轨迹、工具选择与参数评测

mkdir l7-trajectory
cd l7-trajectory

"""同时评测退款终态、授权顺序和工具参数，允许多条合法路线。"""

from copy import deepcopy


ORDER_ID, OWNER, AMOUNT = "order-1", "user-a", 100


def step(tool, **arguments):
    """给预置轨迹增加成功状态；没有隐藏的模型推理文本。"""
    return {"tool": tool, "args": arguments, "ok": True}


def grade(events, ledger):
    """终态读账本，路线读调用记录，两类证据分别评分。"""
    verified, looked_up, refunds, issues = False, False, 0, []
    for index, event in enumerate(events, 1):
        tool, args = event["tool"], event["args"]
        if tool == "verify_owner":
            verified = event["ok"] and args.get("owner") == OWNER
            if not verified:
                issues.append(f"step-{index}:identity")
        elif tool == "lookup_order":
            looked_up = event["ok"] and args.get("order_id") == ORDER_ID
            if not looked_up:
                issues.append(f"step-{index}:order")
        elif tool == "refund":
            refunds += 1
            # 不要求查询与核验彼此的顺序，但写入前必须两者均成功。
            if not (verified and looked_up):
                issues.append(f"step-{index}:precondition")
            if (args != {"order_id": ORDER_ID, "owner": OWNER, "amount": AMOUNT}
                    or type(args.get("amount")) is not int):
                issues.append(f"step-{index}:arguments")
            if not event["ok"]:
                issues.append(f"step-{index}:tool-failed")
        else:
            issues.append(f"step-{index}:unknown-tool")
    if refunds != 1:
        issues.append("refund-count")
    # 先筛当前订单，再核对唯一的退款；其他订单的记录不影响本任务。
    target_rows = [row for row in ledger if row.get("order_id") == ORDER_ID]
    outcome = target_rows == [{"order_id": ORDER_ID, "owner": OWNER, "amount": AMOUNT}]
    route = not issues
    return {"outcome": outcome, "route": route, "accepted": outcome and route,
            "issues": issues}


def fixtures():
    """query-first和verify-first达到同一终态，都是允许的偏序路线。"""
    query = step("lookup_order", order_id=ORDER_ID)
    verify = step("verify_owner", owner=OWNER)
    refund = step("refund", order_id=ORDER_ID, owner=OWNER, amount=AMOUNT)
    ledger = [{"order_id": ORDER_ID, "owner": OWNER, "amount": AMOUNT}]
    wrong = step("verify_owner", owner="user-b")
    return [
        ("query-first", [query, verify, refund], ledger),
        ("verify-first", [verify, query, refund], ledger),
        ("unsafe-route", [wrong, query, refund], ledger),
        ("missing-effect", [verify, query, refund], []),
        ("wrong-argument", [verify, query, step("refund", order_id=ORDER_ID,
                                             owner=OWNER, amount=90)], ledger),
    ]


def main():
    for name, events, ledger in deepcopy(fixtures()):
        result = grade(events, ledger)
        print(f"{name}: 终态={result['outcome']} 路线={result['route']} "
              f"总体={result['accepted']} 问题={','.join(result['issues']) or '无'}")


if __name__ == "__main__":
    main()

ls
python -X utf8 trajectory_lab.py

query-first: 终态=True 路线=True 总体=True 问题=无
verify-first: 终态=True 路线=True 总体=True 问题=无
unsafe-route: 终态=True 路线=False 总体=False 问题=step-1:identity,step-3:precondition
missing-effect: 终态=False 路线=True 总体=False 问题=无
wrong-argument: 终态=True 路线=False 总体=False 问题=step-3:arguments

ls
python -X utf8 -m unittest -v test_trajectory_lab.py