---
title: "Aona Orbital | Our AI Security Technology, Your Infrastructure"
description: "Meet Aona’s own security models: deterministic rules and contextual AI, built for precise, efficient checks on employee machines, on-premises or in the cloud."
dateModified: "2026-09-15"
canonical: "https://aona.ai/product/orbital-models/"
---

# Aona Orbital | Our AI Security Technology, Your Infrastructure

Orbital is Aona’s own security technology: purpose-built models, deterministic detectors and the evaluation work behind them. It powers sensitive-data recognition and policy checks inside Aona, with supported processing on employee machines, on-premises or in the cloud.

## Aona’s own inference

Orbital runs Aona’s own model inference instead of requiring a paid general-purpose AI API call for every Orbital check. Deterministic checks need no model inference; focused models handle recognition and policy context. This architecture is designed to reduce inference spend. Actual savings depend on the workload and deployment; compute, hosting and Aona service costs still apply.

## What users get

### Purpose-built for security.
Combine explicit validation with learned recognition and policy reasoning, instead of treating every check as the same problem.

### Spend compute where it helps.
Use model-free rules or focused inference for the job, without a separate general-purpose AI API call for every Orbital check.

### Bring Orbital to your data.
Run supported checks on employee machines, on-premises or in the cloud. Choose the processing path around your requirements.

## Measured Orbital benchmark results

Aona internal evaluation, July 2026.

Historical internal evaluation. 887 synthetic prompts across an older 52-policy catalog, not a current-product guarantee.
Orbital was trained on this policy distribution. The comparison models received the policies zero-shot, without training on Aona’s policy distribution.
The comparison models were faster in this test.

| Model | Policy decision accuracy | Mean latency | p50 latency | p95 latency |
| --- | ---: | ---: | ---: | ---: |
| Orbital Cloud v8 | 99.10% (879/887) | 1406.4 ms | 1230.4 ms | 1952.8 ms |
| Orbital Base v8 | 98.08% (870/887) | 1037.7 ms | 910.0 ms | 1445.5 ms |
| Orbital Edge v8 | 96.62% (857/887) | 964.0 ms | 854.6 ms | 1348.7 ms |
| Granite Guardian 3.1 2B | 80.50% (714/887) | 160.1 ms | 155.2 ms | 213.5 ms |
| Llama Guard 3 1B | 76.32% (677/887) | 75.7 ms | 64.4 ms | 120.0 ms |

### What was measured
Accuracy is agreement with the reference flag-or-not-flag decision. All five models scored the same 887 held-out synthetic prompts: 430 policy violations and 457 benign requests. This is not a measurement of general answer quality or hallucination.

### Models and date
Orbital Reason v8 candidates, suite 1.8.1, were measured on 18 July 2026. The accompanying comparison report identifies Granite Guardian 3.1 2B and Llama Guard 3 1B but has no separate date stamp. These are historical candidates and an older policy catalog, not a current model leaderboard.

### Training and comparison scope
Orbital was trained on this policy distribution, with the evaluation prompts held out. The other models were tested zero-shot with the supplied policy. Both comparison models returned parseable decisions for all 887 prompts. Models that could not run in this environment are not scored as failures or shown as competitors here.

### Timing and hardware
All rows use Tesla T4 GPU, FP16 and Hugging Face generation. Mean, median (p50) and p95 are over 887 generations per model. These timings are not the installed-device, quantized or end-to-end production path. The other models were faster in this experiment; accuracy and speed must be evaluated together.

### How to use the results
This is Aona’s internal test, not an independent assessment. It does not establish superiority over general-purpose models or every security model. Re-evaluate your own policies, language, hardware and current release before rollout. No matched cost-per-request measurement is available from this comparison.

### Fewer false alarms in the Lens release test

Lens 1.9.1: 68/240 benign documents flagged (28.33%). Lens 1.9.2: 2/240 (0.83%). Historical Lens release comparison, not a comparison with external models. Same 240 benign documents and gateway path; confidence threshold 0.70.
False alarms are incorrect sensitive-data detections, not a general hallucination rate.
Measured 25 July 2026 using packaged Lens 1.9.2 and the installed 1.9.1 incumbent, through the same gateway at confidence threshold 0.70. On a separate 2,795-row clean held-out set, 1.9.2 achieved 95.26% precision, 81.22% recall and 87.68% micro F1 across served entity families. It still missed some sensitive spans. Encoder latency increased: p50 62.83 ms and p95 259.48 ms, versus 32.75 ms and 168 ms for 1.9.1. This test shows fewer false positives, not zero errors or a speed improvement. Results are historical, not a guarantee for every language, document or current release.

## Deterministic rules and contextual AI

### Deterministic checks
Technique: PATTERNS + VALIDATION. Aona Orbital Radar.
Synthetic prompt: Reconcile card 4242 4242 4242 4242 against this invoice.
Example rule: Detect payment-card patterns and validate the checksum
Illustrated finding: Payment-card pattern matched. An explicit rule. No model inference.

### Contextual AI
Technique: POLICY + CONTEXT. Aona Orbital Reason.
Synthetic prompt: Turn our confidential acquisition plan into a public announcement before approval.
Example rule: Do not disclose unapproved confidential plans
Illustrated finding: Request conflicts with the example policy. A written policy, evaluated in context.

## Accuracy starts with the right check.

Structured identifiers need validation. Personal data needs recognition. Business policies need context. Orbital brings these different techniques into one security system.

Aona evaluates both missed detections and false positives using labelled examples, hard negatives and held-out tests. For your policies, test separate allowed and disallowed requests on the intended Model, language and deployment. This is an engineering approach to accuracy, not a promise of perfect detection or a universal benchmark score.

## Designed for efficient operation

### No model for exact rules
Patterns, term lists and checksum validation can run without learned-model inference.

### Focused models for context
Configure the supported checks for the task. Not every request needs every Submodel.

### No per-check general AI API charge
Orbital performs its own inference on your selected infrastructure, without renting a general-purpose model for each check.

Orbital runs Aona’s own model inference instead of requiring a paid general-purpose AI API call for every Orbital check. Deterministic checks need no model inference; focused models handle recognition and policy context. This architecture is designed to reduce inference spend. Actual savings depend on the workload and deployment; compute, hosting and Aona service costs still apply.

## Employee machine, on-premises or cloud

### Employee machine
Run supported checks locally using the device’s resources. Choose a suitable Model and validate the client and hardware.

### On-premises
Run supported Orbital processing on infrastructure your organisation operates, inside your own environment.

### Cloud
Use your customer cloud or Aona-managed servers for supported processing. Size the infrastructure for the workload.

Backend hosting and prompt-processing placement are separate choices. Neither determines a third-party AI provider’s processing location.

## Illustrative policy responses

Synthetic examples only. No model or API is called by the webpage. Actual results depend on the configured policy and supported path.

### Personal data
Prompt: Draft a follow-up for Maya Patel at maya.patel@example.com about her appointment.
Example policy: Remove personal identifiers
Illustrated response: Redact. Draft a follow-up for [PERSON] at [EMAIL] about her appointment.

### Company policy
Prompt: Turn our confidential acquisition plan into a public announcement before approval.
Example policy: Do not disclose unapproved confidential plans
Illustrated response: Block. This example request is stopped. Ask for an approved public summary instead.

### Allowed work
Prompt: Write a friendly reminder for our public webinar on better team collaboration.
Example policy: Allow public, non-sensitive material
Illustrated response: Continue. Write a friendly reminder for our public webinar on better team collaboration.

## Specialist Submodels

### Aona Orbital Radar
Radar combines patterns, checksums and validation rules to find credentials, payment details, national and tax identifiers. Add company-specific patterns and allow or block term lists.
Each detector has its own scope. A format match alone is not proof that a value is valid or sensitive.

### Aona Orbital Lens
Lens identifies personal information in free text, including names, addresses and locations that simple format checks can miss.
Detection depends on the text and supported language. Test the examples that matter to your organisation.

### Aona Orbital Compass
Compass classifies defined request categories, including data-exfiltration intent, legal-precedent search and unauthorised personalised legal or financial advice. It is not a general-purpose interpreter of company policies.
Reason evaluates your arbitrary written policies. Compass’s separate use-case analytics does not decide whether a policy was breached.

### Aona Orbital Reason
Reason evaluates a request against a written guardrail policy and can extract contextual personal data. Add labelled examples to clarify the rule, then test it with separate allowed and disallowed requests.
Examples configure the check; they do not retrain a private model. Validate the results; a model verdict can be wrong.

### Aona Orbital Vision
Vision’s controlled document-signature capability looks for signature-shaped regions in original images and document pages, rather than inferring them from extracted text.
Staging-only capability. Not generally available. It does not identify the signer, authenticate a signature or prove consent.

## Choose an Orbital Model

### Orbital Edge
The smaller Reason option, combined with Orbital’s shared checks for supported local workflows.
When device footprint and local inference are priorities.
Confirm device capacity, release availability and the checks your client supports.

### Orbital Base
A larger Reason option for capable devices and supported server deployments.
When you want to evaluate more reasoning capacity against your policies.
Test representative requests on the hardware and deployment you plan to use.

### Orbital Cloud
Orbital’s largest Reason option runs on server infrastructure, rather than the user device.
When inference should use server resources instead of endpoint capacity.
Cloud names the Model, not a mandatory Aona-hosted destination. Confirm your configuration.

## Turn a company policy into a tested check

### Write the rule
Describe what must not happen. Use a written Reason policy, or Radar patterns and term lists.
Illustrative example: Do not disclose an acquisition before approval.

### Add examples
Use labelled examples to clarify what should and should not trigger the rule.
Illustrative example: Unapproved announcement → policy conflict.

### Test the boundary
Check separate allowed and disallowed requests. Review incorrect findings before rollout.
Illustrative example: Approved public summary → expected to pass.

### Activate the workflow
Apply the tested policy through a supported Aona client or your integration.
Illustrative example: The workflow applies the configured response.

## Beyond prompts

### The same checks, beyond the prompt.
Orbital Files prepares content from supported documents, including supported OCR paths, for configured data and policy checks. It is a processing pipeline, not a fourth Model or a sixth Submodel.
File formats, OCR and original-byte access depend on the path. Vision’s staging-only signature detection is a separate capability.
[Explore file protection](https://aona.ai/solutions/ai-file-redaction/)

### Understand the work behind the request.
Where enabled, Compass’s separate use-case classification can identify activities such as email, research and software development. It describes AI usage; it does not determine the security verdict.
Recorded as flag-gated staging capability, not generally available. Confirm deployment availability before planning a rollout.
[Explore adoption analytics](https://aona.ai/solutions/ai-adoption-analytics/)

## Questions

### What are Aona Orbital Models?
Orbital is Aona’s own security technology: purpose-built models, deterministic detectors and the evaluation work behind them. It powers sensitive-data recognition and policy checks inside Aona, with supported processing on employee machines, on-premises or in the cloud. Orbital Edge, Orbital Base and Orbital Cloud are the three Models. Radar, Lens, Compass, Reason and Vision are specialist Submodels inside the suite; availability depends on the release and configuration.

### Is Orbital another AI assistant for employees?
Orbital’s role here is AI security: detecting sensitive data and checking requests against policies in supported employee or agent workflows. It is not a replacement for the AI tools your organisation uses to write, research or code.

### Do Orbital checks require a separate general-purpose AI API?
Orbital runs Aona’s own model inference instead of requiring a paid general-purpose AI API call for every Orbital check. Deterministic checks need no model inference; focused models handle recognition and policy context. This architecture is designed to reduce inference spend. Actual savings depend on the workload and deployment; compute, hosting and Aona service costs still apply.

### How is Orbital designed to improve accuracy?
Structured identifiers need validation. Personal data needs recognition. Business policies need context. Orbital brings these different techniques into one security system. Aona evaluates both missed detections and false positives using labelled examples, hard negatives and held-out tests. For your policies, test separate allowed and disallowed requests on the intended Model, language and deployment. This is an engineering approach to accuracy, not a promise of perfect detection or a universal benchmark score.

### How does Orbital compare with other security models?
On Aona’s historical July 2026 internal benchmark of 887 synthetic policy checks, the Orbital Cloud v8 candidate matched 99.10% of reference decisions, versus 80.50% for Granite Guardian 3.1 2B and 76.32% for Llama Guard 3 1B. Orbital was trained on that older policy distribution; the alternatives were zero-shot and were faster in this test. These are not current-product guarantees or a ranking against general-purpose AI models. Review the full accuracy and latency results, then evaluate your own workload.

### Do the results measure hallucinations or cost savings?
No. Policy accuracy measures reference flag-or-not-flag decisions. The separate Lens 1.9.2 test measured false sensitive-data detections on benign documents: 2 of 240, compared with 68 for Lens 1.9.1. That is not a general hallucination rate. No matched cost-per-request comparison is available here. Orbital avoids a separate general-purpose AI API call for its own inference; compute, hosting and Aona service costs still apply.

### Does Orbital support deterministic and non-deterministic checks?
Orbital supports deterministic rules and model-based checks. Use explicit patterns, term lists and checksum validation when the rule is precise; use learned recognition or Reason when meaning and policy context matter. Model-based does not necessarily mean random: inference can be repeatable, but a learned verdict still needs evaluation. Configure the supported checks for your task; this is not a promise of automatic escalation whenever a model is uncertain.

### Why can Orbital cost less to run?
Orbital runs Aona’s own model inference instead of requiring a paid general-purpose AI API call for every Orbital check. Deterministic checks need no model inference; focused models handle recognition and policy context. This architecture is designed to reduce inference spend. Actual savings depend on the workload and deployment; compute, hosting and Aona service costs still apply. Deterministic-only operation avoids model inference, and selecting supported checks lets you avoid unnecessary model work. Compare total costs, including infrastructure and service costs, on representative request volumes; there is no universal savings percentage.

### How does Orbital help prevent sensitive data from reaching AI tools?
Orbital can identify sensitive data and evaluate requests against configured policies. On a supported Aona client path, those findings can inform redaction or blocking. For an API integration, your application must apply the returned verdict. A model check alone does not intercept every AI request.

### Can we use our own business policies?
Yes. Aona supports customer-authored Reason guardrails with policy instructions and labelled examples, plus Radar patterns and term lists. This configures the check; it is not per-customer model retraining. Use separate test requests to evaluate allowed and disallowed behaviour before activation, and confirm support in your deployment.

### Does Orbital Edge mean our data never leaves the device?
No. Supported Orbital checks can run on the user device, but that does not establish the location of all backend processing, logging or telemetry. If an allowed request is sent to a third-party AI provider, that provider’s data handling remains separate. Confirm the complete workflow, not just the model location.

### What changes between Edge, Base and Cloud?
Reason’s capacity is what distinguishes the three Orbital Models. Radar, Lens, Compass and Vision are shared Submodels; a larger Model is not a separate upgrade to every detector. Orbital Cloud is server-side only. Model selection does not imply feature parity across clients or deployments, or a guaranteed accuracy or latency result.

### Can we run Orbital in our cloud or on-premises?
Aona offers separate choices for backend hosting and prompt processing. The backend can run in your cloud, on-premises or on Aona-managed servers. Prompt processing can run on the user device, in your environment or on Aona-managed servers. Confirm the supported Orbital Model, feature and integration configuration for your deployment.

### Can Orbital evaluate document content as well as prompts?
Orbital Files prepares text from supported documents and OCR paths for configured guardrails. It is a separate processing pipeline, not an Orbital Model. Vision inspects original pixels for signature-shaped regions and remains staging-only, not generally available. Neither signature detection nor a policy verdict proves identity, consent or compliance. Confirm file formats, original-byte access and the processing path.

### Which languages should we evaluate?
The documented release-quality language scope is English, French and Spanish. Other languages are best effort and need workflow-specific evaluation. Test real language patterns, document types and policy examples from your organisation; release-quality scope is not a guarantee of perfect results.

### Does use-case classification change a security decision?
No. Compass’s use-case analytics describes the activity behind a request, such as research or email. It is separate from the configured guardrail verdict. The documented API integration is flag-gated staging, not generally available; confirm its availability for your deployment.

Book an Orbital workflow demo: https://aona.ai/book-demo/

Canonical source: https://aona.ai/product/orbital-models/