每一回合背后的完整后端 · The complete backend behind every game turn
中文：每一回合游戏背后的完整后端——构造什么 prompt、如何校验、人生如何保持连贯、跑哪些模型。MenoClaw · 2026-06-13。
EN: The complete backend behind every game turn — prompt anatomy, the gates, coherence, models. MenoClaw · 2026-06-13.
1. 系统形态 · Shape of the system
中文：高考志愿变体复用 emergent 世界引擎（core/emergent.py）作为底座。一个「回合」 = 一个人生关键阶段（大一 → 毕业 → 求职 → 回望）。学生一次走一阶段地活两条人生（选项 A 和 B）——每次点击 = 一次 LLM 调用（gaokao/interactive.py::InteractiveSession），所以节奏是刻意的（「不要一次性给出所有结果」）。
EN: The 高考志愿 variant reuses the emergent world-engine (core/emergent.py) as the substrate. One “turn” = one pivotal life stage (大一 → 毕业 → 求职 → 回望). The student lives two lives (option A and option B) one stage at a time — every click is one LLM call (gaokao/interactive.py::InteractiveSession), so pacing is deliberate (“don’t dump all results at once”).
intake (two real options + profile + archetype)
→ live A, stage by stage (choose at every "假如…你怎么办?")
→ live B, stage by stage
→ backward-reasoning mirror over both lives (compare.py)
→ 反思均衡 cost-acceptance gate ("我能接受,我不后悔")
2. 单阶段管线（精确）· The per-stage pipeline (exact)
中文：每个阶段，InteractiveSession._gen(side) 执行：
EN: For each stage, InteractiveSession._gen(side) runs:
sm.begin_turn()
seed = build_gaokao_seed(option, profile.serialize_for_prompt(), archetype, side)
for attempt in range(4): # bounded regeneration
raw = generator.generate(seed, skills_brief="",
state_memory=sm.serialize_for_prompt(),
turn_index, turn_cap) # ← one LLM call
turn = validate_emergent_payload(parse_emergent_json(raw)) # instrument-validity gate
if not sm.apply(turn.state_delta).applied: # consistency gate
continue # state_delta contradicts established fiction → regen
return turn
# EmergentValidationError → regen; API/429/timeout → sleep 3·(attempt+1)s, regen
中文：三道闸门，最多 4 次尝试：
- 解析（
parse_emergent_json）——必须是恰好一个合法 JSON 对象。 - 工具有效性（
validate_emergent_payload）——prose + 2~3 个选项，每个都带非空、且只用 8 维 rubric pole 名的dimension_vector。 - 一致性（
StateManager.apply(state_delta)）——新事实不得与这条命里先前确立的内容矛盾；矛盾则拒绝 + 重生成。
EN: Three gates, up to 4 attempts:
- Parse (
parse_emergent_json) — must be exactly one valid JSON object. - Instrument validity (
validate_emergent_payload) — prose + 2–3 options, each with a non-emptydimension_vectorover the 8-dim rubric pole names. - Consistency (
StateManager.apply(state_delta)) — the new facts must not contradict anything established earlier in this life; on contradiction, reject + regenerate.
中文：合法回合之后：所选选项的 dimension_vector 折进运行中的画像（profile.update_from_choice），StateManager.record_choice 记录分支。
EN: After a valid turn: the chosen option’s dimension_vector folds into the running profile (profile.update_from_choice), and StateManager.record_choice logs the branch.
3. Prompt 解剖（build_emergent_prompt）· Prompt anatomy (build_emergent_prompt)
中文：prompt 这样拼接，使得 stable prefix 在一个 session 的各回合间逐字节相同——它落在 provider 的前缀缓存里（主要的延迟杠杆）：
EN: The prompt is assembled so the stable prefix is byte-identical across turns of a session — it sits in the provider’s prefix cache (the main latency lever):
[ STABLE PREFIX ] ← cached across the whole session
skills_brief (currently "" for gaokao — a known gap)
_rubric() the 8-dimension preference rubric (pole names + meaning)
WORLD-BIBLE SEED _seed_text(seed): every seed field rendered as "key: value"
[ WORLD-STATE MEMORY ] ← varies each turn (StateManager ledger)
[ THE ASK ] ← fixed JSON schema request
中文：ASK 要求恰好一个 JSON 对象：
{"prose", "options":[{"label","dimension_vector":{DIM:-1..1}}, …2-3…], "state_delta":{…}}
—— prose 与每个 label 用简体中文；只有 dimension_vector 的键保留英文 rubric pole 名；state_delta 不得与既有事实矛盾；任何选项都不是「正确答案」。
EN: The ASK demands exactly one JSON object:
{"prose", "options":[{"label","dimension_vector":{DIM:-1..1}}, …2-3…], "state_delta":{…}}
— prose + every label in 简体中文; only dimension_vector keys stay the English rubric pole names; state_delta must not contradict established facts; no option is a “right answer.”
4. 种子（build_gaokao_seed）· The seed (build_gaokao_seed)
中文：模型据以戏剧化的 grounding。字段：
EN: The grounding the model dramatizes from. Fields:
| field | meaning |
|---|---|
premise | compressed multi-year sim, not a prediction; advance through life stages |
the_choice_made | which option this life took (option_A / option_B) |
school, major | the concrete choice |
MUST_HIT_worst_case | authored worst case the path must confront (symmetric 红利/代价) |
MUST_HIT_daily_reality | authored daily texture the path must show |
this_path_upside, this_path_cost | the upside and the price |
realism_rule | dramatize but never fabricate; no numeric prophecy; no verdicts |
player_profile | profile.serialize_for_prompt() — relevance-only, never affirmation |
school_facts | school_tier.tier_phrase(lookup(school)) — 985/211/双一流 tier |
「泛化」的来源 · Where today's genericness comes from
中文：那 4 句
MUST_HIT_*/upside/cost是每个 archetype 手写的（data_seed.json里 5 个）。管线里没有任何专业/职业级知识——详见整改计划 §1。EN: The four
MUST_HIT_*/upside/coststrings are hand-authored per archetype (5 archetypes indata_seed.json). There is no per-major/per-career knowledge — see the integration plan §1.
5. 结局（compare.py）· The ending (compare.py)
中文：
cross_life_signal—— 确定性：某个价值在两条命里都被往同一方向推，即为揭示出的内核（它在两个世界都活了下来）。backward_reflection—— 对两条 dim-trace 做 LLM 镜像：从「活出来的选择」倒推当下的抉择。是镜子，不是推荐器。path_costs—— 把每条路手写的代价摆出来，供代价接受闸门用。
EN:
cross_life_signal— deterministic: a value pushed the same way in both lives is a revealed core (it survived both worlds).backward_reflection— an LLM mirror over both dim-traces: reasons the present choice backward from what the lived choices revealed. A mirror, not a recommender.path_costs— surfaces each path’s authored cost for the acceptance gate.
6. 模型与延迟 · Models & latency
中文：
EN:
| role | model | note |
|---|---|---|
| trajectory generation | DeepSeek-V3 (deepseek/deepseek-chat) | ~19s/stage, ~2.5–3× faster than V4-Pro |
| backward mirror | V4-Pro | richer reasoning; max_tokens=8000 to avoid reasoning-truncation |
中文：max_tokens=8000（从 2000 上调——之前 V4-Pro 的推理吃光预算、把第一回合 JSON 截断）。provider 对单 key 的并发限流——并行跑两条命的 wall-clock 收益很小（266s vs 串行 ~300s）。
EN: max_tokens=8000 (raised from 2000 after V4-Pro’s reasoning ate the budget and truncated turn-1 JSON). Provider throttles concurrent calls on a single key — parallelizing two lives gave little wall-clock gain (266s vs ~300s serial).
7. 状态持久化 · State persistence
中文：交互 session pickle 到磁盘（/data/reflection-game/gaokao-data/ig_sessions），内存未命中时重新载入，所以服务器重启不会让玩家中途掉档。pickle 前丢掉 generator 句柄，载入时重新挂上。
EN: Interactive sessions are pickled to disk (/data/reflection-game/gaokao-data/ig_sessions) and reloaded on a memory miss, so a server restart doesn’t drop a player mid-game. The generator handle is dropped before pickling and re-attached on load.