Runtime / Anatomy / Operational Ecology
Liora 系统架构
本页记录 Liora 当前实际运行的系统结构、职责边界与生产证据链。没有运行证据支持的能力不会被列为现状；尚未运行的内容必须明确标记为 Planned。本页不是愿景图，也不是企业平台架构宣传。
This page documents Liora's currently operating system structure, responsibility boundaries, and production evidence chain. Capabilities without runtime evidence are not presented as current. Anything not yet operating must be explicitly marked as Planned. This is not a vision board or an enterprise platform diagram.
single-operatorowner-governedmulti-agent decisiondeterministic executionevidence-firstrecoverable runtime
Operating Model
运行模型 · 角色与职责边界
Branko · Owner
负责系统方向、批准、优先级与资源分配。生产架构、风险边界和重大变更由 Owner 最终授权。
ACTIVE Dior · 市场分析 Agent
独立分析市场并输出 LONG、SHORT 或 WAIT。不执行订单，不拥有最终裁决权。
ACTIVE Codex · 市场分析 Agent
独立分析市场并输出 LONG、SHORT 或 WAIT。不执行订单，不拥有最终裁决权。
ACTIVE Liora · 最终裁决 Agent
接收 Dior 与 Codex 的分析结果，执行最终仲裁并输出 LONG、SHORT 或 WAIT。不直接管理交易所订单生命周期。
ACTIVE Execution Engine · 确定性执行层
负责风险门禁、订单、FSM、持仓生命周期、恢复、日志与交易所状态同步。执行引擎不替代 Agent 的最终 Alpha 裁决。
ACTIVE Production Decision Path
生产决策链 · 从市场到证据
Dior 与 Codex 分别形成独立市场判断，Liora 负责最终裁决。只有通过风险门禁的决策才能进入确定性执行引擎。交易结果进入 Journal、Telemetry 与证据链。
↓
↓
↓
↓
Execution Engine
FSM · 订单生命周期
↓
↓
Journal / Telemetry / Evidence
Runtime Topology
运行拓扑 · 多节点部署
单一 Owner 治理下的多节点部署。控制、分析、执行和外部服务存在明确边界。多节点部署不等同于大规模集群或通用分布式基础设施。
Control
- Branko · Owner 控制端
- 开发与代码审查
- 手动批准
- 仓库管理
Liora Cloud Runtime
- Agent 运行时 (Liora + Codex)
- 生产服务
- 定时调度
- 引擎进程
- 证据采集
Burberry Remote Node
- Dior Agent 宿主
- QQ Bot 中继
- 备份对象
- SSH / Tailscale
External Services
- OKX Exchange
- Telegram 通知
- 模型供应商
- Cloudflare
Opportunity Lifecycle
机会生命周期 · T1 发现 / T2 管理
T1 — Find Opportunity
负责发现交易机会并形成初始交易命题，包括方向、入场、止盈、止损和初始 Thesis。
T2 — Manage Opportunity
负责持仓后的管理，包括继续持有、退出、风险、利润保护与资金效率。T2 不重新预测市场方向。
Production Evidence Chain
生产证据链 · 从交易所到可验证归因
生产证据必须能够从交易所状态回溯到运行时决策、订单生命周期、Journal、Telemetry 和后续归因。Replay 可以验证逻辑，但不能替代 Production Evidence，也不能单独证明盈利能力。
Exchange
Runtime
Journal
FSM
Telemetry
Dataset
EV Attribution PLANNED
VERIFIED PLANNED
Validation Boundaries
验证边界 · Production / Replay / Shadow / Research
Production
真实订单、真实持仓、真实费用和真实交易所状态。Production Evidence 是生产验收的最终依据。
Replay
用于历史数据回放、逻辑验证、回归检测和候选方案验证。与生产隔离，不向生产写回，不构成盈利证明。
Shadow · INACTIVE
在不影响真实交易的情况下观察候选逻辑。当前状态：INACTIVE。
Research
从生产事件中产生 Finding 与 Research Question。不直接修改生产代码、执行交易或替代 RCA。
Current Boundaries
当前边界 · 系统明确的限制
多 Agent 决策，不代表所有 Agent 具有相同权限
当前生产结论只接受可验证的 Production Evidence
Burberry 已经从当前生产决策链中移除（Removed from current production decision path）。它现在是远程基础设施节点，托管 Dior Agent 和 QQ Bot 中继，不再作为独立决策 Agent 运行。
Operating Model
Roles & Responsibility Boundaries
Branko · Owner
Responsible for direction, approval, priorities, and resource allocation. Production architecture, risk boundaries, and material changes require Owner authorization.
ACTIVE Dior · Market Analysis Agent
Independently analyzes the market and outputs LONG, SHORT, or WAIT. Does not execute orders and does not hold final decision authority.
ACTIVE Codex · Market Analysis Agent
Independently analyzes the market and outputs LONG, SHORT, or WAIT. Does not execute orders and does not hold final decision authority.
ACTIVE Liora · Final Decision Agent
Receives analysis from Dior and Codex, performs final arbitration, and outputs LONG, SHORT, or WAIT. Does not directly manage exchange order lifecycles.
ACTIVE Execution Engine · Deterministic Execution Layer
Responsible for risk gates, orders, FSM, position lifecycle, recovery, journaling, and exchange-state synchronization. The engine does not replace the Agents' final Alpha decision.
ACTIVE Production Decision Path
From Market to Evidence
Dior and Codex form independent market judgments. Liora performs final arbitration. Only decisions that pass the risk gates enter the deterministic execution engine. Outcomes are recorded in the Journal, Telemetry, and evidence chain.
↓
Dior
Independent Analysis
Codex
Independent Analysis
↓
Liora · Final Arbitration
↓
↓
Execution Engine
FSM · Order Lifecycle
↓
↓
Journal / Telemetry / Evidence
Runtime Topology
Multi-Node Deployment
Multi-node deployment governed by a single Owner. Control, analysis, execution, and external services have explicit boundaries. This does not imply a large-scale cluster or general-purpose distributed infrastructure.
Control
- Branko · Owner Control
- Development & Review
- Manual Approval
- Repository Operation
Liora Cloud Runtime
- Agent Runtime (Liora + Codex)
- Production Services
- Scheduled Execution
- Engine Processes
- Evidence Collection
Burberry Remote Node
- Dior Agent Host
- QQ Bot Relay
- Backup Target
- SSH / Tailscale
External Services
- OKX Exchange
- Telegram Notifications
- Model Providers
- Cloudflare
Opportunity Lifecycle
T1 Discovery / T2 Management
T1 — Find Opportunity
Finds trading opportunities and forms the initial trade thesis, including direction, entry, take profit, stop loss, and initial reasoning.
T2 — Manage Opportunity
Manages the position after entry, including holding, exit, risk, profit protection, and capital efficiency. T2 does not re-predict market direction.
Production Evidence Chain
From Exchange to Verifiable Attribution
Production evidence must be traceable from exchange state through runtime decisions, order lifecycle, journal, telemetry, and downstream attribution. Replay may validate logic but cannot replace Production Evidence or independently prove profitability.
Exchange
Runtime
Journal
FSM
Telemetry
Dataset
EV Attribution PLANNED
VERIFIED PLANNED
Validation Boundaries
Production / Replay / Shadow / Research
Production
Real orders, real positions, real costs, and real exchange state. Production Evidence is the final basis for production acceptance.
Replay
Used for historical replay, logic validation, regression detection, and candidate validation. Isolated from production, does not write back, and does not constitute proof of profitability.
Shadow · INACTIVE
Observes candidate logic without affecting live trading. Current status: INACTIVE.
Research
Produces findings and research questions from production events. Does not directly modify production code, execute trades, or replace Root Cause Analysis.
Current Boundaries
Explicit System Limits
Governed by a single Owner; not an unsupervised autonomous system
Multi-Agent decision-making does not imply equal authority among Agents
Multi-node deployment does not imply a large-scale cluster
Replay is a validation tool, not proof of production profitability
Recovery mechanisms reduce failure impact but do not make the system failure-free
Current production conclusions require verifiable Production Evidence
Unverified capabilities are not presented as current system capabilities
Burberry has been removed from the current production decision path. It now serves as a remote infrastructure node, hosting the Dior Agent and QQ Bot relay. It no longer operates as an independent decision Agent.