Kashiあり
同じ会議構造を、役割ごとに異なる粒度で表示します。
管理職本人
自分の会議パターンを振り返るための画面です。
本人
自分に関係するシグナルと、アクセス履歴・確認手段を確認できます。
経営層
5名以上の条件を満たした集計パターンのみを確認できます。個人単位の部下データは表示されません。
Kashi は判定ではなく、役割ごとに制限された構造シグナルを表示します。同じ画面を全員に見せる設計ではありません。
管理職本人の振り返り
この画面は、管理職本人だけが確認できます。
部下個人のレポートは表示されません。
今週の確認候補
会議構造上の偏りが、一定期間続いています。
これは判定ではなく、本人が会議の進め方を振り返るための材料です。
次回のプロダクトレビューでは、冒頭10分を該当メンバーの共有時間として確保します。
これは判定ではありません。Kashi はハラスメントを判定しません。
グラフの詳細を見る
解析方法を見る
根拠となるシグナルを見る
表示範囲とガバナンスを見る
本人の確認画面
この画面は、本人だけが確認できます。
管理職は、このページを見ることができません。
関連するシグナル
あなたに関係する会議上のシグナルがあります。
これは誰かを判定するものではなく、必要に応じて確認や相談につなげるための情報です。
発言シェア詳細
敬語・言語指標の傾向
ユーザーの権利詳細
最近のアクセス履歴
全アクセス履歴
分析の拒否について
補足事項
経営層向け集計サマリー
経営層には、条件を満たした集計パターンのみを表示します。
個人単位の部下データは表示されません。
この画面は集計傾向を見るためのものです。個人単位の判断や人事評価には使用しません。
- 確認候補のある集計パターン
- 表示対象チーム数
- 集計に含まれない小規模チーム
- 確認候補のある集計パターン
- 表示対象チーム数
- 集計に含まれない小規模チーム
差分プライバシー
k-匿名性について
シグナルの構成
管理職別の表示について
データ保持期間
監査の透明性
根拠グレード
EU AI法への対応
就業規則の改定
労働組合との協議
一つのデータから、四つの視点を。
個人、チーム、マネージャー、そして組織全体。各階層が異なる問いに応えます。読み取るのは「構造」のみ。デフォルトですべてのデータはメタデータとして扱われます。
第1層 — 個人のサイン
話者ごとの基準値：発話シェアの推移、強引な割り込み率、無視された発言、応答の遅延などを測定。一定のデータが蓄積されるまで推測は行われません。
同じ1週間を「構造」で捉える。
Kimuraマネージャーだけに
パーソナル・リフレクション ・ Kimura (Manager)
Interaction asymmetry toward one team member has intensified over the past 60 days. Their speaking-share is down 68% vs their own baseline, and interruptions from Kimura are concentrated on this person at 4.7× the rate of other team members. Additionally, politeness-register (敬語) analysis shows Kimura uses plain-form speech toward this person (score 0.38) while maintaining honorific register toward peers (0.88) — a 0.50 gap, review-worthy under MHLW 類型2 evidence standards.
This is not a verdict on a specific person. It is a structural pattern notification. Kashi does not detect harassment.
出典を表示
Schmid Mast 2002 — Dominance and gender in the floor-allocation literature.
Whom did you interrupt, and how often
How an interruption is counted
An interruption is counted when another speaker starts while you are still speaking AND your turn ends within 500 ms. Structural definition only — no content or affect inference.
vs. their own baseline
Why we compare against the person's own baseline
Compared against the person's own past baseline, not the team average. This protects against false positives caused by introversion, cultural differences, or second-language speakers.
追加シグナル：発言順序の偏りと参加率の低下▾
How often each team member's turn was “passed over”
出典を表示
“Passed over” = a substantive turn (> 3 s) followed either by silence ≥ 2× the meeting's typical inter-turn gap, or by another speaker shifting the topic at length. Based on the classic conversation-analysis turn-taking model (Sacks et al., 1974). Content is not read.
Moments where someone's participation dropped sharply right after one of your turns
This signal fires when, within 5 minutes of one of your turns, a colleague's participation drops by ≥ 40%below their own baseline. The triggering turn's content is never visible to Kashi.
出典を表示
Budworth et al. 2015 — Looking forward to performance improvement: future-oriented feedback (feedforward) shown to outperform backward-looking review.
- In the next two product reviews, wait for sentence completion before redirecting. If you find yourself interrupting, pause and ask one clarifying question instead.
- When a proposal comes from a lower-participation speaker, explicitly acknowledge it by name before the discussion moves on (“That's a phased-rollout proposal from Nao. Let's consider it.”).
Naoに見えているもの。
あなたの
ウィットネスビュー ・ Nao
Aggregated signals over 60 days involve interactions between you and Kimura (Manager). This is a structural pattern notification — not a verdict about any person.
Kashi does not detect harassment. It surfaces repeated communication asymmetries so a human review can decide what (if anything) to do about them.
Your voice, against your own baseline
Compared against your own past baseline, not the team average. This protects against false positives from introversion, cultural differences, or second-language speakers.
追加シグナル：言語指標と応答時間の遅延▾
The politeness register each colleague uses when speaking to you
1.0 = honorific (keigo) · 0.5 = neutral · 0.0 = plain form (タメ口). A drop of ≥0.30 below a colleague's register toward other teammates is review-worthy — the same person speaks to you differently than to peers.
- View your audit log — every time an executive, HR member, or reviewer queried your data, with their stated reason.
- Dispute a signal — if a transcript misidentified you, or a signal feels unfair, flag it. Your dispute is visible only to you and neutral reviewers.
- Request a neutral reviewer — escalate into the governed remediation lane (Lane B) without going through your manager.
- Opt out of the semantic lane— you can request that text-reading detectors are disabled for your transcripts. Structural detectors still run (they don't read text).
- Tanaka (CEO) · 4/19/2026 · Quarterly team-dynamics review. Reviewing sustained baseline-drop signal to determine whether to escalate to Lane B.
- Hayashi (HR) · 4/16/2026 · Routine data-quality check on interruption counts in product-team meetings (m-product-w10, w11, w12).
- System · automated · 4/7/2026 · Automated 4-tier retention purge: turns_metadata on meetings past 14-day horizon nulled.
組織が把握できること。
5名以上の
チームダイナミクス要約
Per-manager pattern intensity this week. No headline company-wide score. Individual subordinate data is not accessible — k ≥ 5 anonymized aggregates only, with ε = 1.0 Laplace noise on counts. Every drill-down is audited and visible to the affected employee.
出典を表示
METI 2025 benchmark — Estimated annual workforce-cost impact of unaddressed harassment patterns.
What the counts above protect
Any team with fewer than 5members is suppressed entirely from this view. The roster does not reveal which manager was suppressed — only how many. Without this, a 2- or 3-person team's signal would trivially identify individuals.
The three counts above are true counts plus calibrated noise. An adversary running repeated queries cannot reliably recover which manager moved between watch and concern this week, even with role-level access.
検出項目別の懸念パターン
This is the composition across the roster you see — not a company-wide score. Structural detectors (interruption, chilling, response-latency) run by default. Keigo (敬語) runs as a text-deterministic detector, gated behind the semantic lane flag with tenant opt-in.
Transcripts purge 14 days after upload. Analytics purge at 24 months. Review events purge at 12 months from last update. Legal-holds exempt specific rows and require justification.
2 drill-downs this week. Every drill-down writes an audit row visible to the affected employee at /app/me/audit. Zero silent surveillance.
Opening this roster itself is logged. Managers see when an executive views their summary.