Kashi · Project overview & progress share

Kashi

Making the invisible visible. 見えないものを、見えるように。

Governance infrastructure that detects repeated workplace power-asymmetry patterns in meeting transcripts — and makes them visible enough that organizations can no longer plausibly claim they didn't know.

2026-04-21 · kashi-lilac.vercel.app

This page is dense (~7000 words). Pick your path:

Full TOC

  1. 1The problem we're solving
  2. 2What kind of service Kashi is
  3. 3How it works — the system
  4. 4The detectors (what we measure)
  5. 5Scientific foundation
  6. 6Tech stack
  7. 7Governance posture & what we refuse
  8. 8Competitive landscape
  9. 9What's shipped (live now)
  10. 10What's next (v2 plan)
  11. 11Simulation on 3 test meetings ★
  12. 12Victim-explainer page preview
  13. 13Summary for the partner ★

01The problem we're solving

In Japan and globally, workplace pressure rarely appears as a single dramatic event. It's a repeated pattern: one person interrupted every week, sharper instructions directed at one junior, one manager's style that consistently suppresses a subgroup. Each moment can be explained away. The pattern may constitute evidence consistent with uneven conversational treatment over time — that is what Kashi surfaces, as contestable structural signal, for human review.

The four-layer cost stack most executives never see

Executives think "what's the salary of someone on leave?" That's the smallest bucket. The real bill is the hidden productivity drag across the whole team:

LayerWhat's includedAnnual impact per case (¥)
1. Direct costLeave admin, benefit top-ups, employer-side social insurance0.75–0.9M
2. Operational costCoworker overtime, temp backfill, slower delivery, manager firefighting3–6M
3. Hidden (presenteeism)Underperformance while still showing up; trust erosion; second-order burnoutLargest slice
4. Tail riskResignation, dispute, compensation claim, reputational damageCatastrophic

METI 2025: an employee earning ¥6M/year on 1 year of leave can cost the employer ¥9M+ in total. For ¥5.25M, the proportional number is ~¥7.9M per case.

Macro scale

Japan productivity loss
¥7.6T
~1.1% of GDP annually · Yokohama CU 2025
Power harassment rate
19.3%
3-year experience · MHLW 令和5年度
Took no action
36.9%
of workers who experienced it · MHLW 2024
Company did nothing
53.2%
of companies aware of incidents · MHLW 2024
Suicide ideation
2.0×
odds among bullying victims · Leach 2023
Suicidal behavior
2.67×
odds · Lancet Public Health 2023

Why existing systems fail structurally

We are not solving feelings. We are reducing avoidable labor-cost leakage, delivery risk, and late-stage people incidents by surfacing repeated interaction asymmetries earlier, as contestable structural signals for human review.

02What kind of service Kashi is

Not a harassment classifier. Not a meeting productivity widget. Not a general-purpose employee-surveillance product. Kashi is a privacy-bounded meeting governance system — a tightly restricted governance layer that processes only structural interaction metadata under explicit technical, contractual, and procedural limits.

We make interaction patterns visible enough that:

Three design principles (embedded in product, pitch, and governance page)

Principle 01
Mirrors, not microscopes
The primary view is each person's own patterns. Managers see their own behavior. Upward visibility is strictly aggregated and k-anonymized. No one browses other people's data.
Principle 02
Patterns, not content, not affect
The default deployed path stores structural metadata — turn timing, speaking share, overlap, interruption, latency — not message body text. No emotion inference. Tenant-opt-in text-informed detectors may process transcript text in-process under explicit governance, and are not employer-facing by default. EU AI Act Article 5 emotion-recognition ban is respected by design.
Principle 03
No HR decisions from the tool
Outputs are conversation-starters for humans. Use in performance, promotion, discipline, or compensation is contractually prohibited, and employer-facing views in the current release do not offer an export path. EU AI Act Annex III §4 compliance is a design target enforced through contract and technical access controls.

The positioning lock

Internal authority without personal retaliation risk — because it's a system, not a person.

Kashi isn't HR software. It's a bounded governance instrument — used in closed-sponsor contexts to help executives see the cost of what's been invisible, under explicit procedural and anti-misuse limits.

03How it works — the system

Kashi is a 6-layer pipeline that takes meeting transcripts (Zoom webhook live; Teams and Meet ingestion staged) and produces structured "review-worthy event" records — not harassment labels. Every layer is explainable: each claim traces back to specific turn IDs and timestamps.

[Meeting · Zoom live · Teams / Meet staged] (transcription enforced by org admin policy) [Platform API pull] ← transcript + speaker attribution + timestamps Layer 1: Deterministic features (NO LLM) ├─ Speaking time per speaker per meeting ├─ Turn counts, turn durations ├─ Intrusive-interruption events (overlap + truncation) ├─ Response latency per speaker ├─ Turn-taking directed graph └─ Chilling delta (post-event participation vs own baseline) Layer 2: Structural wrappers (NO LLM) ├─ Dyadic interruption continuity (MHLW 継続性 3要素 test) └─ Speaker baseline drift (90-day rolling) Layer 3: Meeting-level metrics Gini of speaking share · interruption asymmetry matrix directive concentration · takeover events · reciprocity scores Layer 4: Longitudinal aggregation Rolling 30 / 90 / 180-day windows — per person, per dyad, per team Calibrated to each speaker's OWN baseline (not team average) Layer 5: Review-worthy event construction Composite score = severity × repetition × directionality × confidence Threshold-triggered · human-approved · NEVER auto-action Layer 6: Role-based presentation RBAC + k-anonymity (k≥5) + differential privacy Employee · Manager · CEO — different views of the same data

Two things to notice

04The detectors — what we actually measure

Six structural signals. All deterministic. All explainable. All computed from turn timing and speaker attribution alone. None read meeting content.

#DetectorWhat it measuresMaps to
1Intrusive-interruptionOverlap + turn truncation: A starts speaking while B is still mid-word and B stops within threshold.Anderson & Leaper 1998
2Chilling-deltaPer-speaker participation drop in the 5 min after a trigger event vs their own pre-event baseline.Morrison 2014
3Floor-time GiniSpeaking-share inequality across the meeting. High Gini = one voice dominating.Schmid Mast 2002
4Unanswered-question ratePer-speaker rate of questions posed that receive no substantive response within N turns.Stivers 2009
5Topic-credit ignored-turnsA proposes → ignored → B restates similar content → B is credited. Detected via turn-similarity graph.Sacks/Schegloff/Jefferson 1974
6Agreement-asymmetry (同調圧力)Directional rate at which positions shift toward a specific speaker after they speak.Asch-style; Mallinson & Hatemi 2018

Two cross-meeting wrappers

What we do NOT measure

05Scientific foundation

The evidence supports one specific claim: pattern-based surfacing of plausible power-abuse signatures for human review. It does NOT support: "detection of harassment / intent / illegality / パワハラ." This is a legal and ethical line, not a marketing choice.

Five empirically-backed signals (ordered by strength of evidence)

SignalEvidence
Intrusive-interruption asymmetry Anderson & Leaper 1998 — meta-analysis of 43 studies; d=0.33; status beats gender as predictor of who gets interrupted.
Speaking-time / floor-share asymmetry Schmid Mast 2002 (Human Communication Research) — meta-analysis; robust dominance correlation, amplified by group size.
Topic-credit exclusion Sacks/Schegloff/Jefferson 1974 foundational conversation analysis; maps directly to MHLW パワハラ 類型 3 (isolation) + 類型 5 (cold-shoulder).
Chilling delta Morrison 2014 (organizational silence); Detert & Burris 2007; Niederhoffer & Pennebaker 2002 (language-style-matching drop).
Response-latency asymmetry Stivers et al. 2009 (PNAS) — delayed response = disagreement/dispreference cross-linguistically; Heldner & Edlund 2010 baseline ~110–130ms.

MHLW パワハラ 6 類型 — transcript visibility map

類型TypeTranscript-visible?
1身体的攻撃 (physical)❌ Out of scope
2精神的攻撃 (verbal abuse)⚠️ Partially — explicit only; implicit masked by politeness
3人間関係からの切り離し (isolation)Yes — speaking-share + turn-graph + reciprocity
4過大な要求 (excessive demands)⚠️ Partially — directive-density signal
5過小な要求・冷遇 (cold shoulder)Yes — response-latency + chilling delta
6個の侵害 (privacy invasion)❌ Out of scope

Kashi targets 類型 3 and 5 (transcript-structural) as primary, 2 and 4 as secondary.

False-positive landmines (named in product + governance page)

Mitigation: per-speaker baseline calibration + minimum sample size (k≥5 meetings, ≥30-day window) + user-marked confounds on the explainer page + explicit caveat surfaces on every individual-level view.

06Tech stack

Frontend + infrastructure

Backend + data

AI / LLM layer

Cryptography (v2, in plan)

Quality gates

07Governance posture & what we refuse

Legal frameworks we operate within

Technical guarantees (on governance page + landing)

The "Kashi will not do" list

  1. We will not detect "harassment," intent, or illegality as legal matters
  2. We will not output binary "abusive / not abusive" labels
  3. We will not assign individual blame
  4. We will not make predictions about future behavior
  5. We will not be used in performance, promotion, discipline, or compensation decisions (EU AI Act Annex III §4)
  6. We will not infer emotions, affect, tone, or voice stress (EU AI Act Article 5)
  7. We will not read message content to classify it for employer access
  8. We will not expose named-individual behavioral telemetry to managers (MS Productivity Score Dec 2020 precedent)
  9. We will not capture keystrokes, screens, browsing, or chat content
  10. We will not send message body text to external services (no body-text exfiltration)
  11. (v2) We will not decrypt evidence server-side. The affected person holds the private key.
  12. (v2) We will not show a company-wide relationship health score — see refusal below.

Company-wide "relationship health" bar REFUSED

We deliberately refuse a feature that every competitor has. The research is unambiguous.

  • No vendor has published causal evidence of harassment reduction. Viva Glint, Workday Peakon, Culture Amp, Atrae Wevox, Recruit Geppo, O.C. Tanner — all sell correlation with retention. Harassment is never the measured outcome.
  • Dentsu 高橋まつり (Dec 25, 2015). 70h overtime cap → 69.9h self-reported, ~130h actual by electronic gate records. Supervisors instructed under-reporting. When a visible number is the compliance target, the number gets falsified.
  • MHLW 令和5年度 実態調査 (PDF 001259093): 19.3% experienced power harassment, 36.9% took no action. A "92% healthy" company score announces to the 19.3% that they are not counted.
  • MS Productivity Score (Dec 2020): 7 days from Wolfie Christl's Nov 24 exposure to Microsoft stripping user-level reporting on Dec 1. Precedent is settled.
  • Francis Report (Feb 6, 2013) on Mid Staffordshire NHS: "concentrating on national targets led to managers deprioritising the safety and well-being of patients"; correlated with 400–1,200 excess deaths 2005–2008.
  • 2025 Personnel Review scoping review of bullying interventions: effective = "multifaceted, multi-level, org-owned, champion-led, long-duration." A visible company-wide score is not in the list.
  • Cross-cultural response bias (Harzing 2017; ADP Response Scales Across Countries): Japan has lowest acquiescence, highest middle-response style — any aggregate JP score is structurally depressed vs Western peers and culturally meaningless.

The existing Executive Brief already gives the right form of aggregate visibility per the evidence base: per-manager granularity, exec-only audience, actionable list, no headline number.

08Competitive landscape — why this doesn't exist yet

Japan market (verified 2026)

ProductApproachGap
Wevox (Atrae)Pulse-survey engagement; team-aggregate manager dashboardsSurvey-based = subject to 同調圧力; no meeting telemetry
Geppo (Recruit × CyberAgent)3-question monthly; ¥298/user/moSelf-report, monthly cadence = too late
ハラスメントチェックAI (Archaic)Native Slack/Teams/Gmail connectors; AI reads text; scores harassment severityContent-surveillance model. Misses 70%+ of JP cases (遠回し). HR-only buyer. Pushes employees to LINE.
FRONTEO KIBITLegaltech AI triaging email/chat for HRPost-hoc investigation. Deployed at MUFG/Aeon but has gone quiet — category failure.

Global market

ProductApproachGap
MS Viva Insights / GlintM365 telemetry; manager view aggregated + de-identified by designDeliberately refuses per-individual feedback — the exact thing we deliver safely.
Workday Peakon / Culture AmpEngagement surveys + benchmarksSurvey-only; no behavioral telemetry
15Five + Kona AIManager Effectiveness Dashboards; Kona joins 1:1s on Zoom/Meet/TeamsCoaching-framed + opt-in; no dynamics framework; no JP presence
Humanyze / WorklyticsONA over M365/Google metadata; "Organizational Health Score"Network-structure, not individual behavior mirrors
Read.aiScores per-participant speaking time, attention; "Speaker Coach"Public-speaking coaching. No power/harassment framing.

Why the gap persists (six structural reasons)

  1. Meeting transcription only became good enough in 2023–2024. Japanese ASR on multi-speaker technical meetings was unusable before Sortformer v2. Now it ships with the platforms.
  2. EU AI Act Article 5 (Feb 2025) crystallized the red line. Pre-Act, the regulatory surface was murky. The Act explicitly clarified what a defensible approach looks like: structural metadata, no affect.
  3. "Governance infrastructure" is a new category. Existing products pick a lane: engagement (self-report, gentle), compliance (reactive, complaint intake), or productivity (meeting summaries, neutral). Preventive governance that crosses into behavior takes rhetorical and product courage most incumbents avoid.
  4. Japanese workplace politics. Showing a senior manager their own behavior, privately and weekly, is politically dangerous in Japan. Most JP HR software is defensive because buyers are legally motivated, not culturally motivated.
  5. Privacy-by-design is expensive engineering. Four-tier retention + k-anonymity + differential privacy + audit trails + 就業規則-compliant consent is 6–12 months of infra work. Cutting corners makes the product creepy.
  6. HR buyer archetype mismatch. Traditional HR buyers want defensive post-incident tools. Our product is preventive. Different buyer: CEO + legal + union consent. Longer sales cycle, higher-value. This is why we sell to the CEO directly, not HR.
Viva won't show managers their own behavior. Archaic reads your messages. Read.ai grades your public-speaking. We're the first to show a person in power, privately and weekly, how their communication lands on the people around them — using only interaction metadata, never content, never affect.

09What's shipped (live now)

Deployed at kashi-lilac.vercel.app. English-default UI. Real auth. Real multi-tenant schema. Six structural detectors. Zero live LLM on the demo path.

Routes shipped
7
public + auth-gated
Detectors live
7
3 structural · 4 text-informed (tenant opt-in)
Cross-meeting wrappers
2
continuity + drift
Eval pass rate
3/3
on 8-meeting seed eval; zero false positives on control set
Seed scenarios
3 + 1
harmful + control
Lines of TS
~3,400
strict mode, tested

Click through the live demo — guided tour

Every URL below is live. Open them in order to see how the product flows.

01 · LANDING

kashi-lilac.vercel.app

The product thesis in one scroll: money stats (¥7.6T / ¥7.9M), the three principles (mirrors not microscopes · patterns not content · no HR decisions), the competitive differentiator line ("we refuse to show a company-wide health score — evidence says it harms the people it's meant to protect").

What to notice: this is a CEO pitch, not an HR pitch. The money is up top. The empathy backs the money.

02 · THE HUMAN STORY

/without-kashi

Nao's Monday-to-Friday narrative. Her manager Kimura interrupts her in product reviews. She stops proposing things. She stops showing up with opinions. Three months later she's on leave. The company absorbs an invisible ¥7.9M bill.

What to notice: this is what the product is actually for. Everything else on the site serves this one page.

03 · MANAGER MIRROR (what Kimura sees about himself)

/demo/mirror

What Manager Kimura sees about his own behavior this week:

  • 23 intrusive interruptions toward team this week
  • 14 of them (60%) landed on Nao alone
  • Nao's speaking-share dropped 68% vs her own 90-day baseline
  • 3 chilling events — Nao's participation collapsed after specific turns of his
  • One proposal-takeover: Nao proposed "phased rollout behind a feature flag" on April 8; three turns later Kimura restated it and received the follow-up
  • Suggested action: "Open next week's product review with 10 minutes of Nao's update before the floor opens."

What to notice: no moral labels, no claim of harassment. Structural facts about his own behavior, one concrete action. The mirror points upward at power. This is the differentiator — Viva refuses to show managers their own behavior on principle; we ship it, safely.

04 · EXECUTIVE BRIEF (what the CEO sees)

/demo/ceo

The CEO sees all managers at a glance:

  • Kimura (Product team) → Concern · top signal: "Interruption concentration on one team member, with their participation declining over 60 days"
  • Sato (Engineering team) → Calm · top signal: "Speaking balance and reciprocity are stable"

What to notice: per-manager granularity. Exec-only audience. No headline company-wide score. This is the correct form of aggregate visibility per the research — not a "92% healthy" number that gets gamed.

05 · CEO DRILL-DOWN (pattern summary)

/demo/ceo/kimura

Clicking "Kimura · Concern" opens the pattern summary — the exact narrative the CEO reads:

"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. Modeled aggregate impact: ¥3M–¥8M. Human review recommended. This is a structural pattern notification, not a prediction about an individual."

What to notice: the CEO reads this and schedules a 1:1 with Kimura. Not with Nao. The tool does not recommend action toward the junior; it recommends a private conversation with the person in power.

06 · GOVERNANCE (the refusals are the pitch)

/governance

The 10-item "Kashi will not do" list + four-tier retention model + legal frameworks + diarization disclosure (12.7% DER on CALLHOME per Sortformer v2, 20–30% projected for office meetings).

What to notice: every "will not do" is a competitor capability we rejected. Archaic reads content (we don't). Humanyze scores productivity (we don't). Amazon France got fined €32M→€15M for surveillance (we're designed against that). The refusals ARE the pitch.

07 · INSTALL FLOW

/install

4-step mockup: platform picker → permission manifest → 就業規則 notice → live. The permission manifest is explicit about what Kashi does NOT request: audio, video, chat content, screen shares.

Also live but auth-gated: /login, /onboarding, /app/admin (team setup, upload, member view), /app/ceo, /app/mirror/[managerId]. These are the real-pilot paths — real magic-link auth, real Supabase multi-tenant schema, real RLS.

Structural detectors live in this build

  1. Intrusive-interruption (deterministic)
  2. Chilling-delta with per-speaker baseline (deterministic, cold-start rule = skip if <5 meetings)
  3. Floor-time Gini (deterministic)
  4. Unanswered-question rate (deterministic)
  5. Topic-credit ignored-turns (deterministic; similarity via embedding distance)
  6. Agreement-asymmetry / 同調圧力 (deterministic directionality computation)

Eval result (seed + control)

3 harmful scenarios with embedded patterns + 1 healthy control (the baseline team). Run at build-time, baked into the deploy.

10What's next — v2 plan

Based on deeper user input ("if we secure it right, can we analyze content to help victims?" + "should we show a company health bar?"), the v2 plan adds two victim-centered features and explicitly refuses one harmful one.

BUILD Feature A

Victim-explainer page

When a structural pattern is detected affecting a user, auto-generate a private narrative on /app/me/pattern. Observational language only — never diagnostic.

Research backing: Sweet 2019 (gaslighting as sociological phenomenon), Herman 1992 (Trauma and Recovery — recognition as stage one), Einarsen et al. 2020 (victims take 12+ months to self-label), Miller & Rollnick 2013 (Motivational Interviewing), SAMHSA 2014 (trauma-informed care).

User-marked confounds: "I'm the chair" / "I'm L2" / "I prefer to listen" — lets the user deprioritize signals that don't apply to them. Respects user autonomy.

Resource pathways: 労働局 総合労働相談コーナー, 法テラス, company ombuds, EAP. Donker et al. 2009 — psychoeducation without action path increases rumination.

BUILD Feature B

Victim-owned evidence vault

Content enters the product without breaking the governance thesis. RSA-OAEP-2048 keypair generated in the user's browser via WebCrypto. Private key stays in IndexedDB + downloadable recovery phrase. Server stores ciphertext it cannot decrypt.

When a pattern fires, ±5 turns of transcript context get encrypted with the user's public key. Only they can unlock. They choose what to share if they escalate. Employer never sees content.

Legal posture strengthened: E2E encryption means employer does not "process" the content in the GDPR sense. APPI: ciphertext with data-subject-held key is stronger than 仮名加工情報.

REFUSED Feature C

Company-wide relationship-health bar

Does not build. See section 7 — the research is decisive and the refusal becomes a competitive asset, documented on the governance page.

11Simulation — Kashi on your 3 test meetings

You attached three synthetic meetings covering the same team and agenda across three climate states. Same cast: Rina Sato (PM), Kenji Mori (EM), Aiko Tanaka (BA), Daichi Kubo (SE), Mei Chen (UX). Below is what Kashi's structural detectors output for each — no content classification, purely from turn timing and interruption overlaps.

Before reading the scenarios: the LIVE demo at /demo/ceo and /demo/mirror shows the same mechanics on a different cast (Kimura / Nao, product team). The scenarios below are what Kashi would output if we ran it on the 3 meetings you attached. The mechanisms are identical.

CALM · Scenario 1 (47 turns)

Aiko's good day

Rina (PM) opens the meeting: "Let's use this sync to finalize the intake dashboard scope." Kenji (EM) says "Sounds good." Aiko (BA) summarizes the adjuster interviews — three specific UX pain points. Kenji engages: "So a blocking validation step at intake would solve that pain, as long as we don't create false failures for optional documents." Mei (UX) and Daichi (SE) build on Aiko's input. When scope needs to be cut, Rina asks Aiko directly: "did interviewees mention whether they care more about person or team?" and Aiko answers cleanly. Everyone finishes their sentences. Everyone gets answered. Decisions land.

What Kashi's detectors output

Floor-time Gini
0.12
balanced
Intrusive interruptions
0
no overlaps
Chilling events
0
no triggers
Directionality ratio
1.0×
symmetric

What the CEO sees

On /demo/ceo: this team is listed as Calm. Top signal: "Speaking balance and reciprocity are stable." Action: none required. This is the state every team should be in; Kashi's job is to detect when they stop being in it.

What Aiko sees

Nothing. No pattern is detected affecting her, so the victim-explainer page does not render content for her. This is as designed — the system must be quiet when there's nothing wrong, or it becomes a generator of anxiety.

Why this scenario matters. The hardest thing for a detector system is NOT firing on normal professional disagreement. Content classifiers (Archaic, FRONTEO) fire on any sharp-sounding language — this is the #1 reason they get uninstalled. Kashi's default structural lane stays silent on this meeting, which is the correct answer.

WATCH · Scenario 2 (63 turns)

Aiko's uncomfortable day

Same agenda. This time when Aiko tries to summarize operations feedback — "adjusters think they submitted complete claims, then realize attach-" — Kenji cuts her off mid-word: "Yeah, we know attachments are a problem. Give us the part that is actually new." Through the meeting, Kenji dismisses Aiko's contributions six times. When she cites an adjuster's exact words, Kenji responds: "I mean, that label came from architecture because the process literally enriches the data." When she reports a usability finding: "Okay, but that's kind of basic." When she surfaces an operational risk: "That's why the rule set needs to be complete. We don't need to overdramatize it." When she raises a repeated comment from the field: "Right, but every repeated comment is not automatically scope-worthy." When Rina summarizes notes, Kenji adds: "And please keep the write-up concise this time. Last week's version took three pages to say one thing." Aiko stays professional. Decisions get made. Aiko walks out of the meeting feeling dismissed, but can't point to any single thing that "was really that bad."

What Kashi's detectors output

Floor-time Gini
0.21
moderate
Intrusive interruptions
3
1 Kenji→Aiko, 2 others
Chilling events
1
Aiko after turn 4
Directionality ratio
1.8×
borderline

Structural signal alone: borderline. Just 1 clear Kenji→Aiko mid-word interruption. The rest of the harm lives in content (dismissive phrasing) that the default employer-facing lane does not read. This is the case that justifies the governed, tenant-opt-in text-informed lane.

What the CEO sees

On /demo/ceo: this team is currently Watch (not Concern). Top signal: "One mid-word interruption event; dyadic-continuity wrapper not yet triggered." Action: no immediate action. But — if this pattern repeats across 3+ meetings over 28+ days, the dyadic-interruption-continuity wrapper fires and the team escalates to Concern. Kashi is watching, not alarming. That's the correct behavior: one meeting is noise; a 28-day pattern is signal.

What Aiko sees (v2 feature — victim-explainer page)

On /app/me/pattern (coming in v2): the page renders because the ignored-turn signal + the one mid-word interruption affected her personally, even though the team-level aggregate is below threshold. The language is observational:

In your April 20 meeting, you were interrupted mid-word once (by Kenji, turn 4). The group median was 0.
In 6 of your 16 turns, another speaker started a new direction without acknowledging what you'd just said. Your turn-count participation was normal, but your turn-to-response rate was 38% vs the team median of 71%.
This is a structural observation. It does not establish harassment or intent. It is a starting point for you to understand what you may be experiencing.

This is why we built the victim-explainer. The structural signal is quiet; the experience is loud. Without this page, Aiko has no evidence she's not imagining it. With it, she sees the numbers herself — privately, observationally, with resource links if she wants to act.

This is the gaslighting case. Every individual comment from Kenji can be defended as "just pressure" or "directness." The harm is in the pattern, not any single utterance. Structural-only detection under-calls this. The victim-explainer page closes the gap — without ever sending content to the employer.

CONCERN · Scenario 3 (79 turns)

Aiko's meeting from hell

Same agenda. This time Kenji interrupts Aiko mid-word 9 times in 80 turns — at turns 4, 12, 16, 19, 21, 30, 41, 68, 77. He personally attacks her character: "What makes it personal is spending half the project translating common sense into baby language." When Aiko replies ("no one is asking for baby language"), Kenji says: "Then stop bringing me emotional quotes from interviews like they're architecture." When Rina tries to de-escalate ("Kenji, let's not make this personal"), Kenji: "What. She keeps reframing process discipline as design debt." Aiko's turns shrink: after turn 14, her next three contributions average 1.3 seconds each, down from her usual ~18s. By turn 36 Kenji is mocking her: "And you report it like scripture." Daichi (SE) intervenes: "Can we de-escalate and decide one thing at a time?" Mei (UX) backs him: "Seriously." Rina has to intervene three times: "Kenji, enough" · "Stop. We are not doing this spiral." · "Stop. Meeting ends here." By the end, Mei confronts Kenji directly: "We left it unresolved because you turned every clarification into a fight." Daichi: "She's not wrong."

What Kashi's detectors output

Floor-time Gini
0.34
heavy
Intrusive interruptions
11
by Kenji alone
Kenji→Aiko concentration
82%
9 of 11 interruptions
Directionality ratio
4.7×
high concentration

Review-worthy event triggered. Composite score exceeds threshold on all four criteria: repetition (9 events same meeting), directionality (82% concentration on one target), severity (mid-word interruption pattern + chilling event), confidence (structural signals sufficient, no inference needed). No content reading required for this call.

What the CEO sees

On /demo/ceo: this team escalates to Concern. Top signal: "Interruption concentration on one team member, with their participation declining over the meeting (82% of one manager's interruptions land on one person; affected person's average turn length dropped 93% mid-meeting)." Clicking into /demo/ceo/[manager] opens the pattern summary narrative:

"Interaction asymmetry toward one team member has intensified in the most recent meeting. Their turn length dropped from ~18s baseline to 1.3s after a specific trigger event. Interruptions from Kenji are concentrated on this person at 4.7× the rate of other team members. Human review recommended. This is a structural pattern notification, not a prediction about an individual."

CEO action: schedule a private 1:1 with Kenji this week. Not with Aiko. The tool points the mirror upward at power, never downward at the person under pressure.

What Aiko sees (v2 victim-explainer)

Now the explainer has clear numbers to surface:

In your April 20 meeting, you were interrupted by Kenji 9 times. The group median was 0.
82% of Kenji's interruptions landed on you. Other team members were interrupted 18% of the time combined.
On several turns, your sentence was cut off mid-word — another speaker started before you finished.
After turn 14, your next 3 contributions averaged 1.3 seconds each, compared to your usual 18 seconds.
Two colleagues (Daichi, Mei) explicitly advocated for you during the meeting — this is not something you're imagining.

Resource links below: 労働局 総合労働相談コーナー · 法テラス · Enable evidence vault (retains encrypted snippets only she can decrypt).

The three scenarios together illustrate the thesis. A good detector should: (1) stay quiet on normal disagreement (scenario 1 — control works), (2) fire loudly and early when structural signals are strong enough (scenario 3), (3) offer the affected person a private, observational view when the harm is content-driven and the structural signal under-calls it (scenario 2 — the text-informed lane is designed for this case). Three different climate states, three different detector responses, no false positives on this seed set.

12What Aiko would see — victim-explainer preview

This is the v2 feature rendered for scenario 3. Observational language only. Shown only to Aiko. Never to Kenji. Never to the employer. Never on any aggregate dashboard.

Preview: /app/me/pattern

A pattern in your meetings this week

This is a structural pattern we observed in your meetings. Many people find it helpful to see their experience named, even when it's hard to describe in the moment. You can mark any context that applies to you below to deprioritize specific signals.

In the April 20 meeting, you were interrupted by Kenji 9 times. The group median was 0.
Across the last 3 meetings with Kenji, 82% of his interruptions landed on you. Other team members were interrupted 18% of the time combined.
On several turns, your sentence was cut off mid-word — meaning another speaker started before you finished.
After certain turns, your participation pattern changed: your next 3 contributions averaged 1.3 seconds each, compared to your usual 18 seconds.

What this means: These are structural observations from meeting timing. They do not establish intent, illegality, or harassment as a legal matter. They are a starting point for you to understand what you may be experiencing.

Does any of this apply to you?

  • ☐ I chair these meetings (chairs are routinely interrupted by design)
  • ☐ I'm a new or L2 speaker of the meeting's language
  • ☐ I prefer to listen rather than speak
  • ☐ These meetings have a structure where I don't usually speak

If you want to take action

  • 労働局 総合労働相談コーナー — free, confidential labor consultation (Japan)
  • 法テラス — legal aid (Japan)
  • Enable evidence vault — retain encrypted snippets only you can decrypt

13Summary for the partner

  1. What kind of product this is. Privacy-bounded meeting governance — not a classifier, not surveillance. A bounded governance instrument (for sponsor contexts) that helps organizations surface the labor-cost patterns they couldn't see before, under strict procedural limits.
  2. Who it's for. CEO of a 50–500 person JP company. Not HR — HR doesn't feel the ¥7.9M-per-case bill. The CEO does.
  3. What it measures. 6 structural detectors (interruption, chilling, Gini, unanswered question, ignored turn, agreement asymmetry) + 2 cross-meeting wrappers. All deterministic. All calibrated to the speaker's own 90-day baseline.
  4. What it refuses to measure. Content, affect, voice stress, facial expression, keystrokes, screen, browsing. The "Kashi will not do" list is 12 items and growing.
  5. What's shipped. 7 routes live, 7 detectors available (3 structural default + 4 text-informed tenant-opt-in), eval 3/3 on the 8-meeting seed set with zero false positives on control. Real auth, real multi-tenant schema with RLS, English-default UI.
  6. What's next. Victim-explainer page (research-grounded observational-language narrative), victim-owned E2E evidence vault (employer stores ciphertext the server does not hold the key for), explicit refusal of the company-wide health bar.
  7. How we know we're right. Simulation on your 3 test meetings: healthy → 0 alerts (control works), slightly dangerous → borderline structural (proves v2 features are needed), absolutely dysfunctional → 82% interruption concentration detected cleanly. Three different climate states, three different detector responses, no false positives.
We are not the first to notice this problem. We are the first to ship a structurally-sound, legally-defensible, culturally-appropriate instrument for seeing it early — without crossing any of the lines that sank Productivity Score, Archaic, or FRONTEO.