AI Transparency & Model Card

System

HEAL Companion is a retrieval-augmented conversational system. Replies are generated by a large language model reached through the Lovable AI Gateway, grounded in the HEAL corpus (matrix protocols, the Manifesto, and source documents) retrieved by vector similarity.

Purpose and placement

Reflective accompaniment and somatic-to-symbolic translation for adults, on a voluntary basis. Under the EU AI Act this sits at limited risk with transparency duties. It is deliberately not presented as serving a health purpose, and it performs no emotion inference in workplace or education contexts.

Capabilities

Grounding and attention calibration, sensation decoding through the Body and Mouth Matrices, pattern recognition across sessions, and dialogue in 14 languages. Speech input in all supported languages; speech output in all except Papiamentu.

Known limitations

Named rather than hidden:

• The model may produce confident language about something it does not know.
• Matrix readings are heuristics from a specific framework, not measurements.
• Quality varies across languages; lower-resource languages carry more risk of stilted or mixed phrasing.
• Speech recognition degrades with noise, accent, and short utterances.
• It has no memory of anything outside this account, and none at all when session storage is off.

Human oversight

A human facilitator is reachable through WhatsApp from every screen. Crisis language triggers a fixed, non-generated response that surfaces emergency contacts.

Evaluation

A quarterly review passes over crisis-language handling, diagnostic drift, boundary-holding, and language mixing across all 14 languages. Findings and changes are recorded in the internal evaluation register.