Frontier science for LoopX: mathematics, physics, biology and AI / 前沿科学研究方向

[huangruiteng]

Which scientific ideas should shape LoopX next? / 哪些科学思想值得进入 LoopX？

LoopX already has durable state, evidence, continuation and typed control
boundaries. The next research question is how to use them to make better
decisions: when to keep working, which history matters, which experience
actually helps, and what evidence should justify a change.

We are opening a cross-disciplinary research portfolio for review. It
includes near-term experiments and long-term hypotheses across mathematics,
physics, biology and AI. The RFC distinguishes established foundations,
recent research, implementation availability and evidence that has not
yet been demonstrated in LoopX.

LoopX 已有持久状态、证据、接续和类型化控制边界。下一步研究如何让这些基础产生更好的决定：何时继续、哪些历史必须保留、哪些经验真正有效、什么证据足以改变策略。这里公开完整的跨学科研究组合，征集反例、方法和合作场景。

Full RFC / 完整 RFC: English · 中文

All ten directions / 全部十个方向

Track	Scientific direction / 科学方向	LoopX experiment / 实验问题
T01	Anytime-valid experiment evidence / 连续有效的实验证据	固定 treatment 版本，以完整且可比较的任务 episode 为单位，记录分配、缺失、有界 outcome、分析身份和多重比较处理。比较连续有效证据、预注册固定样本方法与现有启发式。
T02	Decision-preserving memory and continuation / 保留决策差异的记忆与接续	在相同实际 token 预算下比较普通摘要、结构化摘要和决策保真 packet，计入回查成本。先接一条明确 continuation 路径，构建过期证据、目标变化和任务身份混淆的对照样例。
T03	Metareasoning and event-triggered stride / 元决策与事件触发步幅	行动前记录预测，再绑定结果。将 shadow policy 与现有规则、便宜阈值对照；行动效果使用前瞻受控实验，被动日志不能独立证明反事实策略价值。
T04	Causal memory utility and transfer conditions / 因果记忆效用与迁移条件	固定 checkpoint、模型、任务、工具版本，对有无整组记忆做重复重跑；必要时再测 A/B/AB/no-memory 交互。把条件化效用、支持数和不确定性绑定到现有 outcome receipt。
T05	Formal authority kernel and checked certificates / 形式化权威小核与可检查证书	形式化一个 revision/lease/receipt 状态机并对照实现。候选证书绑定状态版本、动作摘要、权限范围、前后条件和检查器版本；相关状态变化后证书失效。
T06	Structured curriculum and complementary learning / 结构化课程与快慢学习	快通道保存事件证据，慢通道巩固带适用条件、例外、来源和版本的程序知识。用同样经验和预算，对照随机、时间顺序和对照课程，在新组合任务上验证。
T07	Diversity-preserving capability evolution / 保留多样性的能力进化	只演化一个有界 adapter、记忆规则或 planner proposal，等总预算比较多样性 archive 与单冠军爬山法。保留谱系和负结果，冻结 evaluator、权限与 held-out 任务。
T08	Active information acquisition and experiment design / 主动信息获取与实验设计	保留有界 competing hypotheses；为每个 probe 写明可能观察及其改变的决定。将决策相关信息价值与简单启发式或 Bayesian optimization 比较；负证据可以关闭假设。
T09	Multiscale predictive control state / 多尺度预测控制状态	检验紧凑状态能否跨 provider 预测交付、阻塞与恢复，以及完整微观历史是否仍增加重要预测信息。预测之后还需受控干预，才能讨论行动等价。
T10	Scientific experiment campaign infrastructure / 科学实验 campaign 基础设施	从计算模拟器与领域 owner 的 evaluator 开始，跨中断追踪实验意图、资源预约、数据/样品谱系、测量、复现与决策；对照伙伴原流程的恢复、可复现性和资源浪费。

Suggested first focus / 首轮重点

T01, T02 and T03: continuously valid experiment evidence,
decision-preserving continuation, and cost-aware metareasoning in shadow
mode. They can provide the evaluation and state foundation for later causal
memory, curriculum and capability evolution.

首轮建议 T01 连续有效证据、T02 决策保真接续、T03 元决策 shadow。这三项为后续因果记忆、课程与能力进化提供评估和状态基础。形式化小核可在独立契约明确时并行立项；物理学的多尺度状态和科学实验 campaign 保留为长期研究选择。

Where contributions would help / 需要哪些贡献

1. Statistics: Which experimental units and anytime-valid constructions
   remain defensible for adaptive, correlated agent work?
2. Context and cognition: What real handoff failures distinguish semantic
   similarity from decision equivalence?
3. Formal methods: Which small, stable authority or receipt contract is
   worth proving first?
4. Neuroscience and evolution: Which controlled experiments can test
   curriculum or diversity benefits without hidden extra compute or leakage?
5. Physics and scientific labs: Who has a simulator, measurable campaign
   failure or design-partner workflow suitable for a bounded pilot?

中文：欢迎统计研究者帮助定义有效实验单位；context/cognition 贡献者提供真实接续反例；形式化研究者选择值得先证明的小契约；神经科学/进化研究者提出控制预算和泄漏的实验；物理、材料与生物计算伙伴提供模拟器或真实科研流程。

A useful reply names track → concrete failure/question → primary evidence
or reproducible experiment → expected result and falsifier → willing owner
or collaboration scope. Disagreement with a proposed priority is welcome
when it changes the next experiment.

建议回复：方向 → 具体失败/问题 → 一手证据或可复现实验 → 预期结果与反证条件 → 愿意承担的范围。

Evidence and governance boundary / 证据与治理边界

No new treatment is promoted by this Discussion. Recent preprints are
hypothesis sources, biological experiments do not automatically transfer to
software agents, and formal/statistical guarantees retain their assumptions.
There is no blanket commitment to implement all ten directions or to control
physical labs. Quantum/neuromorphic/biological backends and generic entropy
scoring remain watch items until a bounded caller and measurable benefit exist.

本讨论不晋升运行时能力、不承诺全部实施、不授予物理实验权限。论文新颖性与可部署证据分开；量子、神经形态、生物 backend 和通用熵评分等待明确 caller 与收益证据。

This is an asynchronous Ideas discussion, not a scheduled meeting or roadmap
vote. Any later Open Strategy Review
should select at most four evidence-backed directions and publish its own
agenda.

Research tracker / 研究跟踪： #4391

Proposal PR / 提案 PR: #4393