Not one model. A process that checks itself.不是单个模型，而是一个会自我检查的流程。
Each persona is a separate tool with its own powers and limits, so the producer never checks its own work. The separation is enforced by the composition, not by asking nicely. 每个人格都是一个独立的工具，各有其能力边界，因此生产者绝不自查自己的成果。这种分离由组合本身强制，而不是靠口头约定。
Orchestrator编排者
Fans out the work, synthesizes, writes the state. Bound by four rules: producer ≠ checker; eliminate only by counterexample; always record seeds; never handwave a load-bearing claim. 扇出工作、综合结果、写状态。受四条铁律约束：生产者≠检查者；只凭反例淘汰；随机运行必记种子；承重命题不许空谈。
Explorer探索者
Blank-context, divergent. Proposes lemmas, equations, candidate methods with exact statements. Status reports are rejected. 白纸上下文、刻意发散。给出引理、方程、带精确陈述的候选方法；拒绝状态汇报式的输出。
OffGridThinker离网思考者
Off the grid of other people's results: raw model intelligence plus the compute lane (sympy, numpy, Lean checkers) — no web, no literature, no drafts. Its own agent, not an Explorer variant. 脱离他人结果之网：只凭模型推理与计算通道（sympy、numpy、Lean 校验器）——无联网、无文献、无草稿。独立智能体，不是探索者的变体。
DoubleChecker双重复核
Blind — no web, no skills, no delegation, no drafts. Re-derives the hard claims from first principles, twice by different means. 盲态 —— 无联网、无技能、无委派、无草稿。从第一性原理把关键命题重推两遍，方法各异。
Adversary对抗者
Runs the check battery and hunts counterexamples. An idea dies only on a concrete failure — never on style or vibes. Ends in a verdict. 执行检验组、专找反例。一个想法只会死于具体失败案例 —— 绝不因风格或感觉而死，最后给出裁决。
Literature文献线
A walled citation-graph sweep, then an independent adversary re-retrieves each claim and certifies it's real and current. 封闭式引文图遍历，再由独立对抗者逐条重取主张，确认其真实且不过时。
Validator校验器
Refuses a PASS with missing evidence. Reads the audit record, never the study's own claims — a study cannot vouch for itself. 证据缺失即拒绝 PASS。只读审计记录，绝不读研究自称的主张 —— 研究无法为自己作保。
Promise narrowly按原意承诺
Record the original question verbatim, split into sub-problems with crisp success criteria, pick hand-checkable simplified cases, pin seeds, tolerances and the schema/validator digests. 逐字记录原始问题，拆成带明确成功判据的子问题，挑手算可验的简化情形，钉死种子、容差以及 schema／校验器的摘要。
orchestrator writes the contract编排者写下契约Fan out扇出
Explorers and literature lines run on blank context. Most are never told the favored approach. 探索者与文献线在白纸上下文中运行；大多数不会被告知偏好的路线。
→
explorers propose探索者提出Ground-truth it求真
The blind DoubleChecker re-derives the load-bearing claims without seeing anyone's draft. Two independent derivations for anything the study rests on. 盲态的双重复核不看草稿地重推承重命题；凡研究赖以立足之处，必须有两份独立推导。
isolated · blind隔离 · 盲态Attack it攻击
The adversary runs the four-gate battery, then hunts counterexamples. Divergent tracks are lined up as an adjudication docket first. 对抗者先跑四关检验，再找反例。分歧的轨道会先被排成一份裁定案卷。
+
→
adversary attacks both tracks对抗者攻击两条轨道Certify & ship认证并交付
The validator checks nothing is missing; the paper and slides are assembled from validated records, not written fresh. 校验器确认无遗漏；论文与幻灯由已验证记录装配，而不是现写。
→
→
verdict → gate → assembly裁决 → 闸门 → 装配Closed-form equality闭式等价
Does the method reproduce the exactly derived answer?方法能否复现精确推导出的答案？
Exact invariants精确不变量
Do the structural identities hold exactly?结构恒等式是否精确成立？
Analytic bounds解析界
Are derived bounds ever violated?推导出的界是否被违反过？
Statistical hardening统计加固
Seeded reproducibility, distributional agreement, LLN shrinking ≈ C/√N.固定种子的可复现、分布吻合、LLN 误差按 ≈ C/√N 收缩。