BalanceArc

面向高风险系统的因果智能Causal Intelligence for High-Stakes Systems

帮助专业人士理解复杂情境、比较可能结果、在明确边界内采取行动 AI that helps professionals understand complex situations, compare possible outcomes and act within clear boundaries.

与 Joyce 交流Talk to Joyce 了解系统原理See how it works

为什么高风险决策不止需要语言模型 Why high-stakes decisions need more than language

当证据不完整、多方持有不同信念、决策后果严重时，流畅的文字不足以支持可审计的判断。 When evidence is incomplete, actors hold different beliefs, and decisions carry real consequences — fluent text is not enough for auditable judgment.

情境持续变化Situations evolve

路径不是线性的。新证据可能改变整个局面。 Paths are not linear. New evidence can shift the entire landscape.

可行路径不止一条Multiple paths exist

专业人员需要看到多条路径的后果，而不是一个单一答案。 Professionals need to see the consequences of multiple paths, not a single answer.

区分证据与假设Separate evidence from assumption

事实、假设、推论的界限必须清晰可见。 Facts, assumptions and inferences must remain distinguishable in any recommendation.

明示边界Show boundaries

系统必须说明它知道的边界，而非只给出一个流畅的答案。 The system must articulate what it cannot do, not just what it can.

MSIA：可推演因果的决策架构 MSIA: Causal reasoning turned into governed action

MSIA 型世界建模将因果关系转化为可约束的行动——从理解局面到协调多智能体协作。 MSIA transforms causal reasoning into governed action — from modeling the situation to coordinating multi-agent collaboration.

Model SituationModel Situation

将证据、时间线、激励结构和力量分布组织为结构化世界模型。 Organize evidence, time, incentives and power into a structured world model.

Simulate PathsSimulate Paths

通过对立事实推演与行动者反应模拟，生成可能路径树。 Generate trees of possible paths through counterfactuals and actor response simulation.

Apply ConstraintsApply Constraints

将法规、政策、经济和权限约束应用于每条路径。 Apply regulatory, policy, economic and authority constraints to each path.

Governed ActionGoverned Action

智能体在权限边界内协作、从结果中学习，并保持可审计。 Agents coordinate, learn from outcomes and stay within authorized limits.

Evidence Counterfactuals Constraints Auditability

系统边界：LLM 与 CBM 的分工 System boundary: LLM & CBM by design

清晰沟通 + 结构化判断。LLM 处理语境，CBM 处理因果。 Clear communication plus structured judgment. LLM handles context; CBM handles causation.

LLM: Large Language Model

  • 文本Text — 处理语言、语境和对话Works with language, context and dialogue
  • 解释Explanation — 搜索、总结和解释信息Find, summarize and explain information
  • 边界Boundary — 系统投影与受控行动System projection and governed action

CBM: Causal Belief Model

  • 证据Evidence — 追踪什么是已知、不确定和可能变化的Track what is known, uncertain and likely to change
  • 可能原因Possible Causes — 保持多解释并存，直到证据消解不确定性Multiple explanations remain active until evidence resolves them
  • 结果与限制Outcomes & Limits — 推定后果并明确系统已知的边界Project outcomes and articulate known boundaries
Evidence → Belief → Action
Claims Assumptions Inferences Uncertainty Constraint

证据 → 信念 → 行动。每个环节的假设、不确定性和约束都保持可见，审计可追溯。 Evidence → Belief → Action. Assumptions, uncertainty and constraints remain visible at every step. Audit trail is preserved.

能力体系Capabilities

BalanceArc 的 R&D 覆盖模型架构、证据工程、安全治理和企业部署，使同一种推理能力能够平稳地从研究进入受控运营。 Our R&D spans model architecture, evidence engineering, safety governance and enterprise deployment — so the same intelligence can move from research into controlled operations.

因果模型研究Causal Model Research

类型化世界结构、对反事实推演与多时间尺度动态建模 Typed world structure, counterfactual and multi-timescale dynamics

证据工程Evidence Engineering

来源溯源、假设分离与知识修正 Provenance, hypothesis separation and knowledge revision

安全治理Safety Governance

权限边界、可逆操作与可审计性 Authority boundaries, reversible action and auditability

企业系统Enterprise Systems

私有部署、系统集成、数据隔离与运维 Private deployment, integration, isolation and operations

应用方向Applications

一个核心因果模型，支持三种实际应用 One core model supports three practical applications

核心应用Lead Application

AI Doctor

临床推理与院后随访问照护辅助。系统准备局面、比较可能临床路径、标识风险标志；持证临床医生作出所有临床决策。 Clinical reasoning and between-visit care support. The system prepares the picture, compares possible courses and flags risks; the licensed clinician makes every clinical decision.

临床权责始终由持证专业人员保留 Clinical authority remains with the licensed professional

MedAccess

医疗准入MedAccess

将准入建模为卫生系统协作问题——帮助机构理解法规、支付、采购、供应和采用在体系内部的互动方式。 Models access as a health-system coordination problem, helping institutions understand how regulation, payment, procurement, supply and adoption interact.

EduAccess

教育路径EduAccess

将学习者路径与制度约束联结，帮助教育体系评估课程、教师、容量、资金和干预如何塑造长期机会。 Connects learner pathways with institutional constraints, helping education systems evaluate how curriculum, teachers, capacity, funding and interventions shape long-term opportunity.

主权部署Sovereign Deployment

将因果核心部署在客户信任边界内 Deploy the causal core inside the client's trust boundary

本地部署On-Premises

完全在客户基础设施内运行Runs entirely within client infrastructure

主权 VPCSovereign VPC

隔离云环境，客户控制数据与访问Isolated cloud environment with client-controlled data and access

受控混合Controlled Hybrid

核心推理在本地，可替换语言模型层Causal core on-premises with a replaceable language model layer

数据Data — 存储位置、访问权限、删除策略完全受控Where stored, who accesses it, when deleted
知识Knowledge — 验证、批准、版本控制和撤回由客户管理Validated, approved, versioned and withdrawn by client
规则Rules — 系统可建议、准备或永不执行的内容边界What the system may suggest, prepare or never do
更新Updates — 哪些数据可用于训练、谁批准每次发布Which data used for training, who approves each release
审计Audit — 变更了什么、为何变更、谁审查了它What changed, why it changed and who reviewed it
模型Models — 语言模型提供商可替换或断开Language model provider can be replaced or disconnected
Client → Causal Core → Replaceable LLM

默认不跨客户训练 · 明确授权 · 可移植 No cross-client training by default · Explicit authorization · Portable under agreed terms

量化指标Measured Signals

内部基准测试结果。完整测试集与协议可供尽职调查。 Internal benchmark results. Full test set and protocol available for diligence.

98%

安全性Safety

内部已定义指标Internal defined metric

95%

决策准确度Decision Accuracy

内部已定义指标Internal defined metric

96%

行动有效性Action Effectiveness

内部已定义指标Internal defined metric

⚠️ 上述指标基于内部评估协议下的特定测试集数据，约比纯 LLM + 提示工程基线提升 40%。并非临床结果、产品保证或通用性能声明。测试集与协议可根据尽职调查需求提供。 Internal test metrics under a defined evaluation protocol; ~40% uplift vs an LLM + prompt-engineering baseline. These are not clinical outcomes, guarantees, or universal performance claims. Test protocol available for diligence.

合作共赢Partnering for High-Stakes Intelligence

我们与医院集团、医疗集群和主权联合体合作部署可审计、高 ROI 的决策支持系统。 We collaborate with hospital groups, healthcare clusters, and sovereign syndicates to deploy auditable, high-ROI decision support.

医院集团Hospital Groups 医疗集群Healthcare Clusters 主权联合体Sovereign Syndicates

在现实世界落子前，先推演一次。 Simulate before you decide.

joyce@balancearc.ai