You are about to watch Kashi process a single meeting transcript end to end. Real data. Real signals. No mocks. Every number on the following frames is computed from the transcript by the same code running on the live site.
This is what replaces a generic “demo video.” Drive it forward with the ↓ key or Space. Jump to any frame with 1–6.
m-product-w12 · 2026-04-15 · Product team weekly review · 5 participants · 89 turns over 47 minutes
Kashi reads only who-spoke-when + surface grammar. Never the content's meaning. Never tone. Never voice.
Deterministic. No LLM. No content analysis. Just turn timing and speaker attribution. Every number below traces to a specific timestamp in the transcript above.
It does not classify the words as hostile, dismissive, or harassing. It does not infer Kimura's intent. It does not label Nao as a victim. It only counts structural events from timestamps. The interpretation happens at Layer 5, combined with persistence and directionality.
Kimura classifies every addressee by surface grammar — verb endings, nominal prefixes, honorific substitutions. In this single meeting, his politeness register toward each teammate looks like this:
What it maps to: in 労働施策総合推進法 (Power Harassment Prevention Act) case law, 呼び捨て (dropping -san) and タメ口 (plain form to a subordinate) are cited as 類型2 (精神的攻撃) evidence. The keigo detector makes that legally-recognized pattern structurally measurable.
One meeting is noise. The review-worthy-event scorer requires persistence (≥3 meetings over ≥28 days) AND directionality (concentration on one target) AND sustained drop (affected person below their own baseline). This team satisfies all three.
Orange = Nao's actual share · Dashed = Nao's own 90-day baseline (19%). Per-speaker baseline, not team average — defeats introversion / L2 / chair-role confounds.
The review-worthy event fires at Layer 5. Layer 6 then presents different framings to different roles, through RBAC + k-anonymity + differential privacy. No role sees data another role cannot. The employee sees things no manager sees.
“A pattern in your meetings this week.”
“Your week, seen from outside your own head.”
“Kimura (Product team) · Review recommended.”
Reads his Manager Mirror on Monday. Does not get lectured. Does not get called into HR. Sees his own behavior as data, with one concrete action.
On Tuesday, in the product review, he lets Nao finish her sentence. He uses “ミラさん” instead of “ミラ.” It takes 10 seconds longer.
That is the whole intervention.
Sees her pattern page on Monday. Reads the signals. Discovers she is not imagining it. The tool did not diagnose her, did not call it harassment — it described what happened.
She does not escalate yet. But she knows she could. And if she wanted evidence, she has it — encrypted, only she holds the key.
She stays. She comes back to the next meeting.
Reads the Executive Brief over coffee Monday morning. Sees Kimura flagged. Schedules a private 15-minute 1:1 with Kimura for Thursday.
Does not need to involve HR. Does not need to know who Kimura's affected team member is. Does not need to make any HR decision — the tool is prohibited from feeding those.
¥7.9M bill averted. The company never knows it was averted.
That is the whole product. Structural signals from meeting metadata → deterministic pattern construction → role-appropriate framing → one concrete action for the person in power. No content reading. No emotion inference. No HR escalation triggered. The CEO's labor-cost bill drops.