从空白页面点“用合成样例体验”，会直接载入并分析虚构资料。已有工作先备份，不要直接同意替换。样例按钮不会重置你改过的支出方向或阈值；按负数支出、500 阈值核对后，必要时重新检查。

摘要应标明“合成样例 — 不是真实账户检查”：7 笔总交易、3 笔历史、4 条待复核，检查期为 2026-10-03 至 2026-10-09。

2026-10-08 · Electric · −250

偏离历史金额。历史电费为 60、65、62，中位数 62；250 严格大于 3 倍的 186，需要核对变化原因，不代表盗扣。

2026-10-08 · Unknown recipient · −800

历史未出现＋达到大额阈值。描述没有出现在这份文件的更早历史中，且 800 达到 500；工具没有确认收款人身份。

2026-10-09 · Subscription · −15（第一条）

历史未出现＋同日同额重复。日期、描述、金额相同的记录有两条，可能是两笔购买、导入重复或重复扣款，需人工核实。

2026-10-09 · Subscription · −15（第二条）

同样标记历史未出现＋同日同额重复。两条分别列出，因此计为两条提示。

这是 4 条记录、共 7 个原因标签，不是 4 笔诈骗。商户描述变化也可能触发提示。演示备注可写“合成演示，尚未联系银行”；不要把模拟勾选说成已执行银行设置。演示恢复 JSON 也需要针对该虚构资料确认权限，不代表取得任何真实账户授权。

• CSV 使用 UTF-8、逗号分隔、有标题行；至少 3 列，每行列数一致。最多 2 MiB（2 × 1024 × 1024 字节）、1–10,000 笔交易。自动匹配表头只帮助初选，不能代替人工核对。一次一个账户、一种币种；工具不换汇，也不能仅凭表头保证账户或币种一致。
• CSV 日期使用有效的 ISO YYYY-MM-DD，或明确选择月／日／年（MDY）、日／月／年（DMY）。例如 10/09/2026 在两种设置下含义不同，必须按原件核对。CSV 两位年份必须明确选择 1900 或 2000 世纪；工具不猜世纪。不支持时间戳或表格日期序列。
• 描述不可为空，最多 500 字符。金额可选一个统一正负列，并确认支出方向；也可映射两个不同的支出 Debit／收入 Credit 列。分栏金额必须是非负数，每笔只能一侧为非零；不要把两栏都非零的行当作净额自动相减。
• 金额使用小数点、最多两位小数；可识别受支持的币种符号、括号负数、符号与数字之间的空格，以及末尾负号，例如 - 35.00、(35.00)、35.00-。这些格式不代表币种已核实。千位逗号必须完整分组，CSV 单元格需加引号，如 "-1,234.56"；不支持小数逗号。金额绝对值最高 1,000,000,000,000；分栏金额仍须遵守非负与单侧非零规则。
• 阈值大于零、最多两位小数、最高 1,000,000,000,000；支出达到或超过阈值就标记。零及相反方向金额不产生支出提示。

“最新七天”以本次选定交易的最大日期为终点，向前 6 天，含首尾；不是电脑的今天。选定交易中的更早记录才是历史，旧备份、其他文件或未选中的 PDF 交易表不会自动补入。错误未来日期会改变检查期。

没有更早历史仍能运行，但只检查重复与大额，不做“历史未出现”和“偏离历史金额”比较。偏离比较需要同一描述至少 3 笔历史支出；月付账单通常需此前至少 3 次月付记录。描述只统一大小写和空白，不靠 AI 识别身份；重复依据同日、同描述、同金额。提示总数不是尚未解决的数量。

文字型 PDF 的支持边界

1. 可尝试导入带可提取文字的银行 PDF，最多 10 MiB（10 × 1024 × 1024 字节）、30 页。选择文件前，数字日期有月／日歧义时需明确设定 MDY 或 DMY；只有一个 Amount 金额列时始终必须明确选择支出正负方向，Debit／Credit 分栏则可不选。更改这些设置后需重新选择文件。PDF 两位年份或无年份日期须有无歧义的完整对账年份依据，否则拒绝导入。工具只尝试识别交易表；不保证支持所有银行、所有 PDF 格式或完整识别整份对账单。
2. 检测到多个账户或交易表时，只选择一个需要检查的账户／表。阅读缺页或不支持内容等警告，并逐页查看全部选中交易的预览。PDF 不提供逐行排除清单；CSV 才列出被排除的余额／汇总行。能提取部分交易不代表原件完整；无法确认日期、方向、金额或完整性时，不要确认分析。
3. 扫描件、加密 PDF、文字含义或表格结构无法可靠判断的 PDF 会被拒绝，并请你改用银行 CSV；本工具不做 OCR。可向银行索取 CSV，或在本机表格软件中人工整理有权查看的数据，逐行核对原件后导出 UTF-8 CSV。不要仅改文件后缀，也不要把私人流水上传云端转换。

权限勾选不会授予法律权限、核验代理身份或向银行登记关系。可信联系人不等于资金操作权或账户访问权。只如实记录已经完成的核实。

流水和跟进记录在当前页面本地处理，不上传交易、不调用 AI。打开网站仍需加载网站资源；这不等于完全离线，也不能保护已被入侵的设备或浏览器扩展。使用私人受保护设备，不输入银行密码、验证码或完整账号。

工具不连接银行、不读取邮箱、不持续监控；不转账、不冻结或拦截资金、不自动联系银行、不保证追回。它不能判断交易由谁操作，也不能确定授权真假、行为能力、诈骗或法律责任。账户变更和代理权限等信息需本人或合法代理人向银行另行核实。

有疑问时，通过银行官网或银行卡上的官方电话联系合适部门，记录事实。提示只是人工复核线索，零提示不是安全证明。

JSON 备份未加密。它保存被标记交易及原因、导入／复核警告、人工状态/备注、保护措施、银行日志、下次日期和检查摘要；不含原始 CSV／PDF、未标记交易的完整明细、查看权限勾选或列映射。银行原文件与备份需分别安全保管，不公开分享或放入无人管理的共享目录。

页面不自动持久保存。刷新、关闭或设备故障可能丢失未备份工作；浏览器也可能保留会话，因此关闭页面不是安全删除。下载目录、同步软件、设备备份和打印服务可能另留副本。

恢复时先确认有权查看，再选本工具生成的 JSON（最多 2 MiB，不支持任意 JSON）。恢复成功保留跟进记录和备份中的复核警告，但清空银行输入、列映射及导入预览，页面沿用当前语言。恢复失败会保留旧结果；旧备份中缺失的检查元数据会显示未知，不代表重新检查过。

先备份，再决定是否替换。导入、预览或验证失败不会清除旧检查与备注；旧结果不会冒充本次新输入结果。新输入完成有效分析后，替换已有工作仍须确认；取消会保留旧检查。确认替换将清除旧疑点状态、备注、保护措施、银行日志和下次日期。恢复 JSON 后重新导入同一文件也属于新分析，当前不自动合并恢复记录或跨周记录。

打印报告包含摘要、规则/边界、全部已标记交易及其状态/完整备注、保护措施、银行日志和下次日期；不是完整银行对账单，不附原 CSV／PDF 或全部未标记交易。按钮打开原生打印窗口，由你选择打印机或保存 PDF；先检查页数、长备注末尾和敏感信息。保存 PDF 不等于实体打印或向银行提交，窗口行为依浏览器和系统而异。

提示先确认查看权限

仅在确有同意或法律权限时勾选，不要为消除报错随意勾选。纯体验可用内置合成样例。

文件乱码、读不了或行列不一致

CSV 检查 UTF-8、2 MiB／10,000 笔上限、表头、列数和逗号字段的引号。PDF 检查 10 MiB／30 页上限；扫描、加密或无法可靠识别的文件请改用银行 CSV。改后缀不会转换格式。

日期或金额报错

CSV 确认 ISO、MDY 或 DMY，去掉时间；两位年份选择正确世纪。PDF 有歧义数字日期须在选文件前设定 MDY／DMY；单 Amount 列始终须明确支出方向，分栏则无需。PDF 两位年份或无年份日期缺少明确的完整对账年份依据时，改用银行 CSV。修改 PDF 设置后重新选择文件。金额用小数点、最多两位小数和正确千位分组，不用小数逗号。Debit／Credit 分栏每行只能一侧非零，且不可为负。按原件核对，不为通过导入而猜日期或净额。

没有历史比较，或检查期不对

本次选定交易中应包含更早记录；同描述偏离比较至少需 3 笔历史支出。其他 PDF 账户／表不会自动合并。检查最新日期是否过期或有错误未来日期；这不是实时监控。

改设置后还是旧结果，或提示太多

CSV 修改列映射或解析设置后重新生成预览；只改阈值则沿用已确认的预览。PDF 修改日期顺序或支出方向后，必须重新选择文件。逐页核对并确认新预览，再分析。核对历史覆盖、支出方向和阈值，并阅读替换确认；不要为消除提示删除真实记录。新检查完成前，下方可能仍是保留的旧结果，请阅读页面说明。

恢复后不能重新分析，或怕备注丢失

JSON 不含原 CSV／PDF，需重新导入银行文件，核对映射或 PDF 表，再预览并确认。先下载备份，再开始新检查或同意替换；当前不自动合并旧工作。

没有收到提醒，或没有下载 PDF

下次日期仅是记录，请自行设置提醒。打印窗口需自行选择保存 PDF；取消打印不会生成报告文件。

备份超过 2 MB

工具会阻止生成无法恢复的大备份。先保存完整报告，再决定是否减少备注或分批检查，不要直接丢弃记录。

Start on a blank page and click “Try synthetic example” to load and analyze fictional data. Back up existing work before considering replacement. The example does not reset a spending direction or threshold you changed; check negative spending and 500, then reanalyze if needed.

The summary should say “Synthetic example — not a real account review”: 7 total transactions, 3 historical, and 4 flags, covering October 3–9, 2026.

2026-10-08 · Electric · −250

Higher than historical pattern. Earlier amounts are 60, 65, and 62; their median is 62. Spending of 250 is strictly greater than three times that median (186). Check the change; this does not establish an unauthorized charge.

2026-10-08 · Unknown recipient · −800

Not in earlier history + reaches the large-spending threshold. This description is absent from the file’s earlier history, and 800 reaches 500. The tool has not identified who received the money.

2026-10-09 · Subscription · −15 (first entry)

Not in earlier history + same-day, same-amount duplicate. Two entries share the date, description, and amount. They could reflect two purchases, duplicate import data, or a duplicate charge; a person must check.

2026-10-09 · Subscription · −15 (second entry)

Also flagged as not in earlier history + same-day, same-amount duplicate. Both entries are listed separately, so they count as two flags.

That is 4 records and 7 reason labels, not 4 fraudulent transactions. Changed merchant descriptions can also trigger flags. For a demo note, use “Synthetic demo; bank not contacted.” Do not present simulated checkboxes as completed bank settings. Restoring a demo JSON also requires confirming access to that fictional data; it does not authorize access to a real account.

• Use a UTF-8, comma-separated CSV with a header, at least 3 columns, and consistent column counts. Allow 1–10,000 transactions, up to 2 MiB (2 × 1024 × 1024 bytes). Automatic header matching is only a starting suggestion; verify it yourself. Use one account and one currency. The tool does not convert currencies, and headers alone cannot establish that an account or currency is consistent.
• For CSV, use valid ISO YYYY-MM-DD dates, or explicitly select month/day/year (MDY) or day/month/year (DMY). For example, 10/09/2026 means different dates under the two settings; check the original. CSV two-digit years require an explicit 1900 or 2000 century; the tool does not guess it. Timestamps and spreadsheet serial dates are unsupported.
• Descriptions must be nonempty and at most 500 characters. Select one signed amount column and confirm its spending direction, or map two distinct Debit/spending and Credit/income columns. Separate-column amounts must be nonnegative, with only one nonzero side per transaction. Do not automatically subtract two nonzero columns into a net amount.
• Use a decimal point and at most two decimal places. Supported currency symbols, parentheses for negatives, spaces between a sign and digits, and trailing minus signs can be recognized, such as - 35.00, (35.00), and 35.00-. Recognizing a symbol does not verify the currency. Thousands commas must be correctly grouped and the CSV cell quoted, such as "-1,234.56". Decimal commas are unsupported. The maximum absolute amount is 1,000,000,000,000. Separate Debit/Credit columns still require nonnegative values and only one nonzero side.
• The threshold must be greater than zero, with at most two decimal places, up to 1,000,000,000,000. Spending at or above it is flagged. Zero and amounts in the opposite direction do not generate spending flags.

The latest seven days end at the latest date in the selected transactions, starting six days earlier, inclusive—not today on your computer. Only earlier selected records count as history. Backups, other files, and unselected PDF tables do not supply it. An incorrect future date changes the review period.

Without earlier history, only duplicate and large-spending checks run; new-description and historical-pattern comparisons are skipped. Historical-pattern comparison needs at least 3 earlier spending entries with the same description. A monthly bill generally needs at least 3 earlier monthly payments. Description matching only normalizes case and whitespace, not identity using AI. Duplicates match date, description, and amount. The flag count is not an unresolved-item count.

Limits of text-based PDF support

1. You may try a bank PDF with extractable text, up to 10 MiB (10 × 1024 × 1024 bytes) and 30 pages. Before selecting the file, explicitly choose MDY or DMY for ambiguous numeric dates. Every single-Amount-column PDF requires an explicit spending-sign convention; separate Debit/Credit PDFs may omit it. After changing these settings, select the file again. PDF two-digit or yearless dates require unambiguous full statement-year evidence; otherwise import is rejected. The tool attempts to identify transaction tables; it does not promise support for every bank or PDF format, or complete extraction of an entire statement.
2. If several accounts or tables are detected, select only the account/table you intend to review. Read warnings such as missing pages or unsupported content, and review every page of the selected transaction preview. PDF does not provide an excluded-line inventory; itemized excluded balance/total rows are a CSV feature. Extracting some transactions does not establish that the statement is complete. Do not confirm analysis if you cannot verify dates, directions, amounts, or completeness.
3. Scanned, encrypted, or ambiguously structured PDFs are rejected with a request to use bank CSV instead. There is no OCR. Request a CSV from the bank, or manually prepare authorized data in a local spreadsheet and verify each row against the original before exporting UTF-8 CSV. Renaming a file extension does not convert it. Do not upload private transactions to a cloud converter for this task.

The access checkbox does not grant legal authority, verify a representative, or register a relationship with the bank. A trusted contact is not permission to move money or access the account. Only record checks that actually happened.

Transactions and follow-up records are processed in the current page, without transaction uploads or AI calls. Opening the site still loads website resources; this is not a promise of fully offline use or protection against a compromised device or browser extension. Use a private, protected device, and never enter bank passwords, verification codes, or full account numbers.

The tool does not connect to banks, read email, continuously monitor accounts, transfer money, freeze or block funds, contact banks automatically, or guarantee recovery. It cannot determine who made a transaction, whether authority is valid, a person’s capacity, fraud, or legal responsibility. The account holder or authorized representative must separately verify account changes and representative permissions with the bank.

For unfamiliar activity, contact an appropriate bank department using the official website or the number on the bank card, and record facts. Flags are leads for human review; zero flags is not proof of safety.

JSON backups are not encrypted. They contain flagged transactions and reasons, import/review warnings, manual statuses/notes, protection checks, bank logs, next review date, and review metadata. They do not include raw CSV/PDF files, full details of unflagged transactions, access confirmation, or column mappings. Keep the original bank file and backup separately in safe locations; do not share them publicly or place them in unmanaged shared folders.

The page does not persistently autosave. Refreshing, closing, or device failure may lose unsaved work; browsers can also retain sessions, so closing the page is not secure deletion. Downloads, sync services, device backups, and print services may retain other copies.

To restore, confirm authorized access and select a JSON backup made by this tool (up to 2 MiB, not an arbitrary JSON file). A successful restore keeps saved follow-up and backed-up review warnings, clears bank inputs, mappings, and the import preview, and uses your current page language. A failed restore preserves previous results. Missing metadata in older backups is shown as unknown; restoration does not recheck transactions.

Back up before deciding to replace. Import, preview, or validation failures do not clear your previous review or notes, and previous results are not presented as results for new input. After new input produces a valid analysis, replacing existing work still requires confirmation; canceling preserves the previous review. Accepting clears old flag statuses, notes, protection checks, bank logs, and the next date. After JSON restoration, importing even the same file is a new analysis. Restored or cross-week records are not automatically merged.

The report contains the summary, rules/limits, all flagged entries with statuses and full notes, protection checks, bank logs, and next date. It is not a full bank statement and does not attach raw CSV/PDF files or all unflagged transactions. The button opens the native print dialog; choose a printer or Save as PDF and check pages, long-note endings, and sensitive content in the preview. Saving a PDF does not mean it was physically printed or sent to the bank. Dialog behavior varies by browser and system.

Asked to confirm authorized access

Only confirm if you genuinely have consent or legal authority; do not tick it just to dismiss an error. Use the built-in synthetic example to explore.

Unreadable text, unreadable file, or inconsistent columns

For CSV, check UTF-8, the 2 MiB / 10,000-transaction limits, headers, column counts, and quotes around comma-containing fields. For PDF, check the 10 MiB / 30-page limits. Use bank CSV for scanned, encrypted, or unreliably recognized PDFs. Renaming an extension does not convert a file.

Invalid date or amount

For CSV, confirm ISO, MDY, or DMY, remove times, and choose the correct century for two-digit years. Before selecting a PDF, set MDY/DMY for ambiguous numeric dates and always choose a spending-sign convention for a single Amount column; separate columns do not require it. Use bank CSV when PDF two-digit or yearless dates lack clear full statement-year evidence. Select the PDF again after changing its settings. Use a decimal point, at most two decimal places, and correct thousands grouping—not decimal commas. Debit/Credit rows allow only one nonzero side, with no negative values. Check the original; do not guess dates or net amounts just to pass import.

No historical comparison, or an unexpected period

Include earlier history in the selected transactions; pattern comparisons need at least 3 earlier spending entries with the same description. Other PDF accounts/tables are not merged automatically. Check for stale data or an incorrect future date. This is not live monitoring.

Changed settings still show old results, or too many flags

After CSV mapping or parsing changes, generate a new preview. A threshold-only change reuses the confirmed preview. After PDF date-order or spending-convention changes, select the file again. Verify every page of the new preview and confirm before analysis. Check history coverage, spending direction, and threshold, and read the replacement confirmation. Do not delete real transactions to remove flags. Previous results may remain visible until the new review is complete; read the page’s explanation.

Cannot reanalyze after restoring, or worried about lost notes

JSON contains no original CSV/PDF. Import the bank file again, verify mappings or the selected PDF table, then preview and confirm. Download a backup before a new review or replacement. Old work is not automatically merged.

No reminder or PDF arrives

The next date is only a record; set your own reminder. Choose Save as PDF in the print dialog; canceling it creates no report file.

Backup exceeds 2 MB

The tool blocks a backup that would be too large to restore. Save a complete report first, then decide whether to reduce notes or review smaller batches. Do not simply discard records.