Positioning
Kevros occupies the enforcement layer, not the model-alignment layer.
The table is included to prevent category confusion. It is not a claim that other control layers are unnecessary. Model alignment, moderation, content classification, and runtime enforcement answer different questions.
| Capability | Kevros | Training-time alignment | Post-hoc moderation API | Content classification service | Inline guardrail service |
|---|---|---|---|---|---|
| Enforcement model | Fail-closed cryptographic gate | Training-time RLHF | Post-hoc filter | Content classifier | Inline content filter |
| Evidence chain | Hash-chained decision record | Model training artifact | Request or response log | Classification result | Policy decision log |
| Post-quantum signing | ML-DSA-87 (FIPS 204) SLH-DSA-256f (FIPS 205) |
Not the usual control point | Not the usual control point | Not the usual control point | Not the usual control point |
| Formally verified | Formal methods artifacts maintained | Depends on provider | Depends on provider | Depends on provider | Depends on provider |
| Provider coupling | Model-provider independent by design | Usually tied to model provider | Usually service-specific | Usually cloud or vendor specific | Usually gateway specific |
References for public Kevros claims are summarized on the Artifacts page.