Kashi 可視 — technology

01Executive summary · investor read

Kashi is a deterministic meeting-governance pipeline. It takes meeting transcripts (Zoom webhook live; Teams and Meet ingestion staged), runs them through a stack of structural detectors, and surfaces repeated interaction asymmetries for human review — not as harassment labels, but as contestable signals with explicit uncertainty.

The one-sentence architecture

A Next.js 16 + TypeScript + Supabase application with a pipeline of deterministic detectors organized into 3 lanes (structural-only · hybrid text-informed · refused), protected by row-level security, with commitments declared in typed enums and a machine-readable detector registry — and several honest gaps we're naming on the way to pilot-ready.

02The stack

03The detector pipeline

Every transcript flows through the same 6-layer pipeline. Every layer is deterministic. Every layer's output is typed. No hidden model calls, no black-box scoring.

What it means in plain English

• Layer 1–2: the raw signal extraction. Counts, durations, overlaps, who-addressed-whom.
• Layer 3: per-meeting rollup. "In this meeting Gini was 0.34; Kenji's interruptions landed on Nao 82% of the time."
• Layer 4: "is this one bad meeting or a 63-day pattern?" — the longitudinal wrapper.
• Layer 5: the composite that decides whether a pattern is review-worthy, using thresholds (directionality ≥ 3×, baseline drop ≥ 0.4).
• Layer 6: three different views rendered from the same event — private to the employee, private-self-mirror to the manager, k≥5-anonymized to the executive.

04The 7 detectors, one lane at a time

Each detector declares its lane at compile time in src/lib/pipeline/detector-registry.ts. Employer-facing surfaces default to lane 1 (structural-only). Lane 2 requires tenant opt-in via semantic_lane_enabled feature flag.

Lane 1 — structural-only (3 detectors, default employer-facing)

1. Intrusive-interruption

Detects when speaker B starts while speaker A is still speaking AND A's turn ends within a threshold.

// src/lib/pipeline/layer1-deterministic.ts
const OVERLAP_THRESHOLD_MS = 500;      // overlap window
const TRUNCATION_WINDOW_MS = 500;   // A's turn must end within this

export function detectIntrusiveInterruptions(turns: Turn[]): IntrusiveInterruption[] {
  // O(n) scan: for each adjacent pair (B, A+1), check if B.startMs < A.endMs
  // AND A.endMs - B.startMs < OVERLAP_THRESHOLD_MS
  // AND A's turn ends within TRUNCATION_WINDOW_MS → count it
}

Research anchor: Anderson & Leaper 1998 (meta-analysis, 43 studies, d=0.33).

Known confounds: facilitator/chair role · incident-bridge meeting type · overlap-heavy audio. Handled by caveat surface on the Mirror UI; should be handled by meetingType gating (§7 gap).

2. Chilling-delta

Per-speaker participation drop in a 5-minute window after a trigger turn, compared to that speaker's own rolling baseline (not team avg).

const CHILLING_DELTA_WINDOW_MS = 300000;  // 5 min
const CHILLING_DELTA_THRESHOLD = 0.4;    // 40% drop
const COLD_START_MIN_MEETINGS = 5;

Research anchor: Morrison 2014 (organizational silence); Detert & Burris 2007.

Cold-start rule: no baseline signal emitted if the speaker has fewer than 5 prior meetings. Today this rule lives in perSpeakerBaseline() at layer1-deterministic.ts:118 — it returns null when comparable meetings < 5. See §7 gap: should be lifted to global abstention doctrine.

3. Floor-time Gini

Standard Gini coefficient over speaker durations within a meeting. Gini 0 = perfectly equal, 1 = one person dominates.

Research anchor: Schmid Mast 2002 (Human Communication Research meta-analysis).

Lane 2 — hybrid text-informed (4 detectors, tenant opt-in required)

4. Unanswered-question rate

A turn ending in lexical question markers (「か？」/「?」/「かな」) that receives no substantive response within N turns. Short acknowledgment responses (<2s or lexical "yes/なるほど/そうですね") don't count as a response for this detector.

Research anchor: Stivers et al. 2009 (PNAS cross-linguistic turn-taking).

5. Topic-credit ignored-turns

A substantive turn (>3s) from speaker A, followed either by (a) silence ≥ 2× the meeting's median inter-turn gap, or (b) a topic redirect by speaker B that captures credit for a similar proposal later in the meeting. Similarity is computed via embedding distance.

Research anchor: Sacks / Schegloff / Jefferson 1974; maps to MHLW パワハラ 類型 3 + 5.

6. Agreement asymmetry (同調圧力)

Directional rate at which position-statements shift toward a specific speaker. Position-statements detected via lexical cues; shift detected by comparing turn-order positional claims.

7. Keigo (敬語) peer-addressee asymmetry

Per-speaker-per-addressee politeness-register score from surface grammar. Classifies each turn into sonkeigo / kenjougo / teineigo / plain / imperative / mixed. Detects when the same speaker uses meaningfully lower register toward one addressee than toward peers.

// src/lib/pipeline/keigo.ts — excerpt
const SONKEIGO_PATTERNS = [/いらっしゃ/, /おっしゃ/, /なさる/, /お[一-龯]+になる/];
const PLAIN_PATTERNS = [/だよ/, /だね/, /でしょ/, /よね/, /じゃん/];

const PEER_ASYMMETRY_THRESHOLD = 0.25;
const MIN_TURNS_PER_TARGET = 2;

On our Kimura/Nao seed: Kimura scores 0.88 toward Nakamura (honorific) vs 0.38 toward Nao (plain form). Gap = 0.50. The largest gap in our control team (Sato's): 0.0 — zero asymmetries.

Research anchor: Cook 2011 (J. Pragmatics 43(15)); Saito 2011 (J. Pragmatics 43(6)); Pizziconi 2003; Ide wakimae framework. Unique to Kashi. No Western product has this.

Cross-meeting wrappers (in continuity.ts)

• Dyadic interruption continuity — MHLW 継続性 3要素 test: is an (interrupter, interruptee) pair appearing across ≥3 meetings? Emits a persistence score (0-1).
• Speaker baseline drift — detects the longest sustained drop window where a speaker is below their own 90-day baseline. Example seed output: Nao down 68% for 63 continuous days.

Aggregation thresholds (live in aggregate.ts)

const WINDOW_DAYS = 90;
const CONCERN_DIRECTIONALITY = 3.0;  // 3× peer rate → concern
const CONCERN_BASELINE_DROP = 0.4;   // 40% drop → concern
const WATCH_DIRECTIONALITY = 2.0;
const WATCH_BASELINE_DROP = 0.2;
const COST_PER_CASE_MIN_YEN = 3_000_000;
const COST_PER_CASE_MAX_YEN = 8_000_000;

Known simplification: These thresholds are universal. The meeting-type-normalization research says they should be per-meeting-type. Fix in the §7 gap list.

05Data model

Core types at src/lib/types.ts:

// The atomic unit of input.
export type Turn = {
  speakerId: string;
  startMs: number;
  endMs: number;
  text: string;
};

export type Meeting = {
  id: string;
  dateIso: string;
  teamId: string;
  title: string;
  turns: Turn[];
  // Declared (not yet threaded through detectors):
  meetingType?: MeetingType;
  meetingTypeConfidence?: number;
  scoringMode?: ScoringMode;
};

// New types committed 2026-04-21 · awaiting end-to-end threading:
export type DetectorClass =
  | "STRUCTURAL_ONLY"
  | "TEXT_DERIVED_DETERMINISTIC"
  | "HYBRID_TEXT_INFORMED"
  | "GENERATIVE_ASSIST"
  | "REFUSED";

export type EvidenceGrade =
  | "BLOCKED"           // input quality below gate
  | "INSUFFICIENT"      // thin exposure
  | "WEAK"
  | "EMERGING"
  | "STABLE"
  | "HIGH_CONFIDENCE_STABLE";

export type AbstentionState =
  | "NO_COMPUTE"
  | "COMPUTE_NO_INTERPRETATION"
  | "WATCH_ONLY"
  | "INTERPRETABLE_PRIVATE_ONLY"
  | "INTERPRETABLE_ROLE_BOUNDED";

export type ConfidenceBundle = {
  inputQuality: number;
  contextSupport: number;
  exposureSupport: number;
  detectorConfidence: number;
  aggregationSupport: number;
  reasonCodes: ReasonCode[];
  abstention: AbstentionState;
  grade: EvidenceGrade;
};

Database schemas at src/lib/db/types.ts — multi-tenant with org_id scoped RLS on every table. Raw transcript text is stripped before DB write; only length_chars is persisted.

Supabase schema (abbreviated)

orgs(id, name, created_at)
profiles(id, org_id, email, role, created_at)
  -- role enum: admin | ceo | member (see §7 gap: needs 6)
teams(id, org_id, name)
meetings(id, org_id, team_id, title, date_iso, turns_metadata jsonb)
  -- turns_metadata = TurnMetadata[] with text STRIPPED
meeting_metrics(meeting_id, speaking_share jsonb, intrusive_interruptions jsonb, ...)
manager_mirrors(manager_profile_id, week_ending_iso, ...)
pattern_summaries(manager_profile_id, pattern_intensity enum, ...)
user_keys(user_id, public_key_jwk jsonb)  -- evidence vault (planned)
evidence_vault(id, user_id, event_id, encrypted_snippet, encrypted_data_key, iv)

Migrations applied

• 0001_init.sql — schema + initial RLS policies
• 0002_rls_hardening.sql — replaced v1 policies with isolation-verified v2 versions
• 0003_indexes_and_metrics.sql — query-performance indexes + metric materialization

06Security posture (honest snapshot)

What we've actually built vs what we've only declared. Naming the gap is the first step to closing it.

07Gap analysis — where we can improve

This section is the honest answer to "what's not yet done?" All items below are surfaced directly from the Ideas_wave3 technical-dev research library (17 memos) + the business research memos' technical implications. Every item has a path to code.

P0 Critical — pilot-blockers

P1 High priority — next sprint

P2 Medium — this quarter

Strategic 6-12 month horizon

07·5Critical review — if you can only ship 3 things before pilot

§7 lists 23 improvements ranked by priority. But priority lists ≠ sequencing decisions. This section re-ranks the §7 items by pitch impact × pilot necessity × feasibility with current team size, and picks the three that should be Sprint-1 scope after pre-seed funding.

What the gap analysis missed (v3 addition)

• Webhook ingestion path for Zoom/Teams/Meet. Today Kashi runs on uploaded transcripts. For production, platforms push transcripts via webhook. Missing from §7 because it's plumbing, not product — but it's the integration gate for any real pilot. Add as P1 item 24.
• Per-tenant feature-flag admin UI. The detector-registry defines flags (semantic_lane_enabled, etc.). No admin surface to flip them per tenant. A pilot customer can't opt into lane-2 detectors because there's no way to say yes. Add as P1 item 25.
• Cost-of-detector telemetry. We don't instrument how long each detector takes per meeting. At scale that matters for the pricing model defensibility. Add as P2 item 26.

08Roadmap by sprint

Tied to the financing plan (business.html §6).

Velocity assumption

Sprints 1-2 deliverable on the 2-person founding team. Sprint 3+ assumes 1 additional engineer (the security/DevSecOps hire from business.html §5). Sprint 5+ assumes 2 additional engineers + CS lead from the seed round.

09What we will NOT build

The companion to governance's "Kashi will not do" list, at the technology level. These are architectural refusals — not "we haven't gotten around to it yet," but "we will actively not ship this."

• Audio / prosody / voice-stress analysis. EU AI Act Art. 5(1)(f) + Commission Guidelines C(2025) 884 final. Plus the science is chance-level (Damphousse 2008, Eriksson & Lacerda 2007).
• Emotion / affect / sentiment inference. Same Art. 5 prohibition. Full stop.
• Generative summaries of individual behavior. No "Kashi thinks X about Y." The tool describes what happened; humans interpret.
• Company-wide health score. Refused on evidence. See governance page §11 and business plan §7.
• HR-decision export path. No API, no data format, no workflow feeds performance / promotion / discipline / compensation systems.
• Keystroke, screen, or browser monitoring. Out of scope. Always.
• Real-time in-meeting intervention. No "please stop interrupting" popups. Kashi is asynchronous, weekly, private.
• Identity-based classifications. No gender / race / age / religion inference. The detectors are about structure, not demographics.
• Pulse-survey features. Not our category. Wevox, Peakon, Culture Amp cover this and have failed to move harassment outcomes. Adding a survey tab would dilute our positioning.
• A "score your manager" public leaderboard. Goodhart's law predicts this fails in 6 months. Not built.