Narrow the evidence cell in the comparison table

Shorten from verbose prose to compact '946 / ~4,400 / 0' format with
a legend in the row label. Fixes the KitchenSink4Word column rendering
excessively wide.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: nometalalchemist

docs/index.html

@@ -259,7 +259,7 @@ <h1 class="rev d2">Kitchen&shy;Sink<span class="four">4</span>Word</h1>
         <tr><td data-i18n="mxr12">International style names</td><td class="us yes">✔</td><td class="nd">n/d</td><td class="no">✘</td><td class="nd">n/d</td><td class="nd">n/d</td></tr>
         <tr><td data-i18n="mxr13">Works without Word installed</td><td class="us yes" data-i18n="mv.file">✔ file tools; COM optional</td><td class="nd">n/d</td><td class="nd">n/d</td><td class="yes">✔</td><td class="no">✘</td></tr>
         <tr><td data-i18n="mxr14">Live editing while open in Word</td><td class="us yes" data-i18n="mv.undo">✔ one Ctrl+Z per call</td><td class="yes">✔</td><td class="no">✘</td><td class="no">✘</td><td class="yes" data-i18n="mv.inapp">✔ in-app</td></tr>
-        <tr><td data-i18n="mxr15">Published anti-corruption test evidence</td><td class="us yes" data-i18n="mv.evid">✔ 946 tests · ~4,400 adversarial calls · 0 corruptions</td><td class="nd">n/d</td><td class="nd">n/d</td><td class="nd">n/d</td><td class="nd">n/d</td></tr>
+        <tr><td data-i18n="mxr15">Anti-corruption evidence (tests / stress / corruptions)</td><td class="us yes" data-i18n="mv.evid">✔ 946 / ~4,400 / 0</td><td class="nd">n/d</td><td class="nd">n/d</td><td class="nd">n/d</td><td class="nd">n/d</td></tr>
       </table>
     </div>
     <p class="foot-note" data-i18n="mx.note">Capability survey as of 2026-08-28, compiled from public repositories, documentation, and issue trackers. n/d: not documented in the sources surveyed (not a claimed absence). * Official add-ins covers Microsoft's Word AI offerings and the Claude for Word add-in: cloud-based or Team/Enterprise scoped, not local MCP servers. Corrections welcome: open an issue.</p>
@@ -409,8 +409,8 @@ <h3 data-i18n="g.t">Engineered Not to Corrupt</h3>
 "mxr12": "International style names",
 "mxr13": "Works without Word installed",
 "mxr14": "Live editing while open in Word",
-"mxr15": "Published anti-corruption test evidence",
-"mv.s12": "✔ 12 styles", "mv.s8": "✔ 8 styles", "mv.both": "✔ both directions", "mv.sha": "✔ Word-compatible SHA-512", "mv.file": "✔ file tools; COM optional", "mv.undo": "✔ one Ctrl+Z per call", "mv.evid": "✔ 946 tests · ~4,400 adversarial calls · 0 corruptions", "mv.inapp": "✔ in-app",
+"mxr15": "Anti-corruption evidence (tests / stress / corruptions)",
+"mv.s12": "✔ 12 styles", "mv.s8": "✔ 8 styles", "mv.both": "✔ both directions", "mv.sha": "✔ Word-compatible SHA-512", "mv.file": "✔ file tools; COM optional", "mv.undo": "✔ one Ctrl+Z per call", "mv.evid": "✔ 946 / ~4,400 / 0", "mv.inapp": "✔ in-app",
 "mx.note": "Capability survey as of 2026-08-28, compiled from public repositories, documentation, and issue trackers. n/d: not documented in the sources surveyed (not a claimed absence). * Official add-ins covers Microsoft's Word AI offerings and the Claude for Word add-in: cloud-based or Team/Enterprise scoped, not local MCP servers. Corrections welcome: open an issue.",
 "h2.eff": "Context efficiency",
 "eff.lead": "Every MCP server occupies part of an agent's context window before any work starts, and almost none of them tell you how much. Here is our bill, measured and published: when a client loads every tool schema and description up front, the full 170-tool surface costs about 27,000 tokens, an average of roughly 144 tokens per tool. Clients that defer tool schemas, as Claude Code does, pay close to zero until a tool is actually called. Publishing the number is the point: you should be able to see what the sink costs before you connect it.",
@@ -504,7 +504,7 @@ <h3 data-i18n="g.t">Engineered Not to Corrupt</h3>
 "mxr13": "Word 미설치 환경에서 동작",
 "mxr14": "Word에 열린 채 라이브 편집",
 "mxr15": "손상 방지 테스트 근거 공개",
-"mv.s12": "✔ 12개 스타일", "mv.s8": "✔ 8개 스타일", "mv.both": "✔ 양방향", "mv.sha": "✔ Word 호환 SHA-512", "mv.file": "✔ 파일 도구; COM은 선택", "mv.undo": "✔ 호출당 Ctrl+Z 한 단계", "mv.evid": "✔ 테스트 946건 · 적대적 호출 ~4,400건 · 손상 0건", "mv.inapp": "✔ 앱 내",
+"mv.s12": "✔ 12개 스타일", "mv.s8": "✔ 8개 스타일", "mv.both": "✔ 양방향", "mv.sha": "✔ Word 호환 SHA-512", "mv.file": "✔ 파일 도구; COM은 선택", "mv.undo": "✔ 호출당 Ctrl+Z 한 단계", "mv.evid": "✔ 946 / ~4,400 / 0", "mv.inapp": "✔ 앱 내",
 "mx.note": "2026-08-28 기준 기능 조사. 공개 저장소, 문서, 이슈 트래커에서 취합했습니다. n/d: 조사한 자료에서 확인되지 않음(없다는 주장이 아닙니다). * 공식 추가 기능 열은 Microsoft의 Word AI 제품군과 Claude for Word 추가 기능을 묶은 것으로, 클라우드 기반이거나 Team/Enterprise 요금제 전용이며 로컬 MCP 서버가 아닙니다. 정정은 환영합니다. 이슈를 열어 주세요.",
 "h2.eff": "컨텍스트 효율",
 "eff.lead": "모든 MCP 서버는 작업을 시작하기도 전에 에이전트의 컨텍스트 창 일부를 차지하지만, 그 비용을 알려주는 서버는 거의 없습니다. 여기 측정해서 공개한 우리 청구서가 있습니다: 클라이언트가 모든 도구 스키마와 설명을 미리 불러오면 170개 도구 전체 표면이 약 27,000 토큰, 도구당 평균 약 144 토큰입니다. Claude Code처럼 도구 스키마를 지연 로드하는 클라이언트라면 도구가 실제로 호출되기 전까지 비용이 거의 0에 가깝습니다. 숫자를 공개한다는 것 자체가 핵심입니다: 연결하기 전에 싱크대의 비용을 볼 수 있어야 하니까요.",
@@ -598,7 +598,7 @@ <h3 data-i18n="g.t">Engineered Not to Corrupt</h3>
 "mxr13": "Word 未インストールでも動作",
 "mxr14": "Word で開いたままライブ編集",
 "mxr15": "破損防止テストの証跡を公開",
-"mv.s12": "✔ 12 スタイル", "mv.s8": "✔ 8 スタイル", "mv.both": "✔ 双方向", "mv.sha": "✔ Word 互換 SHA-512", "mv.file": "✔ ファイルツール、COM は任意", "mv.undo": "✔ 呼び出し1回 = Ctrl+Z 1段階", "mv.evid": "✔ テスト 946 件 · 敵対的呼び出し ~4,400 件 · 破損 0 件", "mv.inapp": "✔ アプリ内",
+"mv.s12": "✔ 12 スタイル", "mv.s8": "✔ 8 スタイル", "mv.both": "✔ 双方向", "mv.sha": "✔ Word 互換 SHA-512", "mv.file": "✔ ファイルツール、COM は任意", "mv.undo": "✔ 呼び出し1回 = Ctrl+Z 1段階", "mv.evid": "✔ 946 / ~4,400 / 0", "mv.inapp": "✔ アプリ内",
 "mx.note": "2026-08-28 時点の機能調査。公開リポジトリ、ドキュメント、Issue トラッカーから集約しました。n/d:調査した資料では確認できず(欠如の断定ではありません)。* 公式アドインは Microsoft の Word AI 製品群と Claude for Word アドインをまとめた列で、クラウドまたは Team/Enterprise 限定のもの。ローカル MCP サーバーではありません。訂正歓迎、Issue をどうぞ。",
 "h2.eff": "コンテキスト効率",
 "eff.lead": "どの MCP サーバーも、仕事を始める前からエージェントのコンテキストウィンドウの一部を占めます。それなのに、その量を教えてくれるサーバーはほとんどありません。これが計測して公開した当方の請求書です:クライアントが全ツールのスキーマと説明を先読みした場合、170 ツールの全表面で約 27,000 トークン、1 ツールあたり平均およそ 144 トークン。Claude Code のようにツールスキーマを遅延読み込みするクライアントなら、ツールが実際に呼ばれるまでコストはほぼゼロです。数字を公開すること自体が要点です:接続する前に、この流しの値段が見えるべきだからです。",
@@ -692,7 +692,7 @@ <h3 data-i18n="g.t">Engineered Not to Corrupt</h3>
 "mxr13": "无需安装 Word 即可运行",
 "mxr14": "文档在 Word 中打开时实时编辑",
 "mxr15": "公开的防损坏测试证据",
-"mv.s12": "✔ 12 种样式", "mv.s8": "✔ 8 种样式", "mv.both": "✔ 双向", "mv.sha": "✔ Word 兼容 SHA-512", "mv.file": "✔ 文件工具;COM 可选", "mv.undo": "✔ 每次调用一步 Ctrl+Z", "mv.evid": "✔ 946 项测试 · ~4,400 次对抗性调用 · 0 次损坏", "mv.inapp": "✔ 应用内",
+"mv.s12": "✔ 12 种样式", "mv.s8": "✔ 8 种样式", "mv.both": "✔ 双向", "mv.sha": "✔ Word 兼容 SHA-512", "mv.file": "✔ 文件工具;COM 可选", "mv.undo": "✔ 每次调用一步 Ctrl+Z", "mv.evid": "✔ 946 / ~4,400 / 0", "mv.inapp": "✔ 应用内",
 "mx.note": "能力调查,截至 2026-08-28,汇总自公开仓库、文档与 issue 跟踪器。n/d:在所查资料中未见记载(并非断言其没有)。* 官方加载项一栏合并了 Microsoft 的 Word AI 产品与 Claude for Word 加载项:皆为云端或 Team/Enterprise 专属,不是本地 MCP 服务器。欢迎指正,请开 issue。",
 "h2.eff": "上下文效率",
 "eff.lead": "每个 MCP 服务器在干活之前就要占用智能体的一部分上下文窗口,却几乎没有服务器告诉你占多少。这是我们量出来并公开的账单:客户端若把所有工具的 schema 和说明一次性载入,170 件工具的完整表面约需 27,000 token,平均每件工具约 144 token。像 Claude Code 这样延迟加载工具 schema 的客户端,在工具真正被调用之前成本几乎为零。公开数字本身就是重点:接上水槽之前,你应该先看得到价钱。",
@@ -786,7 +786,7 @@ <h3 data-i18n="g.t">Engineered Not to Corrupt</h3>
 "mxr13": "Läuft ohne installiertes Word",
 "mxr14": "Live-Bearbeitung bei in Word geöffnetem Dokument",
 "mxr15": "Veröffentlichte Anti-Korruptions-Testbelege",
-"mv.s12": "✔ 12 Stile", "mv.s8": "✔ 8 Stile", "mv.both": "✔ beide Richtungen", "mv.sha": "✔ Word-kompatibles SHA-512", "mv.file": "✔ Dateiwerkzeuge; COM optional", "mv.undo": "✔ ein Strg+Z pro Aufruf", "mv.evid": "✔ 946 Tests · ~4.400 adversariale Aufrufe · 0 Beschädigungen", "mv.inapp": "✔ in der App",
+"mv.s12": "✔ 12 Stile", "mv.s8": "✔ 8 Stile", "mv.both": "✔ beide Richtungen", "mv.sha": "✔ Word-kompatibles SHA-512", "mv.file": "✔ Dateiwerkzeuge; COM optional", "mv.undo": "✔ ein Strg+Z pro Aufruf", "mv.evid": "✔ 946 / ~4.400 / 0", "mv.inapp": "✔ in der App",
 "mx.note": "Fähigkeitserhebung, Stand 2026-08-28, zusammengetragen aus öffentlichen Repositories, Dokumentation und Issue-Trackern. n/d: in den gesichteten Quellen nicht dokumentiert (keine behauptete Abwesenheit). * Offizielle Add-ins bündelt Microsofts Word-KI-Angebote und das Claude-for-Word-Add-in: cloudbasiert oder auf Team/Enterprise beschränkt, keine lokalen MCP-Server. Korrekturen willkommen: einfach ein Issue öffnen.",
 "h2.eff": "Kontext-Effizienz",
 "eff.lead": "Jeder MCP-Server belegt einen Teil des Kontextfensters eines Agenten, bevor die Arbeit überhaupt beginnt, und kaum einer sagt, wie viel. Hier ist unsere Rechnung, gemessen und veröffentlicht: Lädt ein Client jedes Werkzeugschema samt Beschreibung im Voraus, kostet die volle Oberfläche mit 170 Werkzeugen etwa 27.000 Token, im Schnitt rund 144 Token pro Werkzeug. Clients, die Werkzeugschemata aufschieben, wie Claude Code, zahlen nahezu nichts, bis ein Werkzeug tatsächlich aufgerufen wird. Die Zahl zu veröffentlichen ist der Punkt: Man sollte sehen können, was die Spüle kostet, bevor man sie anschließt.",
@@ -880,7 +880,7 @@ <h3 data-i18n="g.t">Engineered Not to Corrupt</h3>
 "mxr13": "Fonctionne sans Word installé",
 "mxr14": "Édition en direct, document ouvert dans Word",
 "mxr15": "Preuves publiées des tests anti-corruption",
-"mv.s12": "✔ 12 styles", "mv.s8": "✔ 8 styles", "mv.both": "✔ dans les deux sens", "mv.sha": "✔ SHA-512 compatible Word", "mv.file": "✔ outils fichier ; COM en option", "mv.undo": "✔ un Ctrl+Z par appel", "mv.evid": "✔ 946 tests · ~3 200 appels adversariaux · 0 corruption", "mv.inapp": "✔ dans l'application",
+"mv.s12": "✔ 12 styles", "mv.s8": "✔ 8 styles", "mv.both": "✔ dans les deux sens", "mv.sha": "✔ SHA-512 compatible Word", "mv.file": "✔ outils fichier ; COM en option", "mv.undo": "✔ un Ctrl+Z par appel", "mv.evid": "✔ 946 / ~4.400 / 0", "mv.inapp": "✔ dans l'application",
 "mx.note": "Recensement des capacités au 2026-08-28, compilé à partir des dépôts publics, de la documentation et des traqueurs de tickets. n/d : non documenté dans les sources consultées (ce n'est pas une absence affirmée). * Compléments officiels regroupe les offres IA de Microsoft pour Word et le complément Claude for Word : en nuage ou réservés aux offres Team/Enterprise, pas des serveurs MCP locaux. Corrections bienvenues : ouvrez un ticket.",
 "h2.eff": "Efficacité de contexte",
 "eff.lead": "Chaque serveur MCP occupe une partie de la fenêtre de contexte d'un agent avant même que le travail commence, et presque aucun ne dit combien. Voici notre facture, mesurée et publiée : si un client charge d'avance chaque schéma d'outil et sa description, la surface complète des 170 outils coûte environ 27 000 tokens, soit en moyenne quelque 144 tokens par outil. Les clients qui diffèrent les schémas d'outils, comme Claude Code, ne paient presque rien tant qu'un outil n'est pas réellement appelé. Publier le chiffre, c'est justement l'idée : on devrait pouvoir voir ce que coûte l'évier avant de le raccorder.",
@@ -974,7 +974,7 @@ <h3 data-i18n="g.t">Engineered Not to Corrupt</h3>
 "mxr13": "Funciona sin Word instalado",
 "mxr14": "Edición en vivo con el documento abierto en Word",
 "mxr15": "Evidencia publicada de pruebas anticorrupción",
-"mv.s12": "✔ 12 estilos", "mv.s8": "✔ 8 estilos", "mv.both": "✔ en ambos sentidos", "mv.sha": "✔ SHA-512 compatible con Word", "mv.file": "✔ herramientas de archivo; COM opcional", "mv.undo": "✔ un Ctrl+Z por llamada", "mv.evid": "✔ 946 pruebas · ~4.400 llamadas adversariales · 0 corrupciones", "mv.inapp": "✔ dentro de la aplicación",
+"mv.s12": "✔ 12 estilos", "mv.s8": "✔ 8 estilos", "mv.both": "✔ en ambos sentidos", "mv.sha": "✔ SHA-512 compatible con Word", "mv.file": "✔ herramientas de archivo; COM opcional", "mv.undo": "✔ un Ctrl+Z por llamada", "mv.evid": "✔ 946 / ~4.400 / 0", "mv.inapp": "✔ dentro de la aplicación",
 "mx.note": "Estudio de capacidades a 2026-08-28, compilado de repositorios públicos, documentación y gestores de incidencias. n/d: no documentado en las fuentes consultadas (no es una ausencia afirmada). * Complementos oficiales agrupa las ofertas de IA de Microsoft para Word y el complemento Claude for Word: en la nube o limitados a planes Team/Enterprise, no servidores MCP locales. Correcciones bienvenidas: abra un issue.",
 "h2.eff": "Eficiencia de contexto",
 "eff.lead": "Todo servidor MCP ocupa parte de la ventana de contexto de un agente antes de empezar a trabajar, y casi ninguno dice cuánto. Esta es nuestra factura, medida y publicada: si un cliente carga por adelantado cada esquema de herramienta con su descripción, la superficie completa de 170 herramientas cuesta unos 27.000 tokens, una media de unos 144 tokens por herramienta. Los clientes que aplazan los esquemas, como Claude Code, no pagan casi nada hasta que una herramienta se invoca de verdad. Publicar la cifra es justamente la gracia: uno debería poder ver lo que cuesta el fregadero antes de conectarlo.",