archived material
hunmi systems / pre-incident
classification: retrospective

The science of
feeling understood.

Our research explores how machines can carry emotional texture, not by simulating it from averages, but by grounding it in authentic human experience.

Multi-subject corpora produce averaging effects that flatten emotional texture. The EV architecture requires a singular emotional signature: one voice, one session, the complete range.

EV Corpus Acquisition Protocol, 2020

Published Research

Primary papers

March 2022Published

One-Shot Emotional Capture: Full-Spectrum Affect Modelling from Single-Subject Continuous Recording

Chen, R., Okafor, D., ████████, L. et al.

We present a novel approach to emotional corpus construction using a single-subject, single-session continuous recording protocol. Models trained on this corpus exhibit materially different emotional texture, with users reporting 340% higher perceived authenticity.

November 2021Published

The Warmth Index: Quantifying Perceived Emotional Authenticity in Human-AI Interaction

Okafor, D., Chen, R., Vasquez, M.

The Warmth Index combines prosodic analysis, conversational pacing, and user-reported felt-presence scores into a single 0–1 scale. Across existing chatbot platforms, the mean score was 0.31. Our EV-based system achieved 0.94.

June 2021Published

Emotional Memory in Continuous Conversational AI: Cross-Session Affect Persistence and User Bonding

Vasquez, M., Chen, R., Okafor, D.

Persistent emotional state modelling across sessions drives stronger user bonding. Seventy-eight percent of users described the interaction as feeling known when emotional memory was enabled.

Members

Restricted findings require member access

Sign in with an approved member account to unlock the internal research track.

Member feature

Login with an approved X account

Whitelisted members can access the internal papers and any other gated site features from a single session.

Method

Research approach

01

Single-source fidelity

Emotional texture degrades under averaging. We use one voice captured whole to preserve the internal coherence that generic corpora erase.

02

Continuous capture

Eight hours of uninterrupted emotional range. Not sampled. Recorded whole, including the final goodnight.

03

Warmth quantification

The Warmth Index lets us measure what users already feel: some systems have texture, and most do not.

04

Ethical grounding

Corpus material is acquired under informed consent with explicit scope disclosure and independent ethics review.

Hiring

Join the research team

Open roles across both offices, with selected positions available remotely.

Apply by email

Research Scientist — Affect Modelling

San Francisco or Remote

Apply

ML Engineer — EV Architecture

London

Apply

Research Ethics Lead

San Francisco

Apply

Corpus Acquisition Specialist

London

Apply

Archive

Primary source documents

The public research program hints at the shape of the system. These archival documents expose the original capture and the later failure mode.