Kashi 可視

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:

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

Why existing systems fail structurally

• Burden of recognition sits on the person under pressure
• Burden of proof sits on the same person
• Organizations see fragments, not patterns
• Managers get feedback only after serious damage
• HR relies on lagging indicators — complaints, attrition, leave, 自殺
• Buying-pathway mismatch — existing tools sell to HR, who don't feel the productivity-loss bill. The CEO who bears the ¥7.9M-per-case bill isn't in the sales conversation.

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:

• Employees can validate whether a pattern is real, not gaslighting
• Managers see their own behavior as data, not moral judgment
• CEOs see early signals with structured evidence, not late-stage crisis
• Organizations can no longer plausibly claim they didn't know

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

The positioning lock

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.

Two things to notice

• Layers 1 and 2 are deterministic — no LLM, and their structural detectors work from turn metadata, not content. Each alert is designed to be defensible under EU AI Act scrutiny because it reduces to a counted event from timestamps.
• Per-speaker baseline calibration is central. We compare you to your own 90-day rolling baseline, not to the team average. This defeats the introversion / L2 / chair-role false-positive problem that sinks most engagement tools.

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.

Two cross-meeting wrappers

• Dyadic interruption continuity — maps to MHLW 継続性 3要素 test (was it repeated? over time? toward the same person?). A single meeting is noise; a 90-day pattern is signal.
• Speaker baseline drift — tracks each person's 90-day rolling baseline to detect trajectory changes (speaking share dropped 40% over 60 days vs their own prior).

What we do NOT measure

• ❌ Tone of voice, voice stress, prosody
• ❌ Facial expressions, body language
• ❌ Sentiment, emotion, affect (EU AI Act Article 5 red line)
• ❌ Content-level harassment classification (the Archaic / FRONTEO failure mode)
• ❌ Employee Productivity Score-style metrics (MS Dec 2020 precedent)
• ❌ Keystroke logging, screen capture, browser history

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)

MHLW パワハラ 6 類型 — transcript visibility map

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

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

• Introversion — Spark et al. 2020: ~37% of introverts report mistreatment; speaking-share asymmetry can reflect personality
• Japanese meeting norms — ~5.15s silence/min vs ~0.74s US; longer gaps are cultural
• Neurodivergence (ADHD / ASD) — processing differences affect turn-taking
• L2 speakers — slower turns, longer gaps, reduced floor time
• Legitimate chair role — PMs / facilitators have higher directive density by design

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

• Next.js 16.2.4 (App Router, Turbopack) + React 19 + TypeScript
• Tailwind v4 with @theme inline
• shadcn/ui for components
• Recharts for visualizations (BarChart with LabelList, LineChart with ReferenceLine)
• Inter (Latin) as default typeface; Japanese layer uses Noto Sans JP
• Vercel for deploy; preview URLs per branch, prod on kashi-lilac.vercel.app

Backend + data

• Supabase for Postgres + Auth + Row-Level Security + Storage
• Magic-link OTP auth via @supabase/ssr
• Multi-tenant schema with org-level RLS; cross-org leakage blocked at the database layer
• Two migrations applied: 0001_init.sql (schema), 0002_rls_hardening.sql (policy replacements)
• VTT / TXT / SRT / CSV / JSONL transcript parsers
• All detector outputs pre-baked at build time — demo runtime is 100% static render; no live LLM calls on the demo path

AI / LLM layer

• Claude API — claude-sonnet-4-6 for seed authoring and structured classification; claude-opus-4-7 for reasoning-heavy tasks
• Prompt caching enabled — transcripts as stable prefix, classifiers as separate cached calls
• JSON-schema-constrained output only — temperature 0, stable across runs
• Claude is used to author seed scenarios and test transcripts. Claude is NOT used in live detection on the production path. Detectors are deterministic TypeScript.

Cryptography (v2, in plan)

• WebCrypto API for client-side RSA-OAEP-2048 keypair generation
• AES-256-GCM for evidence snippet encryption, wrapped with user's RSA public key
• IndexedDB for private key storage; BIP39-style recovery phrase for backup
• Server is designed not to hold the decryption key — Kashi stores ciphertext that the company cannot read from the server alone (break-glass procedures are documented separately)

Quality gates

• TypeScript strict mode (tsc --noEmit passes with zero errors)
• Unit tests on all Layer-1 detectors with synthetic turn data
• Eval harness: 3 seed scenarios with embedded ground-truth patterns + 1 healthy control. Current pass rate: 3/3 detected, 0 false positives on control.
• Determinism check: running the same input twice produces identical outputs
• Explainability check: every review-worthy event traces back to specific turn IDs and rule triggers

07Governance posture & what we refuse

Legal frameworks we operate within

• 労働施策総合推進法 §30-2 (パワハラ防止法) — effective 2020 (large employers) / 2022 (SMEs). Obligates policy, consultation channel, swift response, complainant-privacy protection, anti-retaliation.
• APPI — pattern metadata is personal information when linkable. We use 仮名加工情報 + public-notice + 就業規則-revision consent basis (not individual opt-in, which is culturally fragile in Japan).
• EU AI Act Article 5(1)(f) — in force Feb 2025. Bans AI systems that infer emotions in the workplace except for medical/safety. Our no-affect stance is a legal red line.
• EU AI Act Annex III §4 — HR tools used for performance / promotion / termination become high-risk from Aug 2026. "No HR decisions from the tool" is now a legal constraint, not just ethics.
• GDPR Article 88 — employee consent generally not valid basis (EDPB 05/2020); works-council consultation advised.

Technical guarantees (on governance page + landing)

• k-anonymity (k≥5) on every aggregate view
• Differential privacy (ε≤1) on executive dashboards
• Auto-suppression when team size <5 or when a single role dominates
• Audit trail on all drill-down actions; viewable by the affected individual
• Four-tier retention: Raw (14 days, QA-only) · Analytics (24 months, role-based) · Review-worthy events (12 months) · Legal-hold (extended, justified)
• Five RBAC roles: Individual · Manager · HR-Compliance · Restricted Investigator · System Admin

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.

08Competitive landscape — why this doesn't exist yet

Japan market (verified 2026)

Global market

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.

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.

Click through the live demo — guided tour

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

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.

• ✅ Harmful team: persistence score 0.45 (toward Nao), sustained-baseline-drop 63 days, 14 Kenji→Nao interruptions vs 9 to all others — pattern detected
• ✅ Healthy team (Sato): zero review-worthy events, Gini 0.11, directionality 1.0× — no false positive
• ✅ Determinism: same input → same output on repeat runs

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.

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.

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.

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.