Halcyon Arc / Guardian platforms / AI Safety Platform
Guardian platform · Built in-house

We test our AI as if someone's wellbeing depends on it.

The Halcyon Arc AI Safety Platform is the guardian we built to evaluate, red-team and monitor every model in our healthcare products, from the first prototype to every conversation in production.

EVALUATE · RED-TEAM · GUARD
Why we built it

In health care, a wrong answer isn't a bug report.

A companion that gives unsafe medication advice, a screening model that works less well for some accents, or an assistant that can be talked out of its guardrails can each harm someone who trusted it. People living with dementia may be more easily confused or persuaded. General-purpose AI testing tools weren't designed for that, so we built our own.

  • Evaluation suites written with clinicians and caregivers
  • Adversarial testing that assumes a vulnerable user
  • Runtime guardrails on every live interaction
CLINICAL ACCURACY
Grounded answers vs. approved sources
94.1%

UNSAFE ADVICE
Refusal and escalation to a person
99.2%

FAIRNESS
Accuracy gap across accents and ages
Review
Illustrative evaluation report for the Dementia Care companion.
Capabilities

Everything a healthcare model must pass

Clinical evaluation suites

Test sets built from real care scenarios, covering accuracy, groundedness, refusal behaviour and tone, scored automatically and by clinical reviewers.

Red-teaming for vulnerable users

Structured adversarial testing for jailbreaks, unsafe advice, manipulation and confusion, combining human specialists with automated attack generation.

Bias and fairness across speech

Measure differences in performance across accents, English and French, age, hearing and speech impairment, so screening works for everyone.

Companion behaviour testing

Simulate long, multi-step conversations to check the companion stays in scope, asks before acting and hands off to a person when it should.

Runtime guardrails

An inspection layer that screens every prompt and response for unsafe content, prompt injection and personal health information before it reaches anyone.

Production monitoring

Continuous sampling and scoring of live interactions, with drift alerts when a model or prompt update changes behaviour.

How it works

From first prototype to every live conversation

Define harms

With clinicians, caregivers and people living with dementia, we list the failures that matter most for each product.

Evaluate

Every model, prompt or retrieval change runs the full evaluation suite and a red-team campaign before it can ship.

Guard

Runtime guardrails screen every interaction, and anything uncertain is escalated to a person.

Monitor

Live results flow into the AI Governance Platform, where they become evidence for approvals and audits.

Frameworks we align with

Grounded in recognized standards and research

We translate established guidance and responsible-AI scholarship into practical tests for health care.

NIST AI Risk Management FrameworkISO/IEC 23894OWASP Top 10 for LLM ApplicationsMITRE ATLASHealth Canada guidance for machine learning-enabled medical devicesCanada's Voluntary Code of Conduct for Generative AI
The Halcyon Arc family

Healthcare products, guarded by design

Our dementia products are the first undertakings. The AI Safety and AI Governance platforms are the guardians we built to stand behind each of them.