Codebase (2026-06-27): this engine now lives in meno-gaokao-counsellor (
core/…), extracted from reflection-game — see gaokao-standalone-plan. Module paths below use the newcore/layout.
这是什么 · What this is
中文：忠实记录这个游戏当前是怎么把「反思均衡」落地的——怎么考虑用户画像、怎么提供信息、怎么撬动那些最少被审视的维度、怎么处理偏见。理论锚点：Ludwig, Mullainathan, Pink & Rambachan (2025)，行为经济学 2.0——算法帮助一个人达到他自己的反思均衡；这个工具只呈现，从不推荐。代码：
gaokao/{ranked,quiz,checklist}.py。EN: A faithful statement of how the game operationalizes 反思均衡 (reflective equilibrium) right now: how it considers the profile, how it supplies information, how it works the least-examined dimensions, and how it handles bias. Theory anchor: Ludwig, Mullainathan, Pink & Rambachan (2025), Behavioral Economics 2.0 — algorithms help a person reach their own reflective equilibrium; the instrument surfaces, never recommends. Code:
gaokao/{ranked,quiz,checklist}.py.
推演循环 · The loop (algorithm)
中文：整个流程围绕一份可排序的清单展开。入口收集 5 个「专业@学校」的排序 + 初始信心 + 画像问卷 + 自我描述（文字/语音）；随后走过 7 个因素站（课程→日常→兴趣→经济→城市→职业→代价），每一站先沉浸地演一段当前第 1 名在该因素上的、有据可查的第一人称场景，再调查第 1 名与一个强力对比项的差距，最后给学生自由重排的机会；一旦第 1 名变了，轨迹就转向新的第 1 名（已预取，瞬时切换）。7 站走完后做稳定性测试，最后给出新排序与总结。
EN:
intake: ranked list of 5 (专业@学校) + initial confidence + profiling quiz + free-text/voice
│
▼
for each of 7 factor-stages (course → daily → interest → finance → city → career → cost):
1. INFORM live a grounded, web-search-backed first-person scene of the current #1 on this factor
2. INVESTIGATE surface how #1 compares to one strong contender on this factor (the "你可能没考虑过…" tip was **retired 2026-06-17** — see [per-stage-anatomy Legacy](/projects/reflection-game/per-stage-anatomy))
3. (student may RE-RANK the list freely)
4. if #1 changed → the trajectory PIVOTS onto the new #1 (prefetched → instant)
│
▼
STABILITY: reframe the current #1 under fear / opportunity / authority (each = a bias) → does it sway you?
│
▼
END: new ranked list · initial→final shift · frames-stable count · 叶晓阳 6-item 反思均衡检查清单 · final confidence
中文：排序清单本身就是被测量的对象；重排是核心交互；每一站的存在意义都是「给出可能撼动排序的信息」。所谓均衡，就是一份在充分信息 + 偏见矫正之后依然站得住的排序。
EN: The ranked list is the measured object; re-ranking is the core interaction; each stage exists to give information that could move the ranking. Equilibrium = a ranking that holds after full information + bias correction.
怎么考虑用户画像 · How it considers the profile
中文：
PersonalityFile（core/profile.py）是每个用户的模型 = 问卷/测验得到的先验 + 文字/语音自述 + 用户每一次选择所携带的维度向量的滑动均值，落在 8 对极性的量表上（求真↔自保 … 严谨↔慈悲）。- 入口处由画像问卷（
core/quiz.py，精选的高信息量单选题）与文字/语音框共同构建；每一站都可实时编辑（🪞「你和 AI 共同理解的你」→/rank_profile），编辑会喂给下一站。 - 通过
serialize_for_prompt()注入每一次调查。护栏：它只影响学生会遇到哪些岔路/细节（相关性）——绝不影响哪个选项显得更对（不迎合 / 不造回音壁）。
EN:
PersonalityFile(core/profile.py): a per-user model = questionnaire/quiz prior + free-text/voice self-description + a running mean of the dimension-vectors of every choice, over the 8-pole rubric (TRUTH↔PROTECTION … RIGOR↔MERCY).- Built at intake from the profiling quiz (
core/quiz.py, curated max-info MCQs) and the free-text/voice box; editable live every stage (🪞「你和AI共同理解的你」 →/rank_profile), so edits feed the next stage. - Injected into every investigation via
serialize_for_prompt(). Guardrail: it steers which forks/details the student meets (relevance) — never which option looks good (no affirmation / echo chamber).
怎么提供信息 · How it provides information
中文：三层，事实优先：
- Grounding 知识库（
grounding.py，859 个专业）：每个专业一张「事实/观点」分离的卡片——课程、职业、路径、薪资、日常真实、踩坑点、转向。 - 联网检索调查（调查员 investigator）：针对确切的（因素，学校，专业）三元组，给出经权威来源核实的数字。
- 招生简介翻译（看懂这个专业）（已退役 2026-06-17）：官方介绍 →（说人话）→ 他们没告诉你的。 全部受具体性契约约束：要从真实细节里演绎；数字/雇主/统计必须来自 grounding 或权威来源，否则只能定性——绝不编造。
EN: Three layers, fact-first:
- Grounding KB (
grounding.py, 859 majors): per-major fact/opinion card — courses, careers, paths, salary, daily reality, failure modes, pivots. - Web-search investigation (调查员 investigator): authoritative-source-verified figures for the exact (factor, school, major).
- Admissions-blurb translation (看懂这个专业) (retired 2026-06-17): 官方介绍 → 说人话 → 没告诉你的.
All under a concreteness contract: dramatize from real specifics; numbers/employers/stats must come from grounding or an authoritative source, else stay qualitative — never fabricate.
怎么撬动最少被审视的维度 · How it works the least-examined dimension
中文：
- 7 因素的序列刻意走那些学生容易低估的维度——日常/群体生活、城市、可逆性、机会成本——而不只是显而易见的「学校 vs 专业」。
- 调查员是主动探针：每一站都把第 1 名和在该因素上差异最大的对比项对照，结尾落在**「你可能没考虑过…」**——把信息精准投向他们尚未掂量的地方。（这正是 LfL（向学习者学习）路线里的主动辨识：不被动旁观，而是主动构造能揭示/呈现差距的选择。）
- 因素与偏见的完整清单见 Ideal-Preference Checklist。
EN:
- The 7-factor sequence deliberately walks dimensions students under-weigh — daily/communal life, city, reversibility, opportunity cost — not just the obvious 学校 vs 专业.
- The investigator is the active probe: at each factor it contrasts the #1 against the contender most different on that factor, ending in “你可能没考虑过…” — information targeted at what they haven’t weighed. (This is the active-identification move from the LfL line: don’t passively watch, construct the choice that reveals/surfaces the gap.)
- The full inventory of factors + biases is the Ideal-Preference Checklist.
怎么处理偏见 · How it handles bias
中文：
- 框架稳定性测试（第 7 站之后）：把当前第 1 名放进三个劝说框架，每个绑定一种偏见——恐惧/损失厌恶、机会/乐观偏差、权威/权威服从——三者都在试图劝学生放弃第 1 名。
- 先测量后揭示：框架先以朴素的「别人可能会这样劝你放弃 X」呈现，不打标签；学生凭直觉反应；之后才揭示这些其实是偏见。（若一上来就贴标签，他们会有意识地抵抗，什么也测不到。）
- 框架不变性 = 一个反思均衡信号（「即使别人反对，我也不会轻易动摇」）；动摇则能定位是哪种偏见撼动了他。结尾会总结「框架稳定数」和他在哪些偏见上动摇了。
EN:
- The framing-stability test (after stage 7): the current #1 is reframed under three persuasion frames, each tied to a bias — 恐惧/损失厌恶, 机会/乐观偏差, 权威/权威服从 — all aimed at talking the student out of their #1.
- Measure-then-reveal: the frames are shown as plain “别人可能会这样劝你放弃 X” with no labels; the student reacts on instinct; only then are the biases revealed. (Labeling upfront would let them consciously resist and measure nothing.)
- Frame-invariance = a 反思均衡 signal (“即使别人反对,我也不会轻易动摇”); wavering localizes which bias moved them. The end summarizes frames-stable + the biases they wavered on.
均衡判据（停止条件）· The equilibrium check (stopping condition)
中文：结尾给出新的排序清单，并请学生对照叶晓阳的 6 项反思均衡检查清单（了解学什么课程 / 知道做什么工作 / 考虑过放弃什么 / 和在读者聊过 / 不轻易动摇 / 排过城市-学校-专业优先级）做自检，并给最终信心打分（对比初始信心）。一个在充分信息 + 偏见矫正之后依然站得住的决定——「我能接受，我不后悔」——就是操作化的反思均衡。
EN: The end gives the new ranked list and asks the student to self-check against 叶晓阳’s 6-item 反思均衡检查清单 (了解学什么课程 / 知道做什么工作 / 考虑过放弃什么 / 和在读者聊过 / 不轻易动摇 / 排过城市-学校-专业优先级) and to rate final confidence vs their initial one. A decision that holds after full information + bias correction — “我能接受,我不后悔” — is the operational reflective equilibrium.
记录了什么（数据）· What’s logged (the data)
中文：每个会话 → session_log.jsonl：5 个选项、初始→最终排序、每一次重排（站次 + 从/到 + 第 1 名是否变化）、每站的因素/第 1 名/对比项/来源、框架测试结果（被哪些偏见撼动）、检查清单、以及 confidence_initial→final。这个工具本身就是数据集。
EN: Per session → session_log.jsonl: the 5 options, initial→final ranking, every re-rank (stage + from/to + 1-changed), per-stage factor/top/contender/sources, frame-test results (biases swayed), checklist, and confidence_initial→final. The instrument is the dataset.