Reflection Game — 高考志愿填报 variant (plan + MVP built)

一句话 · One-line

中文：一个面向高考志愿填报（选学校/选专业）的反思游戏变体：当一个学生在好学校差专业与差学校好专业之间纠结时，卡点不是信息，而是对一个自己还无法切身感受的未来所做的判断。让他们在每个选项下亲历一段模拟的多年人生轨迹，然后从这段亲历经验所揭示的、关于他们自己价值观的东西出发，回溯推理当下的选择。依照反思游戏的核心原则，它浮现他们的偏好——而不做推荐。

EN: A reflection-game variant for 高考志愿填报 (major/school choice): when a student is torn between 好学校差专业 vs 差学校好专业, the blocker isn’t information, it’s judgment about a future they can’t yet feel. Let them live a simulated multi-year trajectory under each option, then backward-reason the present choice from what the lived experience revealed about their own values. Per the reflection-game core principle, it surfaces their preference — it does not recommend.

构建状态 — MVP 已建成（2026-06-11，自主运行）· Build status — MVP BUILT (2026-06-11, autonomous run)

端到端可玩 · Playable end-to-end. Branch gaokao-variant in meno-sh/reflection-game (not yet merged).

中文：整个闭环的可运行 MVP 已存在并能实时运行：录入 → 实时跑出两种未来 → 回溯推理之镜。

EN: A working MVP of the full loop exists and runs live: intake → live BOTH futures → backward-reasoning mirror.

中文：已建成的内容（gaokao/ 包，复用涌现式世界引擎作为人生底座）：

• trajectory.py —— 人生阶段底座：一个「回合」= 一个关键人生阶段（大一→毕业→求职→回望）；模型为每个阶段生成场景 + 2–3 个带价值标签的选项；StateManager 让人生跨阶段保持连贯（已验证：室友、GPA、专业挣扎这条线索都会延续）。遇 API 错误会带退避重试。
• profile.py —— PersonalityFile（问卷 + 自由文本 + 所选维度向量的滑动均值），注入是为了相关性，而非肯定（防回音室护栏）。
• compare.py —— 确定性的跨人生信号（一个价值在两个选项世界里都被往同一方向推 = 揭示出一项核心价值；翻转的那种 = 依情境而定）+ 一个 LLM 镜子，把这些价值作为待检验的假设反射回去，绝不推荐任何一个选项。
• demo.py —— run_pipeline：两个选项世界都跑一遍，结尾 = 对两条维度轨迹做镜像反思。
• web.py —— 可玩的异步网页应用（录入表单 + 8 个价值滑块 → 流式跑出两段人生 → 镜子）。Live demo: https://easy-justify-implementation-piece.trycloudflare.com（临时隧道；目前未加门禁）。
• school_tier.py + real_data/school_index_sample.json —— 真实录取数据（来自公开的 labolado/gaokao_2016-2020）：约 490 所学校，带 985/211/双一流 标记 + 分数线；build_gaokao_seed 用学校的真实层级给每段人生打地基，而不是靠猜。

EN: What’s built (gaokao/ package, reusing the emergent world-engine as the life substrate):

• trajectory.py — life-stage substrate: a “turn” = a pivotal life stage (大一→毕业→求职→回望); the model invents each scene + 2–3 value-tagged options; the StateManager keeps the life coherent across stages (verified: roommate, GPA, the major-struggle thread persist). Retries API errors w/ backoff.
• profile.py — PersonalityFile (questionnaire + free-text + running mean of chosen dimension vectors), injected for relevance, not affirmation (echo-chamber guardrail).
• compare.py — deterministic cross-life signal (a value pushed the same way in both option-worlds = a revealed core value; one that flips = context-dependent) + an LLM mirror that reflects those values back as a hypothesis to check, never recommending an option.
• demo.py — run_pipeline: both option-worlds run, ending = the mirror over both dim-traces.
• web.py — playable async web app (intake form + 8 value sliders → streams both lives → mirror). Live demo: https://easy-justify-implementation-piece.trycloudflare.com (ephemeral tunnel; un-gated for now).
• school_tier.py + real_data/school_index_sample.json — real admission data (from public labolado/gaokao_2016-2020): ~490 schools with 985/211/双一流 flags + cutoffs; build_gaokao_seed grounds each life in the school’s real tier instead of guessing.

中文：关键发现：

• 单回合质量很强，且「没有标准答案」原则成立（镜子点出跨人生的模式，把它框定为假设，什么都不推荐）。
• 延迟是产品的墙： V4-Pro 约 50 秒/阶段；把两段人生并行几乎没帮上忙（在一个 key 上，供应商对并发的推理调用做了限流）。DeepSeek-V3（非推理）约 19 秒/阶段 = 在可比质量下快 2.5–3 倍 → 现已成为轨迹生成的默认（镜子继续用 V4-Pro）。
• 共享的 OpenRouter key 在负载下会 429 → 真要上线的产品需要更高档位 / 多个 key。

EN: Key findings:

• Single-turn quality is strong and the no-right-answer principle holds (the mirror names the cross-life pattern, frames it as a hypothesis, recommends nothing).
• Latency is the product wall: V4-Pro ~50s/stage; parallelizing the two lives barely helped (provider throttles concurrent reasoning calls on one key). DeepSeek-V3 (non-reasoning) ~19s/stage = 2.5–3× faster at comparable quality → now the default for trajectory gen (the mirror keeps V4-Pro).
• The shared OpenRouter key 429s under load → a launched product needs a higher tier / multiple keys.

中文：待定的分叉（等决策）： MVP 会按用户陈述的滑块倾向自动选定每个阶段（一路直跑到底）。下一个大动作要么是 缓存原型路径（近乎即时），要么是 交互式逐阶段 v2（揭示 > 陈述的偏好）。

EN: Open fork (awaiting decision): the MVP auto-picks each stage from the user’s stated slider-leaning (runs straight through). Next big build is either the cached-archetype path (near-instant) or the interactive per-stage v2 (revealed > stated preference).

为什么它契合这套工具 · Why this fits the instrument

中文：反思游戏是一台偏好引出仪器，不是推荐器：它通过让你亲历选择、而非抽象地排序选择，来浮现你心之所向，并且不泄露任何「标准答案」。志愿填报决策恰恰是这样一类情形——陈述出来的偏好不可靠（名校偏见、父母压力、同辈噪声），真正要紧的是亲历出来的偏好。模拟轨迹把一个 18 岁的抽象猜测，变成一个有过经历的判断。

EN: The reflection game is a preference-elicitation instrument, not a recommender: it surfaces what’s close to your heart by making you live choices rather than rank them abstractly, and it leaks no “right answer.” A 志愿填报 decision is exactly a case where stated preferences are unreliable (prestige bias, parental pressure, peer noise) and lived preference is what matters. Simulating the trajectory turns an abstract 18-year-old guess into an experienced one.

功能 · Features

中文：

1. 录入 / 个性化。 学生输入他们的两个真实选项（学校 A + 专业 X 对 学校 B + 专业 Y）、兴趣、约束（城市、家庭、预算、风险承受度）。这给两条轨迹打种子——泛泛的原型不会让人投入。
2. 两条并行的人生轨迹，每个选项一条，作为一段压缩的多年弧线来演（本科 → 实习 → 第一份工作 → ~5–10 年视野）。每条不是固定故事，而是由涌现式世界引擎从一颗薄薄的种子（画像 + 选项 + 真实结果统计）实时生成，在真正要紧的时刻有分叉。
3. 引出价值的分叉，映射到一条（隐形的）偏好脊柱，由既有的 8 维模型改写到人生决策轴——例如激情 ↔ 务实、安全 ↔ 成长、声望 ↔ 契合、靠近家人 ↔ 独立、深度 ↔ 广度、金钱 ↔ 意义。每个生成的选项都带一个经过验证的维度标签（跨玩家研究的效度挂钩）。
4. 亲历后果的质感（对称）。 两条路都获得同等丰富的好处与坏处：好学校差专业 → 强人脉/品牌，但多年学着自己厌恶的东西 → 分叉（转专业 / 熬过去 / 借品牌转行）；差学校好专业 → 你热爱这个领域，但人脉更薄 → 分叉（出类拔萃脱颖而出 / 感到资源不足 / 转学）。学生从两边都感受到这个取舍。
5. 回溯推理的反思结尾。 亲历两段之后，仪器把轨迹所揭示的反射回去——哪段亲历经验有共鸣，哪些价值在压力下占了上风——并把它映射到当下的选择，作为镜子、而非神谕（「你的选择一直偏向契合而非品牌」，而不是「选 B」）。
6. 比较模式（相对单弧游戏的新东西）：玩两条轨迹，然后做一步比较/反思。可选一个「互换」——把某个分叉换个方向重玩——来测试揭示出的偏好有多稳定。
7. 中文优先。 双语基建（?lang）已经把中文作为主界面支持；弧线/种子以中文撰写。

EN:

1. Intake / personalization. The student enters their two real options (school A + major X vs school B + major Y), interests, constraints (city, family, budget, risk tolerance). This seeds both trajectories — generic archetypes won’t engage.
2. Two parallel life-trajectories, one per option, played as a compressed multi-year arc (undergrad → internships → first job → ~5–10 yr horizon). Each is not a fixed story but generated live from a thin seed (profile + option + real outcome stats) by the emergent world-engine, with branching forks at the moments that actually matter.
3. Value-eliciting forks, mapped to an (invisible) preference spine adapted from the existing 8-dim model to life-decision axes — e.g. passion ↔ pragmatism, security ↔ growth, prestige ↔ fit, proximity-to-family ↔ independence, depth ↔ breadth, money ↔ meaning. Every generated option carries a validated dimension tag (the cross-player research validity hook).
4. Lived-consequence texture (symmetric). Both paths get equally rich upside and downside: 好学校差专业 → strong network/brand but years studying something you resent → forks (switch major / endure / pivot career via the brand); 差学校好专业 → you love the field but a thinner network → forks (excel and stand out / feel under-resourced / transfer). The student feels the trade-off both ways.
5. Backward-reasoning reflection ending. After living both, the instrument mirrors back what the trajectories revealed — which lived experience resonated, which values dominated under pressure — and maps that to the present choice as a mirror, not an oracle (“your choices kept favouring fit over brand,” not “pick B”).
6. Comparative mode (new vs the single-arc game): play both trajectories, then a compare/reflect step. Optionally a “swap” — replay a fork the other way — to test how stable the revealed preference is.
7. Chinese-first. The bilingual infra (?lang) already supports 中文 as the primary surface; arcs/seeds authored in Chinese.

数据从哪来 · Where to get the data

中文：

• 录取/可行性数据（哪个选项才现实）：院校/专业录取分数线、招生计划、选科要求 —— 公开的阳光高考 / 各省考试院数据集，或一次性爬取成一张查找表，录入时用它来校验。
• 给轨迹打地基的结果数据（让模拟不是空想）：毕业生就业质量报告（多数高校都会发布）、专业就业率/对口率/薪资（麦可思/MyCOS 式报告），以及亲历经验语料（知乎/小红书/贴吧上关于「好学校差专业 vs 差学校好专业」的帖子、转行故事）—— 作为生成器的检索地基，而非逐字照搬。
• 轨迹生成： 涌现式引擎（Track B）从一颗薄种子 + 画像生成出可信的人生路径，以上面的结果数据为条件（RAG 式），让分叉和后果是真实的、而非编造的。数据约束，LLM 演绎。
• 价值/分叉内容： 改写既有弧线的分叉语法 + 8 维脊柱；撰写 3–5 个志愿填报种子骨架（原型决策形状）来锚定生成器。

EN:

• Admission/feasibility data (which option is even realistic): 院校/专业录取分数线, 招生计划, 选科要求 — public 阳光高考 / 各省考试院 datasets, or a one-time scrape into a lookup the intake validates against.
• Outcome data to ground trajectories (so simulations aren’t fantasy): 毕业生就业质量报告 (most universities publish these), 专业就业率/对口率/薪资 (麦可思/MyCOS-style reports), and lived-experience corpora (知乎/小红书/贴吧 threads on “好学校差专业 vs 差学校好专业”, career-switch stories) — used as retrieval grounding for the generator, not verbatim.
• Trajectory generation: the emergent engine (Track B) generates the plausible life path from a thin seed + the profile, conditioned on the outcome data above (RAG-style) so forks and consequences are realistic, not invented. The data constrains; the LLM dramatizes.
• Value/fork content: adapt the existing arcs’ branching grammar + the 8-dim spine; author 3–5 志愿填报 seed-scaffolds (archetypal decision shapes) to anchor the generator.

相对当前游戏改了什么 · What changes from the current game

中文：

• 时间尺度： 当前的弧线是一个瞬间（约 1 小时，一个场景→决策）；这个需要多年弧线。→ 使用/扩展涌现式世界引擎（Track B）——它的世界状态账本 + 开放式回合生成正是一套人生轨迹底座；emergent_turn_cap（=12）给长度封顶。
• 生产环境上真实 LLM： 轨迹生成要求走 LLM 路径，所以被搁置的 Track-B 阻塞项——Render 环境上的 OPENROUTER_API_KEY——成了硬性前置条件（生产环境目前无 key → 今天用 Mock）。
• 偏好脊柱： 把 8 个维度重新映射到人生决策轴（见上）；保持它们隐形（研究效度的脊柱），但验证每个生成的选项都带标签。
• 新界面： 录入表单；比较型双轨迹流程；与当下选择挂钩的回溯推理结尾（当前的结尾只给玩家画像）。
• 地基层： 一步检索（结果数据 → 生成器上下文），这是已撰写的弧线所没有的。
• 存档/续玩（刚上线）在这里更要紧——多年轨迹很长，所以续玩是必需的。

EN:

• Time-scale: current arcs are one moment (~1h, one scene→decision); this needs a multi-year arc. → use/extend the emergent world-engine (Track B) — its world-state ledger + open-ended turn generation is exactly a life-trajectory substrate; the emergent_turn_cap (=12) bounds length.
• Real LLM on prod: trajectory generation requires the LLM path, so the parked Track-B blocker — OPENROUTER_API_KEY on the Render env — becomes a hard prerequisite (prod is keyless → Mock today).
• Preference spine: re-map the 8 dims to life-decision axes (above); keep them invisible (the research-validity spine) but validate every generated option carries a tag.
• New surfaces: intake form; comparative two-trajectory flow; backward-reasoning ending that ties to the present choice (the current ending only profiles the player).
• Grounding layer: a retrieval step (outcome data → generator context) that the authored arcs don’t have.
• Save/resume (just shipped) matters more here — multi-year trajectories are long, so resuming is essential.

问题以及我们怎么解决 · Problems & how we solve them

中文：

1. 编造 / 假精确。 一条生成的轨迹可能读起来像预言。→ 把每个分叉/后果都锚定在结果语料里（RAG）；在涌现式效度检查里加一道真实性闸门；把整件事明确框定为一场为了感受取舍的模拟，而非预测——浮现不确定性，绝不给数字。
2. 没有标准答案的完整性（核心原则）。 仪器绝不能倾向某一个选项（尤其是「好专业」的陈词滥调）。→ 对称生成（两条路获得同等丰富的好处/坏处）；维度标签的中立纪律（没有分叉泄露一个「开明」的选择）；反思是对他们揭示出的价值的一面镜子，绝非推荐。
3. 效度——模拟的经验真能揭示真实偏好吗？ → 保留隐形的 8 维脊柱以支持跨玩家可比性；用既有仪器验证这个变体（它能否复原同一个人的画像？）；把回溯推理的输出当作学生拿来对照自己检验的假设，而非判决。
4. 长度 / 投入度。 两条多年弧线可能拖沓。→ emergent_turn_cap + 压缩（跳到要紧的分叉）；存档/续玩；让学生在某个分叉解决后跳过一条路。
5. 无真值的质量评估（悬而未决的 Track-B 问题）。 这里同样适用。→ MVP：在几个种子形状上做人工抽查，看真实性 + 中立性；更长期靠悬而未决的 Track-B 质量评估工作。
6. 中文情境的冷启动数据。 → MVP 不需要详尽数据：3–5 个手写种子骨架（常见的 好/差 × 学校/专业 原型）锚定在一个小的精选结果集上，再加生成器做个性化。等第一个可用版本之后再扩展数据层。

EN:

1. Fabrication / false precision. A generated trajectory could read like prophecy. → Ground every fork/consequence in the outcome corpus (RAG); add a realism gate in the emergent validity check; frame the whole thing explicitly as a simulation to feel the trade-off, not a prediction — surface uncertainty, never a number.
2. No-right-answer integrity (the core principle). The instrument must not tilt toward one option (esp. the “good major” cliché). → Symmetric generation (both paths get equal upside/downside richness); the dimension-tag neutrality discipline (no fork leaks an “enlightened” choice); the reflection is a mirror of their revealed values, never a recommendation.
3. Validity — does simulated experience reveal real preference? → keep the invisible 8-dim spine for cross-player comparability; validate the variant against the existing instrument (does it recover the same person’s profile?); treat the backward-reasoning output as a hypothesis the student checks against themselves, not a verdict.
4. Length / engagement. Two multi-year arcs could drag. → emergent_turn_cap + compression (jump to the forks that matter); save/resume; let the student skip a path once a fork resolves.
5. Quality eval without ground truth (the open Track-B problem). Applies here too. → MVP: human spot-checks for realism + neutrality on a few seed shapes; longer-term, the open Track-B quality-eval work.
6. Cold-start data for the Chinese context. → MVP doesn’t need exhaustive data: 3–5 hand-authored seed-scaffolds (the common 好/差 × 学校/专业 archetypes) grounded in a small curated outcome set, + the generator for personalization. Expand the data layer after the first usable build.

MVP 切片（最小的真东西）· MVP cut (smallest thing that’s real)

中文：

• 录入（学生的两个真实选项 + 几个画像字段）→ 种子。
• 涌现式引擎（带 key）从一个精选种子骨架 + 一个小的有地基的结果集，给每个选项生成一条轨迹，带价值标签的分叉。
• 一个比较 + 回溯推理的中文结尾。
• 前置条件：Render 上的 OpenRouter key。建在既有仓库之上（core/emergent.py、世界状态账本、8 维脊柱、存档/续玩）。

EN:

• Intake (the student’s two real options + a few profile fields) → seeds.
• The emergent engine (keyed) generating one trajectory per option from a curated seed-scaffold + a small grounded outcome set, with value-tagged forks.
• A compare + backward-reasoning ending in 中文.
• Prereq: OpenRouter key on Render. Built on the existing repo (core/emergent.py, the world-state ledger, the 8-dim spine, save/resume).

悬而未决的设计抉择（需要人）· Open design calls (need a human)

中文：

• 把 8 个维度重新映射到人生决策轴（哪 8 个？）。
• 回溯推理要多直白（「你的选择偏向契合而非品牌」——直白到什么程度它就不再是镜子了？）。
• 在允许生成器演绎之前，要求多少真实数据（地基/真实性阈值）。

EN:

• Re-mapping the 8 dims to life-decision axes (which 8?).
• How explicit to make the backward-reasoning (“your choices favoured fit over brand” — how directive before it stops being a mirror?).
• How much real data to require before the generator is allowed to dramatize (the grounding/realism threshold).

相关 · Related

• [!] Quality Overhaul Plan (2026-06-13) — root-cause fixes for genericness / grounding / consistency (in progress).
• CN) — the complete per-turn pipeline + prompt anatomy.
• Reflection Game · Learning-from-Learner Theory