
|
|
|||
|
|
: 実証履歴(過去の確定データとの合致回数)
|
|||
|
|
|
|||
|
|
3. 論理の永続化と高速化の正体
|
|||
|
|
|
|||
|
|
意味ID (Semantic ID): 入力クエリを Embedding 空間へ投影し、ベクトル量子化(Vector Quantization)によって生成される固有のハッシュ値です。
|
|||
|
|
|
|||
|
|
高速化の根拠: local_massive_data.json は、この意味IDをキーとした高密度なキャッシュとして機能します。AIの全推論プロセスをスキップし、検証済みの「真理」を直接 O(1) で参照するため、推論時間を物理的にゼロへと近似させます(※推論実行時比較比)。
|
|||
|
|
|
|||
|
|
🚀 革新的な特徴 (V1.6 実装仕様)
|
|||
|
|
Deterministic Assembly(決定論的アセンブル)
|
|||
|
|
|
|||
|
|
最終回答はAIの作文ではなく、検証済みの Raw Data を、システムが保持する Adherents(言語テンプレート) によって物理的に結合します。
|
|||
|
|
|
|||
|
|
Example:
|
|||
|
|
|
|||
|
|
Raw Data: {"ans": "z^5", "a": 0}
|
|||
|
|
|
|||
|
|
Adherent: "The solution is {ans} (a={a})."
|
|||
|
|
|
|||
|
|
Output: "The solution is z^5 (a=0)." これにより、回答段階でのハルシネーションの混入を 0% に抑えます。
|
|||
|
|
|
|||
|
|
🛠 Setup & Roadmap
|
|||
|
|
Micro-MVP 公開(予定)
|
|||
|
|
|
|||
|
|
近日中に minimal_example.py を公開。以下の動作を証明します:
|
|||
|
|
|
|||
|
|
ComplexVerifier: 数学的制約によるAI出力の拒絶
|
|||
|
|
|
|||
|
|
RejectionLoop: AIと検証器の実際の往復回数の可視化
|
|||
|
|
|
|||
|
|
SessionPurge: メモリクリアによるハルシネーション抑制の検証
|
|||
|
|
|
|||
|
|
⚖️ License (APSL v1.0)
|
|||
|
|
商用模倣(Rejection-based Governance Logicの利用)を禁じ、知能の主権を個人の手に留めます。
|
|||
|
|
|
|||
|
|
© 2025 AXIS Project. All rights reserved. STATUS: TOWARD_MVP_IMPLEMENTATION.
|
|||
|
|
|
|||
|
|
💠 AXIS: Advanced Cross-Integrated System (V1.6) - English Edition
|
|||
|
|
── Establishing Intelligence Sovereignty via Deterministic Governance ──
|
|||
|
|
|
|||
|
|
🧩 Technical Definition
|
|||
|
|
AXIS treats AI as a non-deterministic generator (Lathe) while utilizing a deterministic Verifier to maintain total sovereignty over the output.
|
|||
|
|
|
|||
|
|
1. The Lathe & Rejection Protocol
|
|||
|
|
|
|||
|
|
The Rejection Loop: AI solutions are scanned by external verifiers (Python/SymPy). Any contradiction results in an immediate REJECT, session reset, and re-forgery.
|
|||
|
|
|
|||
|
|
Context Purge: empty_cache() physically incinerates the "failed reasoning" from VRAM, ensuring statistical independence between trials and severing the chain of "hallucination drift."
|
|||
|
|
|
|||
|
|
2. 5D Semantic Lattice Implementation
|
|||
|
|
|
|||
|
|
Nodes are managed via the SemanticNode class, utilizing five orthogonal parameters:
|
|||
|
|
|
|||
|
|
s
|
|||
|
|
1
|
|||
|
|
|
|||
|
|
: Physical Actuality | s
|
|||
|
|
2
|
|||
|
|
|
|||
|
|
: Logical Necessity | s
|
|||
|
|
3
|
|||
|
|
|
|||
|
|
: Contextual Dependency | s
|
|||
|
|
4
|
|||
|
|
|
|||
|
|
: Ethical Score | s
|
|||
|
|
5
|
|||
|
|
|
|||
|
|
: Empirical History.
|
|||
|
|
|
|||
|
|
3. Persistence & Acceleration Mechanism
|
|||
|
|
|
|||
|
|
Semantic ID: A unique hash generated via Vector Quantization in the embedding space.
|
|||
|
|
|
|||
|
|
Acceleration: By searching local_massive_data.json first, AXIS skips the entire AI inference process for known truths, achieving near-zero latency compared to standard LLM execution.
|
|||
|
|
|
|||
|
|
🚀 Core Features (V1.6)
|
|||
|
|
Deterministic Assembly
|
|||
|
|
|
|||
|
|
Responses are not "written" by AI; they are physically assembled by binding verified Raw Data into hard-coded Adherent Templates. This ensures 0% hallucination during the final response delivery.
|
|||
|
|
|
|||
|
|
🛠 Roadmap: Micro-MVP Launch
|
|||
|
|
We will soon release minimal_example.py to demonstrate:
|
|||
|
|
|
|||
|
|
Real-time mathematical rejection by ComplexVerifier.
|
|||
|
|
|
|||
|
|
Tracking of the RejectionLoop iterations.
|
|||
|
|
|
|||
|
|
Proof of SessionPurge efficacy in preventing context-drift.
|
|||
|
|
|
|||
|
|
© 2025 AXIS Project. STATUS: TOWARD_MVP_IMPLEMENTATION.
|