How it works
1
Instrument
Add the Sengol SDK to your agent, send it OpenTelemetry spans, or call the REST API.
2
Evaluate
Evaluators score each response against a ControlBook: your regulations, expressed as configuration.
3
Gate
sengol gate blocks a release in CI when the agent misses its thresholds. Runtime guardrails can withhold a failing response.4
Prove
Export an evidence pack. Anyone can check it with the open-source
sengol-verify.Pick your journey
Platform operator
Install the appliance, load the licence, connect identity, assign roles.
AI developer
Evaluate your agent in code, add guardrails, and gate releases in CI.
Compliance and risk
Choose regulations, write the ControlBook, generate reports.
Human reviewer
Work the review queue and record decisions.
Auditor or examiner
Verify an evidence pack offline, without trusting the vendor.
Just looking
Run the full loop on your laptop in about 15 minutes.
What Sengol does not do
- It does not hold or proxy your model-provider keys.
- It does not train models. It governs a judge model that your own platform trains.
- It is not multi-tenant SaaS. A bank that needs hard separation runs a separate install.