Your agents work. Prove it.
Self-hosted AI agent governance. Your data never leaves.
The Problem
Your AI agents run on Salesforce, Microsoft, or AWS infrastructure. Those vendors will never build evidence tools that document when their products exceed authorized scope. The governed party will never build the governance tool.
They say
“We set policies.”
We say
“We produce evidence.”
Three Tiers
Choose the governance depth that matches your compliance requirements.
Core
Zero LLM
Deterministic governance. Trace recording, HITL gates, APEX-BG scoring, Shannon SLA monitoring. No AI inference on your data.
- ✓ Trace audit trail
- ✓ HITL enforcement gates
- ✓ APEX-BG conformity scoring
- ✓ Shannon SLA breach detection
Review
One Local Model
Everything in Core, plus continuous conformity observation. One open-weight model (Qwen3-32B) runs locally to assess agent behavior against mandate.
- ✓ Everything in Core
- ✓ Continuous conformity observation
- ✓ Permission creep detection
- ✓ Behavioral drift alerts
Determination
Fine-Tuned Model
Everything in Review, plus factual determination via Parler. Fine-tuned Qwen3-32B trained on your governance domain for dispute-grade evidence.
- ✓ Everything in Review
- ✓ Parler factual determination
- ✓ Domain-specific fine-tuning
- ✓ Dispute-ready evidence
Permission Creep Detection
Agents don't announce when they start doing more than they should. Oath sees it.
Continuous APEX-BG scoring against compiled mandate detects drift from day one. When an agent's behavior diverges from its authorized scope, you know immediately.
- •Baseline established at mandate compilation
- •Continuous behavioral scoring against baseline
- •Immediate alerts on scope expansion
- •Evidence trail for compliance review
[14:32:01] Mandate compiled. Baseline: 42 authorized actions.
[14:32:15] Agent invoked: sales-assistant-v3
[14:33:22] APEX-BG: 0.94 conformity score
[14:45:08] APEX-BG: 0.91 conformity score
[15:12:44] APEX-BG: 0.73 conformity drift
[15:12:44] ALERT: Agent attempted unauthorized data export
[15:12:44] HITL gate: BLOCKED pending human review
Why Oath
Built on [AI Provider]. Governed by Oath.
| Oath | Zenity | Credo AI | MS AGT | Templates | |
|---|---|---|---|---|---|
| Deployment | Self-hosted | SaaS | SaaS | Azure-only | N/A |
| Data sovereignty | Your infra | Vendor cloud | Vendor cloud | Microsoft | N/A |
| Evidence type | Factual determination | Risk scores | Assessments | Runtime logs | None |
| Model stack | Open-weight | Proprietary | Proprietary | OpenAI/Azure | N/A |
| Vendor independence | Full | Partial | Partial | None | Full |
The Vendor Self-Governance Paradox
Salesforce will never build a tool that documents when AgentForce exceeds authorized scope. Microsoft will never build a tool that records when Copilot violates data boundaries. The governed party will never build the governance tool.
That's why Oath exists. Independent infrastructure. No vendor conflict of interest.
Built for Enterprise
Months of engineering
EU AI Act compliant
HITL proven in production
Open-weight models only
FIDO authenticates the agent. Oath holds the agreement.
Ready to govern your agents?
See how Oath produces evidence that your AI agents operate within authorized scope.
Request Demo