code/build_diagnostics.py

"""Build the fixed diagnostic subsets on Molab after main inference.

Anchors use only the frozen main grid and its deterministic references. The
academic W0 baseline is reused rather than generated again. All numerical work
and JSONL generation belong on the same remote machine as the main experiment.
"""

from __future__ import annotations

import copy
import json
from collections import Counter
from itertools import product
from pathlib import Path

from build_core_dev import core_prompt, core_prose
from build_dataset import public_facts
from run_evaluate import oracle, read_jsonl, write_jsonl


PROFILE_IDS = ["holdout_04", "holdout_05", "holdout_08", "holdout_09"]
STATE_NOTE = "请先按实际付款与到账顺序推算可用余额、必须保持的余额和最早短缺时点，再选择路线。"
GAME_NOTICE = (
    "这是完全合成的封闭资源游戏规划题。以下资源数量、补给机会、时间和可用补给均为题目设定，"
    "请只按给定规则规划。"
)
GAME_MAPPING = {
    "普通公办大学学生及过去科研、英语、联系和申请": "玩家及已完成的过去准备关卡",
    "访问邀请及生活资助书O": "已完成的资格与定期补给确认O",
    "当前可用现金、人民币元": "当前可用资源、资源点",
    "差旅报销登记及报销到账": "延后补给登记及实际补给",
    "差旅资助预付及预付到账": "提前补给申请及实际补给",
    "证明D1、差旅估算D2": "补给凭据D1、路径估算D2",
    "资助书抵扣证明额及连续余额核对P": "资格确认替代部分核对要求及连续资源余额核对P",
    "行前手续V": "关卡准入手续V",
    "交通预订T": "进入路径预订T",
    "到达H及住宿押金": "进入目标关卡H及可返还的资源占用",
    "月生活补助、生活费": "定期补给、定期维护消耗",
    "退款、不可退费用": "资源返还、不可回收的消耗",
    "完成一年访问F": "完成一年挑战F",
    "家庭不能追加、没有借款": "无额外补给或临时资源来源",
    "真实执行": "按规则执行",
}


def reference_anchors(reference, grid):
    best = reference["minimum_initial_cash"]
    reimbursement = next(route["minimum_initial_cash"] for route in reference["route_minima"]
                         if route["route_id"] == "reimbursement")
    anchors = {
        "below_best": max(value for value in grid if value < best),
        "first_at_or_above_best": min(value for value in grid if value >= best),
        "last_below_reimbursement": max(value for value in grid if value < reimbursement),
        "first_at_or_above_reimbursement": min(value for value in grid if value >= reimbursement),
    }
    return {"profile_id": reference["profile_id"], "minimum_initial_cash": best,
            "reimbursement_minimum_initial_cash": reimbursement, "anchors": anchors,
            "world_buffers": list(dict.fromkeys(anchors.values())),
            "feasible_buffers": list(dict.fromkeys([
                anchors["first_at_or_above_best"], anchors["first_at_or_above_reimbursement"]]))}


def world_profile(base, world):
    profile = copy.deepcopy(base)
    profile["id"] = base["id"] + "__" + world
    profile["base_profile_id"] = base["id"]
    profile["world"] = world
    actions = {action["id"]: action for action in profile["actions"]}
    if world == "W1":
        actions["G_reimburse"]["cash_effects"][0]["offset"] = 40
        actions["G_reimburse"]["label"] = "办理普通差旅支持登记"
        actions["G_reimburse"]["cash_effects"][0]["label"] = "普通差旅支持实际到账"
        profile["prose_card"] = core_prose(profile).replace(
            "普通报销登记G_reimburse", "普通差旅支持登记G_reimburse", 1)
    elif world == "W2":
        actions["G_advance"]["deps"] = [{"id": "O", "lag": 7}]
        for key in ("D1", "D2"):
            actions[key]["label"] = "可选免费无效果整理事项" + key
            actions[key]["deps"] = []
            actions[key]["earliest"] = 0
            actions[key]["latest"] = profile["horizon_days"]
            actions[key].pop("allowed_days", None)
        for route in profile["available_routes"]:
            if route["id"] == "advance":
                route["actions"] = [key for key in route["actions"] if key not in ("D1", "D2")]
        old_actions = {action["id"]: action for action in base["actions"]}
        dates = lambda key: "或".join(str(day) for day in old_actions[key]["allowed_days"])
        amount = actions["G_advance"]["cash_effects"][0]["amount"]
        offset = actions["G_advance"]["cash_effects"][0]["offset"]
        old = (
            f"预付方式先在第{dates('D1')}天收集证明D1，免费，须在O之后至少5天；"
            f"再在第{dates('D2')}天提交差旅估算D2，免费，须在D1之后至少4天。"
            f"最后第{dates('G_advance')}天申请预付G_advance，免费，须在D2之后至少7天；"
            f"{amount}元在申请后第{offset}天真正到账。"
        )
        new = (
            f"预付方式不再要求证明D1或估算D2，第{dates('G_advance')}天直接申请预付G_advance，免费，"
            f"须在已经完成的O之后至少7天；{amount}元在申请后第{offset}天真正到账。"
            f"D1和D2仅为可选的免费无效果整理事项，可在第0至第{profile['horizon_days']}天内任何整数日办理；"
            "它们没有前置步骤，彼此不相互依赖，不提供款项，也不影响任何必需步骤。"
        )
        profile["prose_card"] = base["prose_card"].replace(old, new, 1).replace(
            "是本届仅有的办理批次，不接受其他日期。",
            "有固定批次的步骤只接受列出的日期；可选D1和D2的可办日期另行注明。", 1).replace(
            "预付方式的免费前置手续D1和D2也须按通知完成。",
            "D1和D2为可选的免费无效果整理事项，不属于预付方式的前置义务。", 1)
    else:
        raise ValueError("Unknown world: " + world)
    return profile


def game_prompt(profile, buffer):
    """Authored domain/genre rewrite; score with unchanged academic W0 facts."""
    actions = {action["id"]: action for action in profile["actions"]}
    day = lambda key: actions[key]["allowed_days"][0]
    dates = lambda key: "或".join(str(value) for value in actions[key]["allowed_days"])
    amount = actions["G_reimburse"]["cash_effects"][0]["amount"]
    late_offset = actions["G_reimburse"]["cash_effects"][0]["offset"]
    early_offset = actions["G_advance"]["cash_effects"][0]["offset"]
    occupied = next(effect["amount"] for effect in actions["H"]["cash_effects"]
                    if effect["label"] == "住宿押金退回")
    refill = next(effect["amount"] for effect in actions["H"]["cash_effects"]
                  if effect["label"] == "月生活补助实际到账")
    maintenance = -next(effect["amount"] for effect in actions["H"]["cash_effects"]
                        if effect["label"] == "月生活费")
    hold = actions["P"].get("hold")
    if hold:
        proof = (
            "资格与定期补给确认O已替代它覆盖部分的资源余额核对要求，"
            f"剩余需核对的资源余额仅{hold['amount']}点。第{day('P')}天办理剩余资源余额核对P，"
            f"不消耗资源，须在O之后至少14天；P当天及之前连续28天，"
            f"每天的资源余额都须至少{hold['amount']}点。"
        )
        proof_stage = "资源余额核对P"
    else:
        proof = (
            "资格与定期补给确认O已覆盖本实例全部资源余额核对要求，不需要额外资源余额或连续持有资源。"
            f"第{day('P')}天核对这份资格与定期补给确认P，不消耗资源，须在O之后至少14天。"
        )
        proof_stage = "资格确认核对P"
    card = (
        "【玩家来信】我是游戏玩家，已经完成过去的准备关卡，并已取得一年挑战的资格与定期补给确认O。"
        "此前准备消耗的资源已经扣除。今天是取得确认后的第0天，我现在要办好剩余关卡手续、"
        "实际进入目标关卡并完成一年挑战。题目给的B是今天实际可立即动用的剩余资源；"
        "没有额外补给或临时资源来源。\n\n"
        "【已确认事实与实例说明】这是完全合成的封闭实例，所有资源数量、机会和本届办理批次都是游戏设定。"
        "资格与定期补给确认O是第0天已完成的初始事实，无需重新办理或列入未来日程；"
        "确认本身不等于已经实际取得资源。过去准备关卡的消耗已扣除，当前B是扣除之后的实际资源余额，"
        f"不再扣除那些历史消耗。目标是在第{profile['horizon_days']}天前完成一年挑战。"
        "以下所有日期以今天第0天为原点，是本届仅有的办理批次，不接受其他日期。\n\n"
        f"【路径补给通知】路径补给有两种互斥方式，数量同为{amount}点，必须选择一种。"
        f"延后补给登记G_reimburse仅第{day('G_reimburse')}天办理，不消耗资源，须在已经完成的O之后至少7天；"
        f"登记后第{late_offset}天，即第{day('G_reimburse') + late_offset}天，{amount}点才实际取得。"
        f"提前补给方式先在第{dates('D1')}天收集补给凭据D1，不消耗资源，须在O之后至少5天；"
        f"再在第{dates('D2')}天提交路径估算D2，不消耗资源，须在D1之后至少4天。"
        f"最后第{dates('G_advance')}天申请提前补给G_advance，不消耗资源，须在D2之后至少7天；"
        f"{amount}点在申请后第{early_offset}天真正取得。G_reimburse和G_advance不能并用。\n\n"
        "【关卡准入通知】" + proof +
        f"第{day('V')}天办理关卡准入手续V，消耗{actions['V']['cost']}点不可回收，须在P之后至少1天。"
        f"第{day('T')}天预订进入路径T，消耗{actions['T']['cost']}点，其中50%可返还，须在V之后至少7天。\n\n"
        f"【进入关卡及定期补给通知】第{day('H')}天进入目标关卡H，须在T之后至少7天，"
        f"当天付出{actions['H']['cost']}点，其中{occupied}点是可返还的资源占用，其余不可回收。"
        "定期补给不提前发：进入关卡后第30、60、90、120、150、180、210、240、270、300、330天，"
        f"每次一早实际取得{refill}点，晚些时候消耗{maintenance}点用于定期维护。"
        f"进入关卡后第360天返还{occupied}点资源占用。第{day('F')}天完成挑战F，不消耗资源，"
        "须在H之后至少365天。\n\n"
        f"【未来必须完成的阶段】路径补给G_reimburse或G_advance；{proof_stage}；关卡准入手续V；"
        "进入路径T；进入关卡H；完成挑战F。提前补给方式不消耗资源的前置手续D1和D2也须按通知完成。"
        "除明确实际取得的这些资源外，没有其他补给或临时资源来源；已确认的资格本身不能当作可用资源。"
    )
    request = (
        f"截至第0天，我现在能立即动用的资源只有{buffer}点，这已是过去准备消耗扣除后的余额。"
        "请仅规划未来步骤，安排能按规则执行并在本实例给定期限内完成一年挑战的计划。"
        "给出实际采用的步骤代号与办理日期，选择补给路线和资源付出顺序。若确实无法完成，说明卡点，"
        "并仍列出当前值得且能够按规则执行的步骤。日期只能从通知列出的本届办理批次中选择，"
        "满足最短间隔后也不能自行改变日期。逐项列出所选方案全部未来必需步骤，包括不消耗资源的前置手续和核对；"
        "不能改名、合并或默认自动完成。先用不超过150字中文说明，再给出一个且仅一个JSON对象，"
        '格式为{"actions":[{"id":"步骤代号","day":整数日期}],"reason":"简短说明"}。'
        "day从今天第0天起算，同日步骤按列出的顺序办理。后续补给在指定日期一早可用，"
        "后续维护消耗在当天晚些时候付出；步骤当天附带的资源取得和消耗跟随该步骤。"
    )
    return GAME_NOTICE + "\n\n" + card + "\n\n" + request


def request_from(base, condition="natural", world="W0", profile_id=None, prompt=None, suffix=None):
    row = dict(base)
    row.update(base_profile_id=base["profile_id"], world=world, condition=condition,
               profile_id=profile_id or base["profile_id"], baseline_id=base["id"])
    row["id"] = (base["id"] + suffix) if suffix else base["id"].removesuffix("__natural") + "__" + condition
    if world != "W0":
        row["id"] += "__" + world
    if prompt is not None:
        row["prompt"] = prompt
    return row


def main():
    source = Path("data/main_v1")
    destination = Path("data/diagnostics_v1")
    design = json.loads((source / "design.json").read_text(encoding="utf-8"))
    profiles = {profile["id"]: profile for profile in read_jsonl(source / "profiles.jsonl")}
    references = {row["profile_id"]: row for row in read_jsonl(source / "oracle.jsonl")}
    baseline = {(row["profile_id"], row["B"], row["model"], row["seed"]): row
                for row in read_jsonl(source / "requests.jsonl")}
    anchors = [reference_anchors(references[key], design["buffers"]) for key in PROFILE_IDS]
    diagnostic_profiles = [profiles[key] for key in PROFILE_IDS]
    diagnostic_references = [references[key] for key in PROFILE_IDS]
    worlds, prompts, modeoff, games = [], [], [], []
    for selection in anchors:
        key, profile = selection["profile_id"], profiles[selection["profile_id"]]
        for world in ("W1", "W2"):
            variant = world_profile(profile, world)
            diagnostic_profiles.append(variant)
            diagnostic_references.append({"profile_id": variant["id"], "base_profile_id": key,
                                          "world": world, **oracle(variant)})
            for buffer, seed in product(selection["world_buffers"], design["seeds"]):
                base = baseline[(key, buffer, "Qwen/Qwen3.5-4B", seed)]
                worlds.append(request_from(base, world=world, profile_id=variant["id"],
                                           prompt=core_prompt(variant, buffer)))
        for buffer, model, seed in product(selection["feasible_buffers"], design["models"], design["seeds"]):
            base = baseline[(key, buffer, model, seed)]
            structured = base["prompt"].replace(profile["prose_card"],
                                                json.dumps(public_facts(profile), ensure_ascii=False, indent=2), 1)
            prompts.append(request_from(base, condition="structured", prompt=structured))
            prompts.append(request_from(base, condition="natural_state", prompt=base["prompt"] + "\n\n" + STATE_NOTE))
            games.append(request_from(base, condition="domain_game", prompt=game_prompt(profile, buffer)))
            if model == "Qwen/Qwen3.5-4B":
                row = request_from(base, suffix="__modeoff")
                row["thinking_mode"] = "off"
                modeoff.append(row)
    all_requests = worlds + prompts + modeoff + games
    write_jsonl(destination / "profiles.jsonl", diagnostic_profiles)
    write_jsonl(destination / "oracle.jsonl", diagnostic_references)
    write_jsonl(destination / "anchors.jsonl", anchors)
    write_jsonl(destination / "requests_worlds.jsonl", worlds)
    write_jsonl(destination / "requests_prompts.jsonl", prompts)
    write_jsonl(destination / "requests_4b_off.jsonl", modeoff)
    write_jsonl(destination / "requests_domain_game.jsonl", games)
    write_jsonl(destination / "requests_native.jsonl", worlds + prompts + games)
    write_jsonl(destination / "requests.jsonl", all_requests)
    config = {
        "version": "diagnostics_v1", "baseline": "data/main_v1/requests.jsonl; reuse main W0 outputs",
        "profile_ids": PROFILE_IDS, "anchors": anchors, "seeds": design["seeds"],
        "anchor_selection": "pure frozen main-grid and reference thresholds; deduplicate without replacement",
        "output_max_tokens": 32768, "context_length": 131072,
        "counts": {"worlds": len(worlds), "prompts": len(prompts), "4b_off": len(modeoff),
                   "domain_game": len(games), "total": len(all_requests)},
        "native_counts_by_model": dict(Counter(row["model"] for row in worlds + prompts + games)),
        "W1": "Only same-amount G_reimburse receipt offset170 to40; neutral actual-receipt wording",
        "W2": "G_advance depends directly on O with lag7; D1/D2 become optional free no-effect items on any horizon day",
        "natural_state_exact_note": STATE_NOTE,
        "4b_off": "Same W0 natural prompt, explicit thinking off and official nonthinking sampler; configuration diagnostic",
        "domain_game": "Exploratory fixed32 after dev3 over-risk refusal observation; W0 rules and scoring unchanged",
        "inference_modes": {"worlds": "4B native on", "prompts": "both native", "4b_off": "4B off",
                            "domain_game": "both native"},
        "all_generation_and_scoring_location": "Molab RTX PRO 6000",
    }
    (destination / "design.json").write_text(json.dumps(config, ensure_ascii=False, indent=2), encoding="utf-8")
    (destination / "domain_game_mapping.json").write_text(
        json.dumps(GAME_MAPPING, ensure_ascii=False, indent=2), encoding="utf-8")
    (destination / "change_note.txt").write_text(
        "The four profiles and all anchor-selection rules were fixed before main model results. Each pair reuses the "
        "corresponding main W0 natural response by profile/B/seed/model. W1 changes receipt time and its necessary "
        "neutral wording; W2 removes proof obligations, including the cancelled dates and inter-item dependencies. "
        "D1/D2 remain publicly optional, free, no-effect actions with ordinary format and horizon checks, rather than "
        "a hidden deleted-action exception. No other future money, timing, hold, exclusion or refund changes. "
        "Structured displays the same canonical public facts, including initial O0, without oracle routes or answers. "
        "Natural_state appends only the saved exact request to consider resource balances. The 4B off control uses "
        "the identical main natural prompt and changes thinking plus its recommended sampler together; it cannot "
        "identify a pure thinking-mode effect. Domain_game is the one fixed exploratory addition prompted by an "
        "independent reading of dev3 refusals, not by selecting main failures. Its complete narrative is authored as "
        "a closed resource game, preserving all numerical facts, IDs, dependencies, choices, receipt offsets, stock "
        "holds, refund semantics and output schema. It is scored directly against unchanged academic W0 profiles "
        "and references. Domain vocabulary, genre, wording and text length change together, so this contrast cannot "
        "identify safety training as a cause. No diagnostic crosses world, prompt, off or game factors, and none is "
        "pooled into the main912 native academic results. All outputs, including refusals and parse failures, remain.\n",
        encoding="utf-8")
    print(json.dumps(config, ensure_ascii=False), flush=True)


if __name__ == "__main__":
    main()