JHNRD

Part of Awesome HORIZON SHIELD — the checkable index of every dataset, MCP server and ledger we run.

Japan Home-visit Nursing Reimbursement Database

A record of Japanese home-visit nursing reimbursement rules — additions, reductions, and the expiry of physician instructions — in which every requirement carries the document it rests on.

The numbers themselves are public. What this dataset holds is where each number came from.

日本語版はこちら → README.ja.md

間違いを見つけた方へ。 この資料の誤りに気づけるのは、現場で算定している方です。 直し方は CONTRIBUTING.md に日本語で書いてあります。 出典を1つ添えて Issue を立ててください。 「この加算の要件が違う」の一行でも構いません。訂正は不具合として扱います。

Do not make a billing decision on this alone

8 sources currently rest on the statute text itself (statute — the ministerial notice, ordinance or circular). statute is 11 in total; 5 of those are superseded revisions and cannot be used as current authority. They are kept only so that the revision history is legible.

The rest are ministry material that is not the statute text (agency, 8) and private commentary (secondary, 7).

54 requirements are unconfirmed. Anything we could not check against the statute text says so. It is not left blank.

What this dataset can say is: "this reduction has requirements A, B and C, and C has not yet been confirmed against the statute text." It never says "you may bill this" or "this applies to you."

The reason for that line is simple: when a claim is returned, someone has to be accountable, and a dataset cannot be.

Where it stands

33 items. This is not a complete map of the rules. The name runs ahead of the contents, and the item count is put first so that this is not hidden.

Japanese statutory terms are never translated away

Names that come from a statute — 看護体制強化加算 (Ⅰ), 特別管理加算, 告示第95号 — are kept verbatim in this repository, in the data and in this README, with an English gloss beside them where one helps.

They are not replaced by an English name. No English name for them exists in any Japanese statute, so inventing one would create a precision the source does not have. That is the same reason an unconfirmed requirement is published as unconfirmed rather than tidied away.

The disciplines

1. No number without a source. Every unit value and every requirement carries a source.
2. Every source declares its own standing, in three tiers. statute / agency / secondary. Material published by the ministry is not treated as equal to the statute text.
3. An unconfirmed item is never left blank. It carries confirmed: false and an unconfirmed_reason. Blank and unconfirmed are different things.
4. When sources disagree, we do not pick one and go quiet. Both readings go into conflicts, and the conflict stays marked unresolved. The moment one is chosen, the reason for choosing it disappears.
5. Nothing that reads as "you may bill this" is allowed into the data. The validator rejects it.
6. Revisions cut a new version; the old one is not deleted. Otherwise past decisions cannot be audited afterwards.
7. Every requirement is bound one-to-one to an interview question. A requirement with no question attached is a requirement nobody has actually asked about.

These are not aspirations. tools/validate.py checks them mechanically, and CI fails on any breach.

Public MCP endpoint (no key, read-only)

https://jhnrd-mcp.oga-surf-project.workers.dev

The dataset is served over MCP (JSON-RPC 2.0 / Streamable HTTP at POST /mcp) and over plain HTTP (/status.json, /items, /items/<id>, /sources, /search?q=, /unconfirmed, /conflicts, /gaps, /disclosure, /cite). No key. No write path.

The content is CC BY 4.0 already — but a JSON file on its own means every consumer writes a parser, and every parser mixes in its author's reading. Serving it means that happens in one place.

The same line is drawn on the serving side:

• It does not decide whether a claim may be billed. That it never returns such a statement is checked mechanically by the test suite.
• Every response carries the version, the "this does not decide billing" notice, and the conflict-of-interest disclosure. All eleven tools are tested for it.
• It holds no key and is wired to no storage. The moment it is, it starts holding something of someone's.
• The tools that expose the weaknesses come first: jhnrd_unconfirmed, jhnrd_gaps (including searches that came back empty), jhnrd_conflicts (both readings kept).

The copy it serves is generated from the dataset and never fetched at runtime — a CDN edge once kept a stale revision alive and an internal server handed out old numbers. CI fails if the copy drifts from the source.

Once a day, both the version the live endpoint is actually serving and the manifest listed in the official MCP registry are compared against the repository (the live badge above) — because pushing and forgetting to deploy, and publishing a manifest whose metadata the registry silently dropped, have both already happened here.

That check does not fail when the connection cannot be made — being unreachable is not evidence of being stale — but it does fail when an HTTP status comes back and is not 200, because a status is an answer. The first version of this gate conflated the two and spent its life reporting green while receiving 403s. Its logic is exercised against seventeen mocked endpoints on every push, since a gate that runs once a day can stay broken for a day without anyone knowing.

{ "mcpServers": { "jhnrd": { "type": "http", "url": "https://jhnrd-mcp.oga-surf-project.workers.dev/mcp" } } }

See mcp/README.md; the registry manifest is server.json.

Conflict of interest

The HORIZONs Inc., which maintains this dataset, sells a paid service to home-visit nursing providers that is backed by it. Payment is taken as an initial build fee and a monthly fee.

So: the people building this dataset are paid when it is believed to be useful. That has to be stated first, or the numbers above cannot be used. A dataset whose central discipline is that every source declares its own standing cannot exempt itself from declaring its own.

Four things could bend, and each is bound in advance:

Every number in this README is generated from status.json, and CI fails if it is edited by hand. Inflating a figure is not something that can be done quietly.