エクスクルーシブ
Drift Check <mv-drift-check>
中継されたテキストが元の情報からどのようにずれていったかをホップごとに示します。顧客から営業、PM、チケットへと伝わる要件、インタビューからプレスリリース、記事へと渡る引用、要約の要約、転送・翻訳・AI による書き換えを経た文章などが対象です。
文化的リファレンス
伝言ゲーム(子どもの遊び)、伝承(英国では Chinese whispers として知られる)(19世紀、ゲーム)。ひとつの文が列に並んだプレイヤーの間で次々に伝えられ、変化して届きます。楽しさは、最後のバージョンを元の文と比べ、どこでおかしくなったかを一歩ずつたどり直すことにあります。UI では、テキストの中継(営業メモ、AI の要約、書き直し)がそれぞれ 1 つのホップです。コンポーネントは最終版を元の文と比較し、各ホップがどれだけ変えたかをスコア化し、意味が崩れたホップを示し、元の言い回しを復元できるようにします。
仕組み
中継されたテキストが元の文からどうずれていったかを、ホップごとに示します。顧客から営業、PM、チケットへと受け渡される要件、インタビューからプレスリリース、記事へと渡る引用、要約の要約、転送・翻訳・AI で書き直された文面などが対象です。バージョンの順序付きチェーン(元の文が先頭。それぞれにラベル、誰が、どのように、いつを付け、<template data-version> の子要素または versions プロパティで指定)を渡すと、独自の Myers diff で各ホップを単語単位で比較し、各変更に意味的リスクのルールを適用します。対象は、数値や日付の変更、否定の消失や追加、より強いまたは弱い法助動詞(「may need」→「must」)、より広いまたは狭い数量詞(「some」→「most」)、削除された条件や限定(「if」、「only」、「up to」)、入れ替わった名前、そして terms に列挙した用語集の語です。ドリフトスコア(変更された単語の割合に、重大な変更ごとの重みを加えたもの)が判定(「Meaning shifted · 7 high-risk changes across 4 hops · most introduced at hop 2」)の先頭に表示されます。連結されたホップカードのレールは各段階がどれだけ変えたかを示し、意味が崩れたホップにフラグを立てます。最終版には取り消し線、下線、番号付きのテキストタグが描かれ、それぞれがチェーンをさかのぼって、その変更を持ち込んだホップまで追跡されます。任意のホップをクリックするか、任意の 2 つのバージョンを選ぶと、インラインまたは左右に並べて比較できます。フラグが付いた変更は平易な文で一覧表示され、Restore ボタンで元の言い回しを最終版に戻せます(キャンセル可能な mv-restore、元に戻す操作に対応)。純粋な静的メソッド MvDriftCheck.compare(a, b) は、同じ diff、リスク、スコアをサーバー側で返します。
| カテゴリー | データ表示 |
|---|---|
| タイプ | Web Component(<mv-drift-check>) |
| ステータス | 安定版 |
| 同時にインストールされるもの | button, select |
| Keywords | exclusive, culture, diff, word-diff, text-diff, compare, drift, paraphrase, summary, ai-review, translation, requirements, quote-check, fact-check, provenance, hops, restore, server-side |
When to use
- A requirement went from a customer email through sales notes and a PM brief to a ticket, and the team must see what changed on the way
- An editor or fact-checker must compare a published quote with the original interview and find where numbers or hedges changed
- An AI summary, rewrite or translation must be reviewed for lost negations, stronger claims or changed figures before it ships
- A forwarded policy or support macro drifted across teams and the exact hop that changed its meaning must be found and fixed
Avoid when
- One document is edited over time and the question is how much of the signed or approved text survives, and who wrote each part → use Theseus instead
- Two structured records must be reconciled field by field (invoice vs order, form vs ID) → use Cross-Check instead
- The claims of a single text must be checked against sources rather than against an earlier version → use Unsourced instead
インストール
node scripts/add.mjs drift-check --out ./src/marvelousMarvelous UI MCP サーバーを使う AI エージェント:install_components({ slugs: ["drift-check"], target_dir: "<absolute path>/src/marvelous", framework: "react" })。
コピーされるファイル(依存関係を含む):tokens/tokens.css, core/base.css, components/button/button.css, components/select/select.css, core/dom.js, core/element.js, core/focus.js, core/motion.js, components/drift-check/drift-check.js, components/drift-check/drift-check.css。
使い方
基本となるマークアップです。ここから属性、data-*、CSS 変数でカスタマイズしてください:
<div id="dc-demo" style="width:min(100%,72rem);margin-inline:auto">
<style>
#dc-demo { display:grid; gap:1rem }
#dc-demo .dc-bar { display:flex; flex-wrap:wrap; align-items:center; justify-content:space-between; gap:.75rem 1.25rem }
#dc-demo .dc-title { display:grid; gap:.125rem; min-width:0 }
#dc-demo .dc-title small { color:var(--mv-fg-muted); font-size:.75rem }
#dc-demo .dc-title strong { font-size:1rem; letter-spacing:-.01em }
#dc-demo .dc-controls { display:flex; flex-wrap:wrap; align-items:center; gap:.625rem 1.25rem }
#dc-demo .dc-controls .mv-choice { font-size:.8125rem }
#dc-demo .dc-log { display:flex; align-items:center; gap:.625rem; min-height:2.5rem; margin:0; padding:.5rem .875rem; border:1px dashed var(--mv-border); border-radius:var(--mv-radius-lg); color:var(--mv-fg-muted); font-size:.8125rem }
#dc-demo .dc-log b { color:var(--mv-fg); font-weight:500 }
#dc-demo .dc-log code { font-family:var(--mv-font-mono); font-size:.75rem; color:var(--mv-fg) }
</style>
<div class="dc-bar">
<span class="dc-title">
<small id="dc-kicker">Harborline Freight · Feature request FR-2291</small>
<strong id="dc-heading">Bulk export for warehouse reports</strong>
</span>
<span class="dc-controls">
<mv-segmented id="dc-scenario" aria-label="Scenario" value="request">
<button value="request">Feature request</button>
<button value="quote">Press quote</button>
</mv-segmented>
<label class="mv-choice">
<input type="checkbox" role="switch" class="mv-switch" id="dc-review">
<span class="mv-choice-text"><span class="mv-choice-title">Require reviewer to restore</span></span>
</label>
</span>
</div>
<mv-drift-check id="dc-check" terms="EU workspace, Rotterdam">
<template data-version data-label="Customer email" data-by="Dana Whitfield" data-via="Head of Operations, Harborline" data-at="2026-09-02">
We may need the export to handle up to 50,000 rows by the end of Q4. It should not email reports to all users, only to admins in the EU workspace. If the pilot in Rotterdam goes well, we would roll it out to the other 3 warehouses.
</template>
<template data-version data-label="Call notes" data-by="Marco Bellini" data-via="Paraphrased from a call" data-at="2026-09-04">
Customer may need the export to handle 50,000 rows by the end of Q4. It should not email reports to all users, only admins in the EU workspace. If the Rotterdam pilot goes well, they would roll it out to the other 3 warehouses.
</template>
<template data-version data-label="Account summary" data-by="AI assistant" data-via="Summarized by AI" data-at="2026-09-05">
The customer needs the export to handle 50,000 rows by the end of Q4. Reports should be emailed to all users, including admins in the EU workspace. After the Rotterdam pilot, they will roll it out to the other 3 warehouses.
</template>
<template data-version data-label="Product brief" data-by="Priya Nair" data-via="Edited for the roadmap" data-at="2026-09-08">
The customer needs the export to handle 500,000 rows by the end of Q3. Reports should be emailed to all users, including admins in every workspace. After the Rotterdam pilot, we will roll it out to all 4 warehouses.
</template>
<template data-version data-label="Engineering ticket" data-by="Omar Haddad" data-via="Rewritten as a ticket" data-at="2026-09-10">
Export must handle 500,000 rows by end of Q3. Email reports to all users, including admins in every workspace. After the Rotterdam pilot, we will roll it out to all 4 warehouses.
</template>
</mv-drift-check>
<p class="dc-log" id="dc-log" aria-live="off"><span>Pick a hop, switch to <b>Side by side</b>, or restore a flagged span of the ticket.</span></p>
<script type="module">
const $ = (id) => document.getElementById(id);
const check = $("dc-check");
await customElements.whenDefined("mv-drift-check");
const scenarios = {
request: {
kicker: "Harborline Freight · Feature request FR-2291",
heading: "Bulk export for warehouse reports",
terms: "EU workspace, Rotterdam",
versions: null, // the <template> children
},
quote: {
kicker: "Newsroom fact-check · Quote review",
heading: "“Early results” quote, from interview to social post",
terms: "",
versions: [
{ label: "Interview transcript", by: "Dr. Amara Okafor", via: "Lead researcher, recorded interview", at: "2026-08-18",
text: "In our trial, the treatment reduced symptoms in some patients, but we don't yet know whether the effect lasts beyond 12 weeks. We would not recommend it outside a study for now." },
{ label: "Press release", by: "University press office", via: "Edited for the press release", at: "2026-08-20",
text: "In the trial, the treatment reduced symptoms in many patients, though it is not yet known whether the effect lasts beyond 12 weeks. The team would not recommend it outside a study for now." },
{ label: "News article", by: "Kenji Watanabe", via: "Quoted in an article", at: "2026-08-22",
text: "The treatment reduced symptoms in most patients, and the effect lasts for 12 weeks. The team would recommend it outside a study." },
{ label: "Social post", by: "Health Daily", via: "Shortened for social", at: "2026-08-23",
text: "New treatment cures symptoms in most patients for 3 months. Doctors will recommend it." },
],
},
};
const original = check.versions;
$("dc-scenario").addEventListener("mv-change", (e) => {
const s = scenarios[e.detail.value];
$("dc-kicker").textContent = s.kicker;
$("dc-heading").textContent = s.heading;
check.terms = s.terms || null;
check.removeAttribute("from");
check.removeAttribute("to");
check.layout = e.detail.value === "quote" ? "split" : "inline";
check.versions = s.versions ?? original;
log(`Loaded ${check.versions.length} versions.`);
});
const log = (html) => {
const p = $("dc-log");
p.replaceChildren(...html.map ? html : [html]);
};
const el = (tag, text) => Object.assign(document.createElement(tag), { textContent: text });
check.addEventListener("mv-hop-select", (e) => {
const { from, to, comparison } = e.detail;
log([el("code", "mv-hop-select"), el("span", ` ${from} → ${to} · drift ${comparison.score} · ${comparison.counts.high} high-risk`)]);
});
check.addEventListener("mv-restore", (e) => {
if ($("dc-review").checked) {
e.preventDefault();
log([el("code", "mv-restore"), el("span", " declined: a reviewer must approve changes to the ticket.")]);
return;
}
const c = e.detail.change;
log([el("code", "mv-restore"), el("span", " "), el("b", c.removed ? `“${c.removed}”` : `removed “${c.added}”`), el("span", " is back in the ticket.")]);
});
check.addEventListener("mv-drift-found", (e) => {
const { counts, breakHop } = e.detail;
if (breakHop != null) log([el("code", "mv-drift-found"), el("span", ` ${counts.high} high-risk changes · meaning shifted most at hop ${breakHop}.`)]);
});
</script>
</div>API
Attributes
| Name | タイプ | Default | Description |
|---|---|---|---|
from / to | number (version index) | The pair being compared (0 = source). Defaults to source → final. Clicking the source card selects 0 → last, a hop card selects (n − 1) → n; the pickers allow any pair, reversed pairs included. | |
layout | inline | split | inline | Inline paints the “to” version with removed words struck through before the new ones; split shows both versions side by side (stacked in narrow containers), with the same numbered tags on each side. |
terms | comma-separated list | Glossary terms (product names, legal phrases, “EU workspace”…): any change touching one is flagged as a high-risk Term change. | |
readonly | boolean | Hides the Restore buttons (review-only). | |
label | string | Relay drift | Accessible name of the region. |
locale | BCP 47 tag | en-US | Date formatting on the hop cards. |
<template data-version> | children | One per version, in chain order, source first. The template text is the version (indentation and soft wraps collapsed, blank lines kept as paragraphs); data-label (what it is: “Call notes”), data-by (who), data-via (how: “Summarized by AI”) and data-at (ISO date or free text) describe the hop. | |
data-level / data-layout / data-playing | set by the component | On the host: faithful | reworded | shifted (end-to-end verdict), the current layout, and data-playing while Step through runs. |
Properties
| Name | タイプ | Description |
|---|---|---|
versions | Array<{ text, label?, by?, via?, at? }> | The chain, source first (wins over templates). Setting it resets restores and undo; reading it returns copies, restores included. |
value | string | Current text of the final version, restores included (read-only). |
report | { score, changed, counts, breakHop, hops[], changes[] } | End-to-end analysis (read-only): drift score, share of words changed, counts by severity, the hop that introduced most high-risk changes, per-hop scores, and every change with the hops it came from and a full sentence description. |
strings | Partial<Record<string, string>> | Overrides for every visible text and announcement (English defaults; see STRINGS in drift-check.js). |
from / to / layout / terms / readonly / label / locale | reflected | Mirror the attributes. |
Methods
| Name | Description |
|---|---|
static compare(a, b, { terms? }) | Pure, DOM-free comparison usable server-side or in tests. Returns { score (0-100), changed (0-1), severity, counts: { high, medium, low, none }, changes: [{ id, type: replace | insert | delete, removed, added, severity, risks: [{ type, severity, direction, detail }], from: [start, end], to: [start, end], description }], parts: [{ type: equal | delete | insert, text }] }. Also exported as compareText. |
select(from, to) | Compares any two versions (indexes). Does not emit mv-hop-select. |
restore(id) | Reverts change id of the current comparison to the “from” wording in the final version (goes through the cancelable mv-restore). Returns true when applied. |
undo() | Undoes the last restore. Returns false when there is nothing to undo. |
play() / stop() | Steps through the chain hop by hop (about 2 s each), then returns to source → final; any click or key press inside stops it. |
Events
| Name | Description |
|---|---|
mv-hop-select | The user picked a hop or a pair. detail: { index (hop number, 0 for source → final, null for a custom pair), from, to, version, comparison: { score, counts, changes } }. |
mv-drift-found | After the chain is analyzed, when source → final has at least one medium or high-risk change (once per distinct result). detail: { score, counts, breakHop, findings: [{ …change, hops, description }] }. |
mv-restore | Cancelable, before a span of the final version is reverted. detail: { change, from, before, after } (after = the resulting final text). preventDefault() keeps the text as is (e.g. the ticket needs a reviewer). |
mv-change | The final text changed through a restore or an undo. detail: { value, reason: "restore" | "undo" }. |
Content structure
| Name | Description |
|---|---|
template[data-version] | Direct <template data-version> children are the chain (ignored once the versions property is set). |
CSS classes
| Name | Description |
|---|---|
mv-drift-check-head | Score (-score-num, role=meter -meter), verdict (-verdict with data-level), summary line, flagged kinds (-kinds / -kind), actions (Undo restore, Step through). |
mv-drift-check-rail / mv-drift-check-hop | The chained hop cards (role=toolbar of buttons with aria-pressed): -hop-top, -hop-num, -hop-risk, -hop-break (data-break), -hop-label, -hop-by, -hop-via, -hop-bar, -hop-foot; data-level low | medium | high, data-in-range and data-endpoint for the pair being compared. |
mv-drift-check-panel | Comparison: -caption, -toolbar (pickers, -segs layout toggle), -view (-text, -split with -col), -findings (-list of -finding with -chip, -finding-diff, -hopchip, -restore; -more toggles minor changes). |
mv-drift-check-change / -del / -ins / -tag / -gap | Diff marks: a change wraps <del> and <ins> (data-severity high = wavy underline, medium = dashed), a numbered text tag for flagged changes, and a gap marker in split view where words were only added or removed. |
CSS variables
| Name | Default | Description |
|---|---|---|
--mv-drift-check-added | var(--mv-success) | Tint of inserted words. |
--mv-drift-check-removed | var(--mv-danger) | Tint and strike of removed words. |
--mv-drift-check-risk | var(--mv-danger) | High-risk tags, wavy underline, hop flags and the shifted verdict. |
--mv-drift-check-caution | var(--mv-warning) | Medium-risk tags and the reworded verdict. |
--mv-drift-check-faithful | var(--mv-success) | Score bar color when the text stayed faithful. |
--mv-drift-check-text-size | var(--mv-text-base) | Font size of the inline text view. |
Accessibility
The component is a labelled region. The drift score is a role="meter" with a spoken value (“64 of 100, meaning shifted”), and the verdict and summary are plain text. The hop rail is a role="toolbar" of real buttons with aria-pressed and roving tabindex (Left/Right, Home/End, Enter or Space to compare); each card has a full accessible name (“Hop 2: Account summary, AI assistant, summarized by AI, September 5, 2026. Drift score 57, 5 high-risk changes; 72 from the source. Meaning changed most here.”). Diffs never rely on color: removed words are real <del> with line-through and inserted words real <ins> with an underline (wavy for high risk, dashed for medium), each prefixed with visually hidden “Removed:” / “Added:”, and flagged changes carry a numbered text tag (“3 Negation”) plus a hidden “(flag 3: negation dropped)”. Every flagged change is also listed as a sentence (“Changed “It should not email reports” to “Reports should be emailed” at hop 2 (AI assistant): negation dropped. High risk.”) next to its Restore button, whose name quotes the wording it brings back; hop chips jump to the hop that introduced the change. Selecting a pair, restoring and undoing are announced in a polite live region; after a restore focus moves to the next Restore button, so keyboard users keep their place. The pickers are native selects with labels, the layout toggle uses aria-pressed, the minor-changes toggle aria-expanded/aria-controls. Hovering or focusing a change or its finding highlights both. Step through is opt-in, stops on any key or click (and on Escape). Reduced motion (OS or data-motion="reduce"): no view fade, bars jump instead of sliding. Forced colors: tags, chips and flags get system-color borders, the selected hop an Highlight outline, bars use Highlight.