The Nature of Vigilance

Meerkat is independent middleware. It does not generate content, summarise documents, or answer questions. It only verifies. A system that produces output cannot reliably audit that same output. Separation of concerns is not a preference; it is a security boundary.

Our verification pipeline cuts through AI confidence scores the same way. Token probabilities look convincing; independent analysis reveals whether the claim actually follows from the source. The glare is the model's fluency. The sunglasses are the verification layer.

Domain-specific adapters. A hallucinated medication dose is not the same class of error as a fabricated legal citation or an inverted financial figure. Each domain has its own purpose-built verification model, its own extraction rules, and its own risk thresholds. Generic verification is the wrong call.

Starter tier: 10,000 verifications, heuristic checks, sandbox environment. Professional tier: full verification pipeline, faithfulness scoring, production-grade SLA. Enterprise tier: advanced uncertainty detection, domain-specific models, air-gapped deployment, calibrated confidence guarantees. You do not hand a pup a live scorpion.

A confabulating model cannot detect its own confabulation. The error is invisible from inside. You need an external system with a different vantage point. That is the fundamental argument for third-party verification. Geoffrey Hinton said it plainly: "A confabulator is unable to detect its own confabulation."

Meerkat is not a single model. It is an ensemble: multiple independent verification methods working in parallel. Each component is specialised. The trust score reflects agreement across all of them. No single method is sufficient.

Three deployment models: cloud API for speed, hybrid containers inside your VPC, and fully self-hosted for air-gapped environments. Multiple authentication methods: API key, OAuth 2.0, Agent JWT, SSO. The infrastructure adapts to your security posture, not the other way around.

Continuous monitoring, not periodic audits. Real-time drift detection, anomaly surfacing, and adaptive thresholds that respond to changing error patterns. A quarterly compliance review is the equivalent of a sleeping sentinel. The threat landscape does not pause between assessments.

The developer experience: three API calls, five minutes to first verification. Free tier with full pipeline access. Sandbox mode for testing governance rules before enforcement. The integration should feel lightweight even though the verification is comprehensive.

Every verification produces an immutable audit record: what was checked, what passed, what was flagged, what was blocked, and how it was resolved. Full provenance chain from AI output to human decision. The audit trail is the bark. Compliance teams, regulators, and legal counsel hear it clearly.