Logs are valid. Metrics are green. The drift is already happening.
Aranthos is the external behavioral observability layer for autonomous AI in production. Read-only, statistical, explainable.
February 2024. An Air Canada chatbot told a grieving customer about a bereavement refund policy that did not exist.
The customer sued. The British Columbia Civil Resolution Tribunal held the airline liable for what their autonomous system said. The decision is now cited as a reference in Canadian AI liability case law.
The drift was visible in the system's behavior before the lawsuit. Logs were valid. Metrics were green. The signal lived in the gap between what the system was supposed to do and what it actually did.
Aranthos observes that gap.
One number. Five dimensions. Outside the model.
The Drift Health Index turns autonomous AI behavior into a continuous, decomposable score. The number ranges from 0 to 100. The five dimensions name what drifted. Both are observable from outside the system.
onboarding-bot . DHI 78 → 92
External, read-only, statistical. Standard observability boundaries. No SDK. No model access. No prompt storage.
The fleet, then the drift.
Five systems shown. One drifting. One warning. Two stable. One recovered. Aranthos names the ones that drift, then walks the team in. Each card is a system. Each score is a DHI. The drift surfaces before the page goes out.
Capture. Baseline. Score. Detect. Qualify. Route.
Read-only access to signals your stack already exposes. Provider-neutral. No SDK injected, no model retrained, no prompt stored.
Read-only ingest of operational signals.
Encrypted one-way pipeline. No callbacks.
Neutral statistical reference, learned.
5-dimensional DHI cycle, continuous.
Statistical outlier detection.
Drift typed and ranked across 5 dimensions.
Slack, Teams, email, webhook, audit.
Target 32ms median, <250ms p99 end-to-end. Audit-grade evidence trail at every step.
Outside the system, by architecture.
Aranthos observes the outputs your autonomous systems already emit on standard observability boundaries. The platform extracts statistical fingerprints sufficient to detect behavioral drift. Aranthos does not persist prompts or completions. Behavioral fingerprints are the only artifact stored over time.
This is an architectural choice, not a contractual promise.
customer-support-prod-03POST https://your.audit.endpoint- No SDK to install.
- No code change in your autonomous system.
- No model access required.
- No prompts or completions persisted.
- No vendor lock-in. LLM provider neutral.
- No engineering team rotation.
Compliance is jurisdictional. Architecture isn't.
Read-only behavioral observation is the only architecture compatible with autonomous AI audit, regardless of regulatory frame: EU AI Act articles 9 and 12, NIST AI Risk Management Framework, sectoral oversight in finance, gaming, healthcare.
The audit instrument is the same. The architecture is the proof.
Currently engaged on EU AI Act enforcement (August 2026). Security questionnaire and architectural walkthrough available on request.
Thirty minutes.
Your stack.
Your DHI.
We'll walk through what Aranthos would see on your autonomous systems. No SDK, no integration. Just a conversation.