订阅的设计就是让你忘记它。单笔金额小到不触发任何感觉，扣款日藏在月中，涨价时那封邮件混在促销里——**唯一能看见全貌的地方是账单流水**。 贴进来，代码做四件事：按「同一商户 + 金额相近 + 间隔稳定」判出哪些是周期性扣款，用实测间隔的中位数定周期（月付／季付／年付），把每一项换算成**年化金额**并排序，再推出下一次扣款日还有几天。同时比对首末金额，标出**涨过价**的那几项以及一年因此多付多少。 这些数字一个都不经过模型。让模型「看一眼账单估个年费」，它会算错，而且错得很自信。模型只负责它真正擅长的部分：怎么判断一项订阅的退订难度、退订前该先确认什么、以及打电话时怎么说。 还有两类会被单独列出来但**不计入年化**：金额或间隔不稳的（按量计费，往往比固定订阅更贵），和只出现一次的（年费订阅在 12 个月流水里也只出现一次）。$2 解锁逐项退订路径判断、退款窗口的判断方法、以及取消邮件与致电话术。Subscriptions are designed to be forgotten: amounts small enough not to register, charge dates mid-month, price-rise emails buried in promotions. **The statement is the only place the whole picture exists.** Paste it and the code does four things: identifies recurring charges by same merchant, similar amount and stable interval; determines the cycle from the median measured gap; converts each to an **annualised figure** and ranks them; and projects the next charge date. It also compares first and last amounts to flag **price rises** and what each costs per year. None of those numbers pass through the model. Ask a model to eyeball a statement and estimate an annual total and it will be wrong, confidently. The model handles what it's actually good at: judging how hard something is to cancel, what to confirm first, and what to say on the phone. Two groups are listed but **excluded from the total**: irregular charges (metered services, often costlier than fixed subscriptions) and single occurrences (an annual subscription appears once in twelve months too). $2 unlocks per-item cancellation paths, how to judge refund windows, and cancellation scripts.
逐项退订难度分级（App 内可退／须致电／须邮件／须走网页版）与该先确认什么、退款窗口的判断方法（不给你会变的天数，给你去哪查）、取消邮件与致电话术、以及退订后的验证清单——很多「已取消」并没有真的取消。Per-item difficulty (in-app / phone / email / desktop-web only) and what to confirm first, how to judge refund windows (where to look, not a number that changes), cancellation emails and phone scripts, and a post-cancellation verification list — plenty of "cancelled" subscriptions aren't.
本工具只对你粘贴的文本做模式识别，不连接任何银行、不读取账户、也不代你操作退订。识别依据是「同一商户 + 金额相近 + 间隔稳定」，因此**会漏也会错**：商户名被聚合成通道名（财付通／PAYPAL）时无法区分真正收款方，只出现一次的年费订阅判不出来，同名商户的普通消费可能被误判为订阅——请以你在各服务方账户内看到的订阅状态为准。下一次扣款日是按实测间隔推算的估计值，不是官方日期。退订、退款与合同条款以各服务方的条款为准。你粘贴的内容不入库。若账单含卡号，请先删掉再贴——本工具不需要卡号。This does pattern matching on text you paste. It connects to no bank, reads no account and cancels nothing for you. Detection relies on same merchant, similar amount and stable interval, so it **both misses and mis-flags**: aggregator names (PayPal, WeChat Pay) hide the real payee, an annual subscription appearing once can't be classified, and ordinary purchases from a recurring merchant may look like a subscription. Treat each service's own account page as authoritative. The next-charge date is an estimate from measured intervals, not an official date. Cancellation and refund terms are set by each provider. Nothing you paste is stored. Remove card numbers before pasting — none are needed.