它为每个客户做的每一个判断、走的每一步、为什么这么做，运营者随时看得到、复盘得了。从一次回复的来龙去脉，到一个客户的画像演化，再到整盘系统的健康态势——三层都摊开。
Every judgment it makes for every customer, every step it takes, and why — visible and reviewable anytime. From the story behind a single reply, to how a customer's profile evolves, to the health of the whole system — all three layers laid open.
客户发来一条消息，AI 这一次到底想了什么、查了什么、为什么这么回——整个经过像回放一样看得清清楚楚。
A customer sends a message; what the AI thought, checked, and why it replied this way — the whole process plays back, crystal clear.
这条回复最终是发出了、先压住了，还是被拦下了，理由写得明明白白。比如"涉及还没核实的产品说法，按规矩先压住、不乱讲"——你一看就懂它为什么这么处理。
Whether the reply was sent, held back, or blocked — with the reason spelled out plainly. E.g. "touches an unverified product claim, so it's held by rule, not winged" — you instantly see why.
这条回复是基于知识库里哪几条已核实资料说出来的，都能回看；还会标出"这次缺哪块知识"，提示你去补。这是"AI 不会信口开河"最硬的证据。
Which verified pieces of knowledge the reply stood on — all reviewable; it even flags "what knowledge was missing this time" so you can fill the gap. The hardest proof that the AI doesn't make things up.
独立评审会从"靠不靠谱、像不像真人、有没有温度、会不会给压力、产品说法准不准"几个角度打分；第一版不达标时，AI 自动改写一次，改前改后都摆给你看。
An independent review scores it on reliability, human feel, warmth, pressure, and product accuracy; if the first draft falls short, the AI rewrites once — and shows you before and after.
每次决策 AI 都留一段"我这么做对不对、还能怎么更好"的自评，你看得到，也看得到"上次提到的问题这次有没有改进"。这一次用了几次大模型、用量多少、有多少命中缓存省下来，也一并透明（不折算成具体金额）。
Each decision leaves a note — "did I do right, how could I do better" — and you can see whether last time's issue improved. How many model calls this took, the usage, and how much was saved by cache hits are all transparent too (no dollar figure attached).
AI 对每个客户的认知分层级、有出处、标置信。最关键的一条诚实边界：它只观测、只沉淀，绝不拿没把握的猜测去预设客户、左右话术——这既避免贴标签偏见，又让你越用越懂这个人。
Its read on each customer is tiered, sourced, and confidence-marked. The key honest boundary: it only observes and records — it never lets an unsure guess preset the customer or steer the wording. That avoids label bias and helps you understand the person better over time.
你对这个客户下的判断，AI 永远改不动、坚信不疑、原样带进每一次决策。这是最高权威，AI 不会"自作聪明"覆盖。
Your call on the customer: the AI can never change it, trusts it absolutely, and carries it verbatim into every decision. Top authority — never "optimized" away by the AI.
驱动决策Drives decisions每一条都挂着原始聊天证据，找不到出处就直接丢弃。它会进决策，但门槛是"拿得出依据"，不是 AI 随口一说。
Each carries the original chat evidence; no source, no keep. It feeds decisions, but the bar is "can show receipts," not the AI's say-so.
有证据才驱动Drives — with evidence证据还不够、置信还不高的猜测。只记录、供你参考，永远不影响 AI 怎么回复。一眼看清哪些是人定的、哪些 AI 有据、哪些只是参考。
Guesses without enough evidence or confidence. Recorded for your reference only, and never affect how the AI replies. You see at a glance what's human-set, what's grounded, what's just a hint.
只记录 · 不驱动Record only · never drives用心理学公认的"大五人格"五个维度，慢慢摸清每个客户的性格倾向，形成画像供你参考。它只观测、不替客户预设回应；证据不够时如实标注"还不确定"。
Using the well-established Big Five dimensions, it gradually reads each customer's traits into a profile for your reference. It observes only, never presets responses; when evidence is thin it honestly marks "not sure yet."
AI 对客户的一些关键判断，会随多轮聊天累积证据、连成走势线，置信高低、是否已经站稳都标得清清楚楚。同样只供观测复盘，不驱动回复。
Key judgments accrue evidence across turns into a trend line, with confidence and whether it's settled clearly marked. Also observation-and-review only — it doesn't drive replies.
AI 给每个客户记的长期记忆卡——核心事实、近期动态、已过时信息分层——以及还在观察、尚未正式入卡的候选记忆，你都能查、能管。
The long-term memory card the AI keeps per customer — core facts, recent moves, outdated info, tiered — plus candidate memories still under observation, all reviewable and manageable.
把 AI 对这个客户的几项关键评估打成 0–100 的健康分，红黄绿一目了然。注意：这些分只是给你看的体检表，不会反过来左右 AI 怎么说话。
It turns several key assessments into 0–100 health scores, red-amber-green at a glance. Note: these scores are a check-up for you to read — they never loop back to steer how the AI talks.
成交了没有（AI 永远不会自己宣布成交，必须由运营者确认）、主动发出去的内容客户有没有回应、有没有推进关系，加上回复率、对话深度这些日常成效，连成一张看得见的成效账——干得好不好，是数据说了算。
Whether a deal closed (the AI never declares a deal itself — an operator must confirm), whether proactive messages drew responses, whether the relationship advanced, plus reply rate and conversation depth — woven into one visible ledger. How well it's doing is the data's call.
从单个客户抬起头，看整盘：AI 整体表现、这一天的态势、它琢磨出的新想法、怎么自我进化、知识库健不健康、底层有没有稳稳跑着——管理者要的全局视角，都在这里。
Lift your eyes from one customer to the whole board: overall AI performance, the day's posture, the new ideas it surfaced, how it self-evolves, knowledge-base health, and whether the plumbing runs steady — the manager's bird's-eye view, all here.
AI 主动暂缓的三类细分（策略暂缓 / 安全拦截 / 等更多信息）、没核实的产品说法被拦了多少、自我反省的改进率、发送成功与取消的比例——AI 整体表现一屏看住。
The three kinds of holds (policy hold / safety block / awaiting more context), how many unverified product claims got stopped, the self-reflection improvement rate, and send-vs-cancel ratio — overall AI performance in one view.
这一天 AI 跑了多少次、分别停在哪个环节、有多少正等着你拍板——积压一眼看到，及时介入。
How many runs today, where each one paused, and how many await your call — backlog spotted at a glance.
AI 自己琢磨出来的新标签、新关系类型、疑似成交线索，都进"待审"区，你审了才生效——它永远不会自作主张。
New tags, relationship types, and suspected deal signals the AI surfaces all land in a review queue — they take effect only after you approve. It never acts on its own.
AI 从历史里学、提出"这样调会不会更好"的建议，先在影子环境拿历史对话重跑验证（零副作用、不碰真实客户）；只有确实更好、过了安全回归、再经你二次确认才真正上线，还能一键回滚。全程透明可查。
It learns from history and proposes "would this tweak help?", first replaying past chats in a shadow run (zero side effects, no real customers touched); only if it's truly better, clears safety regression, and you confirm does it go live — with one-click rollback. Fully transparent.
每条知识改过几次、谁改的、改了什么都留痕可回溯；哪些还没核实、哪些有矛盾或过时，系统会标出来提示补。AI 写进来的知识一律先标"待核实"，从不自说自话当成真。
Every edit to a knowledge item — how many times, by whom, what changed — is traceable; what's unverified, conflicting, or stale gets flagged for follow-up. Anything the AI writes in is marked "to be verified" first — never self-certified as true.
哪些微信号在线、跟进任务排了多少、信号采集有没有断流（断了会告警）、发送有没有触顶 / 退避的提醒——保障整盘稳稳跑着。
Which WeChat accounts are online, how many follow-up tasks are queued, whether signal intake stalled (alerts if so), and send cap-hit / back-off alerts — keeping the whole board running steady.
全自治不等于没有边界。客户始终只跟 AI 对话；遇到超出 AI 职权或能力的事，它不装懂、不乱答，而是向幕后决策源请示、拿回结论后用自己的口吻转述客户。这一层，就是看懂 AI 怎么守住分寸。
Full autonomy is not the absence of boundaries. The customer only ever talks to the AI; when something exceeds the AI's remit or ability, it doesn't fake it or wing an answer — it consults the decision-maker behind the scenes, then relays the conclusion to the customer in its own voice. This layer is about how the AI keeps within its limits.
批准、驳回、有条件批准、退回再议——每一种都看得到决策者给的口径、授权窗口和约束条件。AI 严格按这个口径转述，不擅自加码。
Approve, decline, approve-with-conditions, send back for more thought — for each you can see the terms, the authorization window, and the limits the decision-maker set. The AI relays strictly to that brief, never adding on its own.
哪些事项触发了请示、由谁拍板、走的什么渠道（决策者微信、后台直接裁决、决策者对话）、结论是什么，全程留痕可回溯；需要时还能改派给备选决策人。
What triggered the consult, who decided, through which channel (the decision-maker's WeChat, a direct call in the back office, or a chat with them), and the conclusion — all traceable; you can also reassign to a backup decision-maker when needed.
某些高价值时刻——客户明确要签约、想到店参观——AI 会主动把真人专属顾问的微信名片递给客户，自己退到辅助答疑。哪些客户已被引荐、它何时退的辅助、每张名片覆盖了多少客户、回应如何，你都看得到。台前顾问负责临门一脚，与幕后决策者是两回事。
At certain high-value moments — a customer clearly ready to sign, or wanting to visit in person — the AI proactively hands over a real advisor's WeChat card and steps back into a support role. Which customers got an intro, when it stepped back, how many customers each card reached, and the responses — all visible. The front-stage advisor closes the last mile, and is separate from the behind-the-scenes decision-maker.
可选 · 默认关闭的辅助模式Optional · assist mode, off by default你用日常的话给后台 AI 下指令（比如"给最近问过价的客户发个新品提醒"），它会拆成一步步要做的事逐个执行，每一步成功没成功都摊开。可以先"演练一遍"只看它打算做什么、并不真发；高风险的计划必须你二次确认才会执行。
Give the back-office AI an instruction in everyday words (say, "send a new-arrival nudge to customers who recently asked about price"), and it breaks the job into steps and runs them one by one, each step's success or failure laid open. You can "dry-run" first to see what it intends without anything actually sending; any high-risk plan runs only after you confirm a second time.
看得清，是为了管得住。看看这些可观测之下，红线如何写进代码、认知方法与自我演化如何工程落地。
Seeing clearly is what makes it governable. See how the red lines are compiled into code, and how the cognition and self-evolution are engineered beneath what you observe.