アーキテクチャ

スキルとナレッジ

MARIA OSがエージェントの能力、ナレッジライフサイクル、スキルガバナンスをどう管理するか。

スキル基盤

スキルはオンデマンドで取得または生成される

Skill Requirement Engineがエージェントの能力をタスク要件にマッチングする。不足するスキルは取得または生成をトリガーする。

skill-fetch

Dynamic Skill Fetch

Requirement resolved from existing Skill Store

Requirement
skill_key:
K4.verify
domain:
contract
min_level:
4
Skill Store
Risk Evaluation
K4.verify / contract
Cost Analysis
K3.infer / finance
Verifier-01Contract Verification
1

Requirement

Skill key, domain, level

2

Store Search

Query matching capability

3

Resolve/Create

Generate if not found

4

Bind to Agent

Phase Gate enforced

No HardcodingConsistent QualityInfinite Extensibility
権限モデル

スキル/フェーズ権限マトリクス

すべてのスキルがすべてのフェーズで利用可能なわけではない。権限マトリクスが各意思決定ステージで使用できるスキルを管理する。

P0
P1
P2
P3
P4
P5
K1
Collect
K2
Curate
K3
Infer
K4
Verify
K5
Plan
K6
Execute
K7
Audit
K8
Govern
AllowedGateBlocked|K6 blocked in observe, K8 requires authority
ライブ可視化

スキルがエージェントに流れる様子を観察

エージェントネットワーク全体でのスキル割り当てと消費のリアルタイム可視化。

Skill Store
K1Collect
K2Curate
K3Infer
K4Verify
K5Plan
K6Execute
K7Audit
K8Govern
Dynamic Refill Pipeline
Market Analyst

Market Analyst

Strategy

K1
K3
Risk Evaluator

Risk Evaluator

Compliance

K4
Data Processor

Data Processor

Operations

K2
Quality Auditor

Quality Auditor

Governance

K7
Refill Log

Waiting for refill events...

5

Total Skills Bound

0

Refills Complete

0

Active Transfers

Agents never stop. Skills arrive before they are needed.

キャパシティ管理

エージェントの実行力は枯渇しない

スキルは継続的に補充される。キャパシティ管理によりエージェントは常に必要なリソースを確保。

skill_ready
A

Capability Mismatch

Agent knows what to do but cannot execute at required level

B

Quality Degradation

Skill is fatigued or no longer fits current context

C

Phase Mismatch

Decision weight exceeds skill strength

D

Environment Change

Skill is correct but premises have shifted

Agents get tired. Environments change. Decisions get heavier.

MARIA OS refills skills, not replaces Agents.

トポロジー

Universe トポロジー

Universe、Planet、Zoneの構造的レイアウト — オペレーティングシステムの空間次元。

MARIA Universe3 planets
Common Infrastructure
Skill Registry
Policy
Routing
Executive Decision11 agents across 3 zones
Design Principles
Verifier and Auditor are independent
Executor always pairs with Phase Gate
Governor at Planet level for consistency
Orchestrator at Zone entry point
スキルフレームワーク

Skills Measure Design Adherence, Not Just Performance

Skills are observation instruments that verify whether a Universe executes according to its designed judgment structure. Not vanity metrics. Governance truth.

Health Skills
Verify the system operates without errors
Success Rate
98.7%
Latency (P95)
45ms
Reproducibility
99.2%
Schema Validity
100%
Stop Rate by Reason
Policy
4.2%Threshold violations
Safety
0.5%Safety rule triggers
Uncertainty
2.8%Low confidence stops

Stop rates are improvement signals, not failures

Risk Profile → Target
High Risk
SR: 99.5%Audit: 100%
Medium Risk
SR: 98%Audit: 99%
Low Risk
SR: 95%Audit: 95%

Targets adapt to risk tier and genesis phase

Every skill is linked to a ValueKey for automatic policy adjustment

Skills measure whether the Universe is betraying its design philosophy.

パフォーマンスプロファイル

Right Targets, Right Context

One size doesn't fit all. Target profiles adapt to your risk tier and data classification. Stop reasons tell you where to improve.

Baseline

low-medium

Conservative

medium-high

Strict

high

Aggressive

low

Baseline Profile
For new universes, standard operations
success
95%
automation
60%
audit
95%
Stop
<10%
Stop Rate Breakdown8.6% total
Policy
4.2%
Uncertainty
2.8%
Safety
0.5%
Data
1.1%
High-value approval required
Policy threshold violation
Improvement
Adjust policy thresholds

Every stop is a signal. Every profile is a strategy.

Post-Deployment

Execution Flow Circuit

Parallel, serial, and loop paths with HITL checkpoints

Trigger
Collect Agent
Parallel
Curate Agent
Verify Agent
Merge Results
Infer Agent
Human ReviewHITL
Plan Agent
Compliance Gate
Execute Agent
Quality Loop
back to Infer
Audit Agent
Final ApprovalHITL
Complete

Flow Legend

Agent
HITL
Parallel
Loop
Gate
Decision

HITL Checkpoints

Human Review
Final Approval
Parallel execution for efficiency
Loop circuits for quality
Human gates clearly marked
ナレッジ基盤
オフラインナレッジパック

信頼できるナレッジ、単に利用可能なだけではない

Offline knowledge is not a convenience feature. It is judgment infrastructure. Three guarantees: content is correct, provenance is traceable, time decay is visible.

5層アーキテクチャ

Package

Evidence

Interpretation

Constraint

Integrity

Package MetadataIdentity & scope
package_idversionscopetime_basis
ポイントインタイムモード
Current

Past decisions can be replayed with their original knowledge context.

信頼性指標

Financial Compliance v2.3

K-2024-001 | basis: 2024-04

142d left

valid

HR Policy Guidelines

K-2024-002 | basis: 2024-01

28d left

warning

Vendor Assessment Criteria

K-2023-015 | basis: 2023-12

14d overdue

expired

整合性チェック
Entry Hashsha256:8f2a...c4d1
Root Hashmerkle:a1b2...3e4f
Signature
verified

Offline knowledge is evidence for judgment, not a reference library. It exists to support decisions and explain them later.

Offline-first for audit resilience. Online for supplementary context.

ナレッジ戦略
オフライン + オンライン

2つのソース、1つの明確な階層

Offline is the authority. Online is the supplement. They are never merged — the reference strategy is unified, not the knowledge itself.

Offline Knowledge

Judgment Basis

Primary source for decisions

Audit Resilience

Evidence that persists

Reproducibility

Replay any past decision

Accountability

Explain why this choice

Primary for Judgment
Online Knowledge

Supplementary Info

Latest developments

External Context

Market conditions

Reference Examples

Similar cases

Advisory Opinion

Not authoritative

Supplement Only
Integration Rule
Judgment: Offline as authoritative source
Context: Online as reference (labeled)
Conflict: Escalate to HITL review
Decision Scenarios
Expense Approvaloffline
Offline

Policy v2.3: Max $5,000 without escalation

Online

Industry avg: $4,200 for similar purchases

Approve based on policy (offline primary)

Humans manage the boundary. AI respects it. When offline and online conflict, the system escalates — it never auto-resolves.

Audit-resistant: Municipality. Finance. Education. Healthcare.

ナレッジメンテナンス

スケジュール化されたナレッジクレンジング

Daily

Real-time conflict check

847

this month

Weekly

Cross-agent consistency

12

this month

Monthly

Full structure review

1

this month

Pipeline
Active
Capture
Scan
3
Cleanse
4
Gate
5
Report

3

Critical

5

Warning

12

Resolved

Cycle-driven knowledge maintenance

ガードレールと意思決定合意

Policy Engine

AI decisions require more than correctness—they need consistency, explainability, and accountability. Policy Engine embeds guardrails into every automated judgment.

Core Modules
Gate Flow
Proposed
Validate
Evaluate
Decision
Result
pass

All constraints satisfied. Proceeding to execution.

Evidence Requirements by Risk Tier
R0Minimal

Unit tests + Logs

R1Moderate

Additional tests + Refactor trace

R2High

Human review + Escalation

R3Critical

Multi-party approval + Staged deploy

Policy Engine permits—not produces—AI decisions.

ドキュメント取り込み

ドキュメントをガバナンスに変換

既存のポリシーが実行可能な意思決定ルールになる

Upload

Policy documents, handbooks, guidelines

Extract

Automatic extraction of decision rules

Integrate

Merge with existing value hierarchy

Sync

Continuous policy synchronization

Supported Document Types

Corporate PolicyEmployee HandbookCompliance GuidelinesOperating ProceduresRisk FrameworksApproval Matrices

Documents in, governance out. No manual rule writing required.