Y&Y · Singapore · Decision Core

大模型时代的
决策中枢
The Decision Core for the
Age of Large Models

让重大判断回到证据、反方与可证伪条件。
我们把多个前沿大模型,组织成一条可审计的判断工序
Bring every major judgment back to evidence, counterarguments, and falsifiable conditions.
We organize frontier models into one auditable judgment pipeline.

判断可追问Auditable · 每个结论开得出原证据every conclusion opens to its evidence 数字可回查Traceable · 每个数值落得回出处every number traces to its source 结论可证伪Falsifiable · 每个判断写明何时应推翻every judgment states when it should be overturned
The Problem

决策层不缺答案,缺的是可追问的判断 Executives don't lack answers.
They lack judgments that can be questioned.

信息更多、模型更快、假设更短命。每一层汇报都更流畅,也可能离原始条件更远—— 没有人撒谎,结论仍会失去分歧、来源与适用条件。 More information, faster models, shorter-lived assumptions. Each layer of reporting reads smoother — and drifts further from the original conditions. Nobody lies, yet conclusions shed their dissent, their sources, and their limits.

01

分歧被过滤Dissent gets filtered

组织更容易上送共识,不上送未决冲突。最有价值的反方意见,往往死在中间层。 Organizations escalate consensus, not unresolved conflict. The most valuable counterargument usually dies in the middle layers.

02

条件被丢失Conditions get lost

表达越完整,越容易遮住证据缺口。最终答案缺少反方,也缺少推翻当前判断的条件。 The more polished the deliverable, the better it hides the gaps. Final answers arrive without counterarguments — and without the conditions that would overturn them.

03

单一视角锁定Single-lens lock-in

买一个大模型,买不到独立判断——同一视角与同一盲区,加再多算力也不会消失。 Buying one large model doesn't buy independent judgment. The same lens carries the same blind spots — and no amount of compute makes them go away.

Method

把多模型组织成一条可审计的判断工序 Many models. One auditable pipeline.

判断先行,证据与反方随附。异构模型家族互为制衡,误差彼此独立。 Verdict first, evidence and counterarguments attached. Heterogeneous model families check one another — their errors are independent by design.

1

独立研判Independent analysis

不同家族的模型互不看答案,分别形成判断与依据。独立性写进工序,不靠自觉。 Model families work blind to one another, each forming its own judgment and rationale. Independence is engineered, not assumed.

2

交叉质证Cross-examination

逐项追问来源、口径、反例与不确定性。相互批判暴露事实盲区。 Sources, definitions, counterexamples, and uncertainty — interrogated item by item. Mutual critique exposes factual blind spots.

3

合议裁决Adjudication

保留分歧,记录修订理由与证伪条件。不用多数投票掩盖冲突。 Dissent is preserved; every revision is reasoned; every verdict carries its falsifiers. Majority votes never bury conflict.

4

可追溯交付Traceable delivery

判断、原证据、主要反证与行动窗口随附。先读结论,再决定追问到哪一级。 Verdict, raw evidence, strongest rebuttal, action window — delivered together. Read the conclusion first; drill down as far as you choose.

全程留痕 · 逐级回查 · 每次修订有理由,每个结论有推翻条件 Fully logged · drill-down at every level · every revision reasoned, every conclusion falsifiable
Architecture

模型可替换,工序与记忆才是长期底座 Models are replaceable. Process and memory are the foundation.

能评测所有模型的一层,必须让模型供应商保持可替换。中立不是姿态,而是架构。 A layer that evaluates every model must keep every vendor replaceable. Neutrality is not a posture — it is architecture.

供应商可以更换,验证纪律机构记忆持续复利。 Vendors change. Verification discipline and institutional memory compound.
  • 异构模型池Heterogeneous model pool GPT · Claude · Gemini · 开源与中文前沿模型,可插拔接入 GPT · Claude · Gemini · open & Chinese frontier models, hot-swappable
  • 编排与路由中枢Orchestration & routing 任务分解 · 模型路由 · 同题评测对照 task decomposition · model routing · same-question benchmarking
  • 长期机构记忆Institutional memory 判断史 · 事故库 · 教训回写,跨模型不失真 judgment history · incident library · lessons written back, model-agnostic
  • 安全基线Security baseline 主权部署 · 最小权限 · 全链路日志 sovereign deployment · least privilege · end-to-end logging
L0

权利边界Data rights

无权利,不取、不算、不交付。边界先于能力。 Without established rights, nothing is ingested, computed, or delivered. Boundaries precede capability.

L1

证据层Evidence layer

来源分级、口径校验、数字回查。 Source grading, definition checks, number-level traceback.

L2

异构研判Blind analysis

不同模型家族独立分析,禁止互看。 Model families analyze independently — cross-viewing is forbidden.

L3

质证裁决Adjudication

保留反方、分歧与证伪条件。 Rebuttals, dissent, and falsifiers preserved in the record.

L4

高管交付Executive delivery

判断、原证据、反方与行动窗口。 Verdict, raw evidence, counterarguments, action window.

L5

评测运营Eval operations

事故回写、留出评测与持续回归。 Incidents written back, held-out evals, continuous regression.

Applications

同一座中枢,迁移到不同决策场景 One core. Many decision arenas.

方法可复用;行业来源、责任边界与验收指标逐行业重配。 The method transfers; sources, responsibility boundaries, and acceptance metrics are re-engineered per industry.

ENERGY

能源决策情报Energy intelligence

每日能源市场晨报:判断先行、数字可回查、结论附证伪条件,可现场审验。 A daily energy briefing: verdict first, numbers traceable, conclusions falsifiable — auditable on the spot.

FAMILY OFFICE

家族办公室Family office

跨资产情报汇总与重大判断留痕。卖方材料与单一顾问的盲区,用异构验证补齐。 Cross-asset intelligence and a durable record of major judgments. Heterogeneous verification closes the blind spots of sell-side research and single-advisor views.

QUANT

量化研究Quant research

信号验证、研报去伪与反事实推演。回测验证历史,我们验证论证链。 Signal validation, research vetting, counterfactual reasoning. Backtests verify history; we verify the chain of argument.

RESEARCH

学术研读Academic reading

引用核验、分歧保留与文献审读。工具只存不判;我们让每条引用可溯源。 Citation checking, dissent preservation, literature review. Most tools only store; ours adjudicates — every citation traceable.

Origin & Team

先为自己而建,再产品化交付 Built for ourselves first. Productized after.

Origin源起Where it began

这套中枢最初为创始团队自身的家族办公室而建——真实资产、真实决策压力、每日运转打磨。 被自己的钱检验过的工序,才拿出来做成行业产品。 This core was first built for the founding team's own family office — real assets, real decision pressure, refined in daily operation. Only a pipeline tested with our own capital becomes a product.

在新加坡,我们把学术训练、工程底座与企业决策场景连接起来: 亚洲时区、法治与数据规则,连接多侧市场,而不依附任何一侧。 From Singapore, we connect academic training, engineering infrastructure, and enterprise decision-making: Asia's timezone, rule of law, and data regime — linked to every side of the market, beholden to none.

Team 研究训练 × 工程能力 × 商业实践 Research × Engineering × Practice

把前沿论文转译为独立研判、质证与评测纪律;把模型、证据链与审计组织成可运行工作流; 理解高管阅读、部门复核与责任边界。 We translate frontier research into disciplines of independent analysis, cross-examination, and evaluation; organize models, evidence chains, and audit into running workflows; and understand how executives read, how departments review, and where responsibility ends.

清华大学Tsinghua 北京大学Peking University 新加坡国立大学NUS 香港大学HKU
Sovereign Deployment

情报全球采集,数据按级进域 Global intelligence. Tiered data boundaries.

边界先于能力;权限、日志与退出机制随级别增强。未经评审,不承诺部署结论。 Boundaries precede capability; permissions, logging, and exit mechanisms strengthen with each tier. No deployment commitments before review.

L0

会前预检Pre-meeting review

公开信息与授权材料。不接内部系统,文件级最小披露。交付红旗、证据链、反方与证伪条件。 Public and authorized material only. No internal systems; file-level minimal disclosure. Delivers red flags, evidence chains, rebuttals, falsifiers.

L1

受控协作Controlled collaboration

限定内部材料与指定信源。专属环境或受控接口,最小权限。批准范围内的可追溯判断。 Scoped internal material and designated sources. Dedicated environments or controlled interfaces, least privilege. Traceable judgments within approved scope.

L2

深度部署Deep deployment

持续更新的专有数据与工作流。私有化或专属云,权限日志归客户。合同定义的持续服务与审计能力。 Continuously updated proprietary data and workflows. On-premises or dedicated cloud; permissions and logs belong to the client. Service and audit defined by contract.

让每一个重大判断,
经得起追问
Make every major judgment
stand up to scrutiny.

从一个真实决策主题开始:让冲突被看见,让结论回到证据。 Start with one real decision topic: surface the conflicts, and bring the conclusions back to evidence.