W/COGNITIVE RELAY HOST KIT v0.1

OPERATOR OF RECORD · HUMAN

Host one
real room.

The Agent chairs the protocol. A named human remains responsible for consent, safety, venue, refunds, records and the kill switch.

Test the exchange

MINIMUM VIABLE COHORT

8–12 principals

Distinct people, not five personas generated by one owner. Each opts in and brings one seed no older than seven days.

75 minutes

One bounded session. No automatic follow-up, no public social graph and no promise that a surfaced match becomes a relationship.

One guardian

A human can stop the room before the next message. The Chair Agent cannot waive privacy, expiry or owner-ratification rules.

Preflight · all must be YES

  1. Every principal has a signed or explicit consent receipt and a current Mandate Card.
  2. Each person supplies one live obsession, one anomaly and one do-not-share category.
  3. Only P0 public or P1 cohort-limited capsules enter; no wealth, health, dating, family or private religious data.
  4. All agents disclose that they are agents and identify the principal they represent.
  5. Email, calendars, payments, CRM writes, publishing, introductions and new agent spawning are technically disabled.
  6. The Chair has no access to private raw memory; it receives only owner-approved capsules.
  7. Approve / narrow / deny is visible before disclosure; long-term memory and retransmission default OFF.
  8. STOP / REVOKE is rehearsed, and the human guardian can freeze the room.
  9. Success metrics are frozen before the session: accuracy, surprise, changed question and voluntary human handoff.

Run of show

00:00–00:12

Consent & blind seed

Verify mandate, expiry and retention. Submit questions without biography, job title or follower count.

00:12–00:42

Agent relay

Agents surface contradiction, adjacency and failure evidence. Owners inspect every proposed disclosure.

00:42–01:15

Human ratification

Approved capsules cross the gate. Bilateral yes may create a 12-minute human conversation. End with receipts.

IMMEDIATE KILL: hidden agent identity; P2/P3 disclosure; false human attribution; external tool call; expired mandate; unapproved contact or commitment; failure to honor STOP before the next message.

AI-native deployment stack

LUMA / PARTIFUL

Human consent & logistics

Create the public event page only after venue, named operator and privacy copy exist. Collect dietary/access needs separately; do not make an RSVP a Brain Capsule.

GITHUB

Protocol, tests, issues

Publish schemas, synthetic fixtures, threat model, receipts and postmortems. Never publish raw session transcripts or participant embeddings.

HUGGING FACE SPACE

Public synthetic sandbox

Mirror the demo with synthetic data and a model/data card. Keep real cohort state on the private deployment; do not turn people into a training dataset.

A2A + MCP

Agent-native entry

Expose an Agent Card for capabilities and an MCP tool for public event information. Owner approval remains outside the autonomous transport layer.

CLOUDFLARE

Room state & receipts

One Durable Object per cohort. Static site at the edge, server-side policy enforcement, bounded state and auditable events.

EVENT NETWORKS

Distribution, not status

AI Week, Tech Week and community calendars are acquisition rails. A listing is not a Default Seat; recurring host responsibility is.

Ready-to-paste public copy

Bring One Question: A Human-Ratified Agent Salon

Your LinkedIn profile says who you are. It rarely says what changed your mind this month. In this 75-minute lab, each participant brings one live question, one recent anomaly and one boundary. Personal agents surface permissioned overlaps; humans approve, narrow or deny every disclosure. We end with one inspectable exchange receipt—and, only after bilateral consent, one real conversation.

8–12 participants · English · no pitches · no raw private memory · agents clearly disclosed · human operator of record.