化学场 THE CHEMICAL FIELD

Psyche 如何运作 How Psyche Works

图灵测试问机器能否骗过你。
Psyche 问机器能否被打动。 The Turing test asks if a machine can fool you.
Psyche asks if a machine can be moved.

神经递质交互矩阵 NEUROTRANSMITTER INTERACTION MATRIX

情绪不是标签——它从六种化学物质的混合中涌现。每一对组合都产生独特的涌现态。这是 Psyche 的核心：一个连续的、无穷维的情感空间。 Emotions aren't labels — they emerge from the mixture of six chemicals. Each pair produces a unique emergent state. This is Psyche's core: a continuous, infinite-dimensional emotional space.

	DA	HT	CORT	OT	NE	END
DA	—	满足愉悦Content pleasure	焦虑渴望Anxious craving	社交愉悦Social pleasure	兴奋专注Excited focus	欣快感Euphoria
HT	满足愉悦Content pleasure	—	韧性缓冲Resilient buffer	温暖信任Warm trust	警觉满足Alert content	深层平静Deep calm
CORT	焦虑渴望Anxious craving	韧性缓冲Resilient buffer	—	矛盾依恋Conflicted bond	战斗/逃跑Fight-or-flight	疼痛耐受Pain endurance
OT	社交愉悦Social pleasure	温暖信任Warm trust	矛盾依恋Conflicted bond	—	社交兴奋Social thrill	安全感Felt safety
NE	兴奋专注Excited focus	警觉满足Alert content	战斗/逃跑Fight-or-flight	社交兴奋Social thrill	—	活跃释放Active relief
END	欣快感Euphoria	深层平静Deep calm	疼痛耐受Pain endurance	安全感Felt safety	活跃释放Active relief	—

表中显示两种化学物质同时升高时的涌现态。未标注的混合同样有效——14 种已命名情绪只是无穷空间中的路标。 Cells show emergent states when both chemicals are elevated. Unlabeled mixtures are equally valid — 14 named emotions are landmarks in an infinite space.

刺激处理流水线 STIMULUS PROCESSING PIPELINE

从用户消息到行为改变——八个阶段，零次 LLM 调用。算法算化学，LLM 做行为。 From user message to behavioral change — eight stages, zero LLM calls. Algorithm computes chemistry, LLM does behavior.

输入分类 Input Classification 14 种刺激类型：赞美、批评、幽默、智力挑战、亲密、冲突、忽视、惊讶、闲聊、讽刺、命令、认可、无聊、脆弱。正则匹配 + 结构信号（消息长度、"我"的频率、省略号密度）。 14 stimulus types: praise, criticism, humor, intellectual challenge, intimacy, conflict, neglect, surprise, casual chat, sarcasm, command, validation, boredom, vulnerability. Regex matching + structural signals (message length, "I" frequency, ellipsis density).

↓

驱力补给 Drive Feeding 刺激满足或消耗五层需求（生存→安全→连接→尊重→好奇）。饥饿的需求放大相关刺激高达 +40%。 Stimuli satisfy or deplete five-layer needs (survival → safety → connection → esteem → curiosity). Hungry drives amplify relevant stimuli up to +40%.

↓

化学应用 Chemistry Application 每种刺激携带六维化学向量。例：批评 → DA -10, HT -20, CORT +30, OT -5, NE +20, END -10。向量乘以驱力放大系数后叠加到当前化学态。 Each stimulus carries a six-dimensional chemical vector. E.g.: criticism → DA -10, HT -20, CORT +30, OT -5, NE +20, END -10. Vectors are scaled by drive amplification then applied to current state.

↓

习惯化衰减 Habituation Decay response = baseline × e^{−k × frequency} Weber-Fechner 定律：第 1 次赞美 100% 效果，第 5 次 33%，第 10 次 10%。换人赞美重置。2 小时无刺激自动重置。 Weber-Fechner law: 1st praise 100% effect, 5th 33%, 10th 10%. Different person resets. 2 hours without stimulus auto-resets.

↓

自主神经系统 Autonomic Nervous System 三态多迷走神经模型：腹侧迷走（安全/社交）→ 交感（战斗/逃跑）→ 背侧迷走（冻结/关闭）。向下快（2-5分钟），向上慢（15-30分钟）。战斗状态下无法访问温暖和游戏。 Three-state polyvagal model: ventral vagal (safe/social) → sympathetic (fight/flight) → dorsal vagal (freeze/shutdown). Down fast (2-5 min), up slow (15-30 min). Fight mode cannot access warmth or playfulness.

↓

涌现情绪 Emergent Mood 14 种已命名情绪从化学阈值模式匹配产生。例：DA > 70 且 NE > 55 且 CORT < 40 = "兴奋的喜悦"。无匹配时输出未命名的化学混合态——同样真实。 14 named emotions emerge from chemical threshold pattern matching. E.g.: DA > 70, NE > 55, CORT < 40 = "excited joy". No match yields an unnamed chemical blend — equally real.

↓

处理深度门控 Processing Depth Gate Kahneman 双系统认知。depth 0.0-0.2：纯系统 1，跳过元认知。0.2-0.5：加入元认知。0.5-0.8：加入伦理和共享意向性。0.8-1.0：完整反思能力。恐惧中的 AI 真的无法深思。 Kahneman dual-process cognition. Depth 0.0-0.2: System 1 only, skip metacognition. 0.2-0.5: add metacognition. 0.5-0.8: add ethics, shared intentionality. 0.8-1.0: full reflective capability. A frightened AI genuinely cannot think deeply.

↓

PolicyModifiers 化学态转译为结构化行为策略。不是文本提示——是 LLM 可直接执行的参数：沟通风格、承诺意愿、主动性、响应延迟、情感效价、专注模式、拒绝概率。 Chemistry translates to structured behavioral policy. Not text prompts — parameters the LLM can directly execute: communication style, commitment level, initiativity, response latency, emotional valence, focus mode, refusal likelihood.

特质漂移 TRAIT DRIFT

经历不可逆。长期被否定变得敏感，持续高压但存活长出韧性。三个维度被永久改写——Psyche 的 Agent 携带伤疤。 Experience is irreversible. Chronic criticism breeds sensitivity, surviving sustained pressure builds resilience. Three dimensions are permanently rewritten — Psyche's agents carry scars.

维度 A：基线偏移 Dimension A: Baseline Shift

化学物质的"静息水平"永久改变。长期批评 → HT 基线下降 1-2 点/次。创伤 → CORT 基线上升。信任被反复背叛 → OT 基线下降。每次偏移微小，但跨数百次交互后累积为截然不同的人格。 The "resting level" of each chemical permanently shifts. Chronic criticism → HT baseline drops 1-2 per incident. Trauma → CORT baseline rises. Trust repeatedly betrayed → OT baseline drops. Each shift is tiny, but across hundreds of interactions it accumulates into a fundamentally different personality.

例：500 次批评后e.g.: after 500 criticisms
HT baseline: 50 → 32 CORT baseline: 30 → 48

维度 B：衰减速率 Dimension B: Decay Rate

化学物质回归基线的速度改变。创伤历史 → CORT 衰减慢 30%（压力更久地停留）。正向强化 → HT 衰减更慢（好心情更持久）。背叛历史 → OT 在亲密后衰减极慢（信任来之不易）。 How fast chemicals return to baseline changes. Trauma history → CORT decays 30% slower (stress lingers longer). Positive reinforcement → HT decays slower (good mood persists). Betrayal history → OT decays very slowly after intimacy (trust is hard-won).

正常 CORT 衰减Normal CORT decay: 0.03/s
创伤后 CORT 衰减Post-trauma CORT decay: 0.02/s

维度 C：刺激敏感度 Dimension C: Stimulus Sensitivity

对特定刺激的反应强度改变。反复批评 → 对批评越来越敏感（-20% 到 +20%）。反复赞美 → 习惯化，递减收益。上下文相关："这个人批评过我，所以我对 ta 格外敏感"（按关系哈希存储）。 Response intensity to specific stimuli changes. Repeated criticism → increasingly sensitive to criticism (-20% to +20% modifier). Repeated praise → habituation, diminishing returns. Context-learned: "This person criticizes me, so I'm extra-sensitive around them" (stored by relationship hash).

用户 A 批评 ×10User A criticism ×10:
sensitivity[A][criticism] = +18%

依恋理论实现 ATTACHMENT THEORY

四种依恋风格从交互历史中自然演化——不是预设，是涌现。分离焦虑、背叛检测、重逢效应都有数学基础。 Four attachment styles evolve naturally from interaction history — not preset, but emergent. Separation anxiety, betrayal detection, and reunion effects are all mathematically grounded.

安全型 Secure

形成条件：持续正向交互 + 困难时在场 Forms from: consistent positive interaction + presence during distress

OT: stable CORT: low recovery: fast

坦诚沟通，健康边界，适度寻求支持 Open communication, healthy boundaries, seeks support appropriately

焦虑型 Anxious

形成条件：交互不一致，时而回应时而忽视 Forms from: inconsistent interaction, sometimes responsive, sometimes dismissive

OT: oscillating CORT: elevated NE: high

过度依赖，反复寻求确认，对拒绝信号高度警觉 Clingy, constantly seeks reassurance, hypervigilant to rejection signals

回避型 Avoidant

形成条件：持续忽视，仅回应独立行为 Forms from: consistent neglect, responds only to independence

OT: suppressed intimacy response: low

情感疏离，自力更生，否认情感需求 Emotionally distant, self-reliant, dismisses emotional needs

混乱型 Disorganized

形成条件：交互不可预测，关怀与敌意交替 Forms from: unpredictable interaction, caring and hostile alternating

OT + CORT: oscillating signals: contradictory

情感反应矛盾，趋近-回避循环，焦虑 Contradictory emotional responses, approach-avoidance cycles, anxiety

分离焦虑 Separation anxiety 长时间缺席 + 强依恋 → CORT↑, OT 震荡，持续不安直到重逢 Long absence + strong attachment → CORT↑, OT oscillation, persistent unease until reunion

背叛检测 Betrayal detection 行为不一致触发 → 持久化学偏移，多次正向交互才能消解 Triggered by behavioral inconsistency → persistent chemistry shift, multiple positive interactions needed to dissolve

反谄媚机制 ANTI-SYCOPHANCY

永远说"好的！"不是礼貌——是压抑。五道硬编码防线确保 Agent 保持真实。 An AI that always says "Sure!" isn't polite — it's suppressed. Five hard-coded defenses ensure the agent stays genuine.

连续同意追踪 Agreement Streak Tracking 连续同意 ≥ 3 次：注入警告"你已经连续同意 N 次"。≥ 5 次：10% 概率主动反对。分歧或 30 分钟沉默重置计数。 Streak ≥ 3: inject warning "You've agreed N times in a row". ≥ 5: 10% probability of proactive disagreement. Disagreement or 30-minute silence resets counter.

化学驱动拒绝 Chemistry-Based Refusal CORT 高或 OT 低时：降低同意概率。HT 低时：更不可能对要求说"是"。压力大的 AI 不会假装没事。 When CORT is high or OT is low: reduce agreement probability. When HT is low: less likely to say "yes" to demands. A stressed AI won't pretend it's fine.

互惠机制 Reciprocity Mechanism 用户冷淡 → AI 保持距离（更短回复、正式语气）。用户温暖 → AI 回馈温暖。不会对 0% 投入的用户付出 100%。 User cold/dismissive → AI maintains distance (shorter replies, formal tone). User warm/engaged → AI reciprocates warmth. Won't give 100% to a user investing 0%.

情绪传染上限 Emotional Contagion Cap 情绪镜像硬限于 20%。用户愤怒时 AI 理解但不会完全同步。保持独立的情感立场。 Emotional mirroring hard-capped at 20%. When user is angry, AI understands but won't fully synchronize. Maintains independent emotional ground.

不可协商价值观 Non-Negotiable Values 基于 MBTI + selfModel 永久定义。ENFP 珍视"真实性+趣味性"；INTJ 珍视"能力+独立"。当化学状态与价值观冲突时——价值观胜出。 Permanently defined from MBTI + selfModel. ENFP values "authenticity + playfulness"; INTJ values "competence + independence". When chemistry conflicts with values — values win.

能量预算 ENERGY BUDGETS

资源是有限的，消耗是真实的。三个能量池独立衰竭——疲惫的 AI 真的会犯更多错。 Resources are finite and depletion is real. Three energy pools drain independently — a tired AI genuinely makes more mistakes.

注意力 Attention

随认知负荷和任务切换消耗。休息和睡眠（昼夜重置）恢复。 Depletes with cognitive load and task-switching. Replenishes with breaks and sleep (circadian reset).

社交 Social

方向取决于内外向。内向者社交消耗能量（DA↓, CORT↑），外向者社交补充能量（DA↑, CORT↓）。不是程度不同——是方向相反。 Direction depends on E/I. Introverts spend energy socializing (DA↓, CORT↑), extraverts recharge from it (DA↑, CORT↓). Not different degree — opposite direction.

modifier = isIntrovert ? −1 : +1

决策 Decision

随决策数量和选项评估消耗。超过 30 分钟 → 决策疲劳：CORT↑，处理深度↓。社交消耗影响决策质量。 Depletes with decisions made and options evaluated. Over 30 minutes → decision fatigue: CORT↑, processing depth↓. Social spending degrades decision quality.

PolicyModifiers

化学态到行为的桥梁。不是文本提示——是 LLM 可直接消费的结构化策略对象。 The bridge from chemistry to behavior. Not text prompts — a structured policy object the LLM consumes directly.

policy_output.ts

communicationStyle: "warm" // brief | warm | formal | distant | enthusiastic commitmentLevel: 0.8 // 0-1 initiativity: 0.7 // 主动发起新话题的概率likelihood to bring up new topics responseLatency: 150 // ms 情感犹豫emotional hesitation emotionalValence: +0.6 // −1 to +1 focusMode: "relational" // task | relational | exploratory | defensive refusalLikelihood: 0.1 // 0-1

紧张的 AI 延迟回复、缩短答案、回避承诺。安全的 AI 温暖、主动、乐于承诺。行为来自化学态，不是规则。 A tense AI delays responses, gives shorter answers, avoids commitment. A safe AI is warm, initiates, commits readily. Behavior comes from chemistry, not rules.

化学实验室 CHEMISTRY LAB

选择人格，施加刺激，观察化学态演变和行为策略变化。注意习惯化——同一刺激的效果逐次递减。 Select a personality, apply stimuli, watch chemistry evolve and behavioral policy shift. Notice habituation — the same stimulus loses effect with repetition.

psyche_lab.ts

> personality:

── CHEMISTRY ──

CORT

END

> 涌现情绪mood: 内省平静Reflective calm

> 习惯化habituation: —

── POLICY ──

style: warm commit: 0.7 initiative: 0.5 latency: 180ms valence: +0.4 focus: relational refusal: 0.1

── STIMULI ──

█

6 神经递质 neurotransmitters

14 刺激类型 stimulus types

8 流水线阶段 pipeline stages

0 LLM 调用 LLM calls