JP Number Grounding

Check that every number, amount, date and time in an AI-written Japanese summary actually appears in its source. Kanji numerals, 万/億, full-width digits, Japanese era years and 午前/午後 are normalized first. MCP server, CLI and Python library. Standard library only, no network, MIT.

Meeting minutes (議事録) and summaries are one of the most common jobs given to language models in Japan. A Japanese municipal survey counted 755 use cases for minutes summarization, second only to greeting drafts (cited here). The known failure: a model turns a proposal into a decision, or adds a deadline nobody said.

Checking the numbers by hand is slow because Japanese writes one value many ways. A script that searches the source for the digits in the summary misses most of them:

source                              summary            naive digit search   this tool
三千五百万円                         3,500万円          "3500" not found     supported
令和7年4月1日                        2025/4/1           "2025" not found     supported
午後2時                              14:00              "14" not found       supported
1割5分                               15%                "15" not found       supported
参加者12名                           12人               found                supported
問い合わせ30件                       30人               found (wrong)        unit_mismatch
次回は5月18日                        期限は5月1日        "5","1" found (wrong) unsupported
2025年4月1日(火)                     2025年4月1日(水)    found (wrong)        W_WEEKDAY_MISMATCH

On the 30 bundled cases this tool flags the right cases 30/30; the naive digit search gets 14/30. The cases and the rules come from the same author, so treat this as a regression suite, not proof.

What it normalizes

• kanji and mixed numerals: 三千五百万, 3千5百万, 1億2000万, 二〇二五, full-width １２

• money, rates and counters: 円, ドル, %, パーセント, 割/分/厘, 人 = 名, 件, 社, 回, か月 = ヶ月 = カ月, 拠点 = か所, kg, km ...

• dates: 2025年4月1日, 2025/4/1, 令和7年4月1日, 令和元年, R7.4.1, 4月1日, 4/1, 2025年4月, and the weekday in (火)

• times: 14:00, 午後2時, 10時半, 午前12時

• era years: 令和6年度 = 2024年度

• approximations: 約, およそ, 程度, 前後 ... allow a tolerance (default 10%)

• words that only look like numbers are skipped: 一宮, 三菱, 統一, 一部, 万一, 同一人物

Status per fact: supported, approx_supported, year_unverified (source gives no year), unit_unverified (one side has no unit), unit_mismatch, unsupported. ok is false when any fact is unsupported or unit_mismatch, or a weekday does not match its date.

It checks that values are present, not that they are used correctly. A summary can quote 15% from the source and attach it to the wrong item.

Related MCP server: japan-data-mcp

MCP server (for AI agents)

{"mcpServers": {"jp-number-grounding": {"command": "python", "args": ["-m", "jpnumground.server"], "cwd": "/path/to/jp-number-grounding"}}}

Both tools are read-only and offline. Suggested agent rule: after writing minutes or a summary from Japanese text, call jp_number_grounding_check and fix or remove every unsupported fact before sending.

CLI and Python

python -m jpnumground transcript.txt minutes.txt   # JSON, exit 1 when a fact is unsupported
python -m jpnumground --extract minutes.txt        # list quantities
python -m jpnumground.bench                         # score on the 30 cases

from jpnumground import check, extract
check("予算は三千五百万円", "予算は3,800万円")["problems"]
# [{'text': '3,800万円', 'normalized': '38000000円', 'status': 'unsupported', 'issues': []}]
extract("令和7年4月1日 午後2時")[0]["normalized"]   # '2025-04-01'

Tests

python -m unittest discover -s tests   # 26 tests

日本語

AIが書いた要約・議事録の数字・金額・日付・時刻が、元の文章(文字起こし・資料)にあるかを確かめます。漢数字・万/億・全角・和暦・午前/午後・割/分をそろえてから比べるので、「三千五百万円」と「3,500万円」、「令和7年4月1日」と「2025/4/1」は同じとみなし、言っていない期限(unsupported)や単位の取り違え(30件→30人)、日付と曜日の食い違いを返します。値があるかを確かめる道具で、意味の正しさは確かめません。標準ライブラリだけ・外部通信なし・MIT。

License

MIT. See LICENSE.